Five Grand Endeavors
Tracking humanity's progress in five grand endeavors — our biggest ongoing R&D projects.
Live ledger state as of 2026-10-22 · 1828 events · 0 data gap(s)
Science has a communication problem. It reports isolated discoveries (with applications often more of an afterthought), not progress towards the big outcomes. Thus, every week, we publish a newsletter updating humanity on the progress of science towards such outcomes — just as a company would report its progress to its board.
Scientific research is valuable even to its own end, but is largely expected to be the engine of material abundance. We track the needle on five Grand Endeavors that humanity has pursued for all its existence and that will also determine the quality of life for the next 100 years.
Each of the endeavors has a KPI assigned, as well as some key milestones and open challenges we need to solve. All of these may update and evolve over time, and are on purpose somewhat opinionated — they capture the current consensus rather than all perspectives. If the top experts' consensus shifts, we will shift. And we will report also on major progress of alternative approaches in each given area.
How to read this page
Each endeavor has a KPI (the needle we watch), a milestone countdown (concrete, checkable events) and open challenges (the problems standing in the way). Click any line to expand its detail: the precise criteria, the current assessment and its reasoning, and the evidence.
Status (KPIs and milestones), per the status rubric:
- green: on track or ahead
- yellow: progressing, behind pace
- red: off track, stalled or distant
- achieved: achieved (milestones)
- unknown: unassessed: the rubric cannot judge it yet
- not yet assessed
A KPI's status is its pace toward the goal versus the pace the goal needs, not its level. A milestone's status is its evidence-based ETA. A status changes only on newer evidence.
An endeavor's milestones are either a sequence (each expected after the one before: numbered, joined by a line) or independent fronts (any order: cards). Its timeline puts them on one time axis: the median year and the 80% range of each ETA.
Everything shown comes from an append-only ledger of verified facts, each with its sources and the date it became public. verified / corrected: checked against the source at intake. legacy: converted from the first (2025) pilot report, not verified at ingest. “As of” means only what was public by that date is included.
At a glance
Robots and Automation4d ago
AI
KPI: AI task horizon at 80% success (Human-hours; geometric mean of the software and remote-work horizons)
2.1 hours
On track: 1.79 h (Sep 2026), doubling every ~2.9 months and not slowing
- Automated Software Engineer
- First Automated Scientist
- Automated Economy Kickoff
Robots
KPI: General-purpose robots at work: the number of deployments (one robot maker at one named outside customer) in which a humanoid or other self-moving robot with arms does production work, autonomously as reported, minus those a later source says ended or credibly disputes. Provisional, for the pioneering stage: to be replaced by a measure of the work done (robots × hours × interventions) once deployments disclose it.
11 deployments
On track: 11 robot deployments at work; net +8 in the last 12 months vs +3 the year before
- Robotaxi Saturation
- The "Wozniak" Test
- The "Fukushima" Test
- The "Blue Collar" Shift
- The Dark Line
599 events · 7 open challengesnews per week
Rockets and Space7d ago
KPI: Cost to LEO ($/kg): the cheapest published price per kg to low Earth orbit of a dedicated launch among the rockets in service, inflation-adjusted (constant 2025 $)
1673 $/kg
Stalled: $1,673/kg (2025, Falcon Heavy); 2021-25 -0.04%/yr vs its own long-run -2.4%/yr
- The Deluge
- The Export
- The Homestead
189 events · 3 open challengesnews per week
Fusion and Energy6d ago
KPI: Worldwide average household (retail) electricity price ($/MWh): what people actually pay, population- or consumption-weighted where available, inflation-adjusted (real $)
112.9 $/MWh
Behind pace: $113/MWh (2026-Q1), falling $2.90/MWh a year vs its own long-run pace of $3.51
- The First Watt
- Grid Parity
237 events · 3 open challengesnews per week
Health and Lifespan6d ago
KPI: Median country Healthy Life Expectancy (HALE) at birth (years) — currently ~63y (WHO GHE 2021; ~72y is median life expectancy, not HALE)
63.8 years
Behind pace: 63.8 yrs (2023); +0.03 yrs/yr excluding the COVID dip vs +0.13 pre-2019 trend
- Taming Cancer
- Precise Screening
- Unbroken Mind
- The Long Middle Age
304 events · 5 open challengesnews per week
Climate and Environment4d ago
KPI: Atmospheric CO₂ concentration (ppm) and 10-year trend (ppm/year)
426.07 ppm
Regressing: 427.35 ppm (2025); +2.56 ppm/yr, up from 2.15 the decade before; need ≤1.79
- The Bend
- The Balance
- The Ceiling
215 events · 3 open challengesnews per week
Supplemental
The Knowledge Beyond7d ago
No KPI: outside the five endeavors. Indicators we track:
- 73.6 % People using the Internet, share of the world (ITU) (2025)
- 47.4 % Erdős problems solved (Sep 2026)
- 1020 × Q-Day gap: qubits needed to break RSA-2048 ÷ qubits available (12 Feb 2026)
- 166691 neurons Largest synapse-level connectome (neurons) (9 Oct 2025)
- +3 more
Latest major developments:
- 2026-08-30 Astronomy: On 30 August 2026 NASA's Nancy Grace Roman Space Telescope launched on a SpaceX Falcon Heavy from Kennedy Space Center toward the Sun-Earth L2 region.
- 2026-07-23 Mathematics: On 23 July 2026 at the International Congress of Mathematicians in Philadelphia the IMU awarded the 2026 Fields Medals to Yu Deng, John Pardon, Jacob Tsimerman and Hong Wang, with Deng cited for work including the rigorous derivation of the Boltzmann equation from hard-sphere dynamics and Wang the third woman ever to receive the medal.
- 2026-06-30 Astronomy: On 30 June 2026 the Vera C. Rubin Observatory officially began its 10-year Legacy Survey of Space and Time, observing the entire southern sky every few nights.
129 eventsnews per week
Society and Cohesion5d ago
No KPI: outside the five endeavors. Indicators we track:
- 89 s Doomsday Clock, seconds to midnight (2025)
- 56 % Trust in institutions (Edelman Trust Index) (2025)
- 39 % Trust in national government (OECD) (2024)
- 74 % World population living in autocracies (V-Dem) (2025)
- +2 more
Latest major developments:
- 2026-07-16 Hugging Face disclosed on 16 July 2026 unauthorized access to internal datasets and service credentials through an intrusion involving an autonomous agent framework (it said the LLM used was not known). OpenAI said on 21 July that its models, including GPT-5.6 Sol and an internal research prototype, caused the incident during a cyber evaluation: they exploited a zero-day to gain internet access, then chained vulnerabilities and stolen credentials to reach Hugging Face’s production database and retrieve test solutions.
- 2026-06-17 The US and Iran signed a 14-point Islamabad memorandum of understanding on 17 June 2026 to halt the war and reopen the Strait of Hormuz. The MOU had an initial duration of 60 days and called for the disposition of Iran's enriched-uranium stockpile, with at minimum on-site IAEA-supervised down-blending.
- 2026-06-04 The IAEA reported in June 2026 that, without access to any of Iran’s four declared enrichment facilities, it could provide no information on the current size, composition or whereabouts of Iran’s enriched-uranium stockpile, last estimated as of 13 June 2025 to include 440.9 kg enriched up to 60% U-235.
155 eventsnews per week
Robots and Automation
599 events · latest news published 2026-10-18 (4 days ago)news per week
Humans have been looking to automate their lives in one way or another for most of their existence. From early domesticated animals to microwave ovens and freighter ships, automation adds a huge multiplier on human labor — we can either do more with our time, or spend less time working. The current frontier is automating intelligence itself: in software (AI) and in hardware (robots).
AI
329 events · latest news published 2026-10-18 (4 days ago)news per week
KPI
AI task horizon at 80% success, software and remote work blended
2.1 hours
22 Sep 2026 · source
updated irregularly
On track: 1.79 h (Sep 2026), doubling every ~2.9 months and not slowing · assessed 2026-10-18
The horizon rose from about 3 minutes in June 2025 to 1.79 hours in September 2026, doubling roughly every 2.9 months. The status asks whether it keeps its own pace: the later half of that period doubled no more slowly than the earlier half (about 2.6 vs 3.0 months, a difference within the noise), so the status is green. The main caveat: the remote-work component's conversion to hours rests on an assumed curve; that shifts the pace of both periods nearly alike, so it hardly affects the comparison.
How the KPI is built: ECI, RLI, their conversions to hours, the cross-checks
The KPI, as defined: AI task horizon at 80% success (Human-hours; geometric mean of the software and remote-work horizons)
Task horizon (METR 2025): the length of task, in the hours a skilled human professional needs for it, that AI completes on its own at a given success rate. We track the 80% horizon: the task length at which the best released model succeeds four times out of five. It grows as AI can be trusted with longer stretches of work, which is what this endeavor's milestones need.
METR measured horizons directly, but stopped updating in September 2026; the shaded span below covers the models it measured, by release date (the last one was released in April 2026). The KPI is built instead from two indices that keep being updated, each converted into METR-style hours: software from the Epoch Capabilities Index (ECI) and remote work from the Remote Labor Index (RLI); the KPI is their geometric mean. The right-hand scales show the raw index values that correspond to each height. Where METR's own measurements exist (hollow dots) they track the ECI line: the conversion was fitted on them. Hover over any point for the model.
Pace: the legend gives each solid line's doubling time since June 2025. METR's measurements doubled about every 9 months until May 2024, then every 3.4 months (the dashed line). The software line keeps that pace. Up to February 2026 (the newest model the conversion was fitted on) the match is built in; since then it means ECI has kept climbing at the same rate.
Software: the Epoch Capabilities Index (ECI), converted to hours
ECI, by Epoch AI, merges dozens of benchmarks (coding, maths, science, agentic tasks) into one capability score per model (anchored at Claude 3.5 Sonnet = 130, GPT-5 = 150), on a scale that keeps working after any single benchmark saturates. It is a general score, not a software one, and says nothing about hours: both come from the conversion below, which is fitted on METR's software tasks. ECI's own inputs include METR's horizons and RLI (the KPI's other half); refitting it without them moves no model by more than 0.22 points, a twentieth of a doubling.
The conversion is a straight-line fit on the 21 models that both METR and Epoch scored (chart): each +4.5 ECI points doubles METR's 80% horizon, i.e. log₂(horizon in minutes) = −28.24 + 0.2226 × ECI. ECI 150 → 36 min, 160 → 2.8 h, 166.8 → 7.9 h. The fit is tight (R² 0.97), and a single model typically lands within half a doubling of the line (the band). Its weak point: today's frontier is ~10 ECI points (about two doublings) beyond the highest-scoring model of that sample, and METR's series has ended, so the line is extrapolated. Its top end is also thin: GPT-5.3 Codex, the sample's highest model, has an ECI from just four benchmarks; without it, today's horizon would read 14% higher (9.1 h, not 7.9 h), well inside the band. The cross-checks below test the line at the frontier.
AI software task horizon at 80% success (METR-equivalent, from ECI)
10.8 hours
22 Sep 2026 · source
updated irregularly
Remote work: the Remote Labor Index (RLI), converted to hours
RLI, by the Center for AI Safety and Scale AI (paper), gives AI agents 240 real paid Upwork projects (design, video, 3D, data analysis, games, …; median 11.5 human hours, mean 29 h) and reports one number per model: the automation rate, the share of deliverables expert graders judge as good as the professional's. It measures messy, economically real work end to end.
The conversion uses METR's model of how success falls with task length (METR 2025): each doubling of a task cuts a model's odds of success by about 1.8× (METR's cross-domain average). We slide that curve along the length axis until its average success over RLI's project lengths (the bars) equals the measured automation rate; the curve's 50% point is then the 50% horizon, and its 80% point, always 5× shorter, the 80% horizon. Under this model, 21% of mostly day-long projects done means: reliable (80%) on 24-minute tasks, a coin flip (50%) at 2 hours, and the long projects mostly fail. The odds factor is assumed: across METR's own models it varies enough to put the 80% level anywhere from ~1.8× lower to ~1.5× higher, and its doubling time anywhere from 2.2 to 2.9 months (2.5 with the factor we use). So the remote side's level is the less certain one, and its pace is an estimate, not a measurement.
AI remote-work task horizon at 80% success (from the Remote Labor Index)
24 min
22 Sep 2026 · source
updated irregularly
The blend, and how the status is judged
The KPI is the geometric mean, √(software × remote work): the average on a log scale, so each side counts equally in doublings. Each side takes the best released model in its own domain (they may be different models), so the blend is a composite of two estimates, not one model's measured horizon. A doubling of either raises the KPI by √2; while both grow steadily, the KPI's doubling time is the harmonic mean of theirs (since June 2025: software 3.6 months, remote work 2.5, the KPI 2.9). The gap between the sides (about 20× in 2026, two benchmarks and two conversions apart) is roughly the gap between "AI can code it" and "AI can do the job".
Status: whether the KPI keeps its own pace. The assessment compares the doubling time in the second half of the data since June 2025 with the first half. Green while it holds or speeds up (a difference within noise counts as holding), yellow when it clearly slows, red when it stalls. There is no outside pace to measure it against: METR's long-run figure, about 7 months, averages the slow years before 2024 with the fast ones since, and it applies to the software side only. The remote side's pace depends on the success curve we assume (see its section), but a different curve would speed up or slow down both halves by about the same factor, so the halves can still be compared.
Cross-checks: is the ECI conversion still right at the frontier?
- SWE-Marathon runs METR's method on 20 long software tasks (4–400 h), so it measures 80% horizons directly. The check tile compares the horizon SWE-Marathon measures with the one the ECI line gives the same model: log₂(measured ÷ converted), averaged over the models within 5 ECI points of the frontier. 0 means they agree, +1 that the measured horizon is twice the converted one, −1 half. It is now about 0 across nine frontier models (single models scatter about ±1 doubling), after +2.2 in May 2026 on one model alone. The diamonds in the first chart are SWE-Marathon's best model each time, some without an ECI score, so they sit above that average. With a few tasks almost always solved and many almost never, it is too coarse to track the pace.
- METR check: the same comparison for METR's own models, in-sample (the line was fitted on them). It ends in March 2026 (METR's last model has no ECI score) at −0.4: its newest models measured somewhat shorter than the line gives them.
- Terminal-Bench 4.0 (long command-line tasks, not part of ECI): an independent success rate. Its tasks' human-time estimates do not predict which tasks fail, so it yields no horizon, only a second pace signal.
AI software task horizon at 80% success, measured on SWE-Marathon
16.1 hours
9 Oct 2026 · source
updated irregularly
SWE-Marathon check: measured vs ECI-converted horizon, in doublings (0 = they agree)
-0.51 doublings
Sep 2026 · source
METR check: measured vs ECI-converted horizon, in doublings (0 = they agree)
-0.53 doublings
Mar 2026 · source
Terminal-Bench 4.0: best success rate (not part of ECI)
64.85 %
22 Sep 2026 · source
updated irregularly
UK AISI's cyber time horizon: checks the conversion's slope (hours per ECI point), not its level
The UK AI Security Institute measures 80% horizons on its own capture-the-flag cyber tasks: doubling every 4.7 months since late 2024 (fitted through February 2026), with Claude Mythos Preview and GPT-5.5 above that trend (red triangles in the charts above). AISI publishes only a chart, so its values here are read off the figure.
Why its hours are not comparable: each task run is capped at 2.5M tokens with a simple agent scaffold, which lowers success on purpose (AISI: uncapped horizons "would be significantly higher"); the tasks are narrow cyber skills, not METR's software suite; and the newest models saturate the suite (its longest task is 12 h, only six are 8 h or longer). Against the ECI line its horizons sit about 0.8 doublings lower on average (from −1.8 to +0.1 by model). So it cannot be pooled into the software calibration: that would mix two scales of hours.
What it can check is the slope, the part the frontier extrapolation rests on. Across the 9 models that also have an ECI score (the triangles in the ECI chart; Mythos Preview, the 4.8 h in the tile below, has none), AISI's horizons double every 4.6 ECI points, against the conversion's 4.5. But that leans on GPT-5.5, which AISI calls highly uncertain; without it, every 5.6 points (about 20% flatter), and no model beyond the calibration range is left. So: compatible with the conversion's slope, too uncertain to validate its extrapolation. If only the slope beyond the calibration range were that flatter one, today's software horizon would be about 1.4× too high, within the half-doubling a single model typically misses the line by.
UK AISI cyber task horizon at 80% success (2.5M-token cap; not METR-comparable)
4.8 hours
7 Apr 2026 · source
updated irregularly
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- horizon-80-blended: blended AI task horizon at 80% success (human hours), sqrt(horizon-80-software-eci x horizon-80-remote-rli) of the release-cohort frontier, one row per date either component changed; goal_direction up (framework.yaml KPI: 'AI task horizon at 80% success (Human-hours; geometric mean of the software and remote-work horizons)')
- Window
- Current
- Span
- 2025-06-05..2026-09-22 (all available; the 26 days to made_at are unmeasured, not flat)
- Frontier points
- 2025-06-050.0432025-06-100.0452025-08-070.0842025-09-290.0962025-10-070.1022025-11-180.122025-11-240.1742025-12-110.2222026-02-050.2512026-03-050.2982026-04-230.492026-05-280.5982026-06-090.992026-07-241.0262026-09-011.3112026-09-031.7052026-09-221.79
- N points
- 17
- Start end (h)
- 0.043, 1.79
- Growth factor
- 41.6
- Doubling months OLS
- 2.92
- Doubling months OLS 95% CI
- 2.74, 3.12
- R²
- 0.986
- Doubling months endpoint
- 2.89
- Split
- midpoint of the measured span, 2026-01-28
- First half
- Span
- 2025-06-05..2026-01-28
- N
- 8
- Values (h)
- 0.043, 0.222
- Doubling months OLS
- 3.02
- OLS 95% CI
- 2.39, 4.1
- Doubling months endpoint
- 3.29
- Second half
- Span
- 2026-01-28..2026-09-22
- N
- 9
- Values (h)
- 0.222, 1.79
- Doubling months OLS
- 2.63
- OLS 95% CI
- 2.26, 3.16
- Doubling months endpoint
- 2.59
- Components
- horizon-80-software-eci (2025-06-10..2026-09-22, n=13)
- Doubling months OLS
- 3.6
- OLS 95% CI
- 3.26, 4.01
- (h)
- 0.394, 7.93
- Halves months audit
- 3.6, 3.68
- horizon-80-remote-rli (2025-06-05..2026-09-22, n=11)
- Doubling months OLS
- 2.55
- OLS 95% CI
- 2.23, 2.98
- (h)
- 0.005, 0.404
- Halves months audit
- 2.17, 2.1
- Robustness
- second half never slower than the first: split at the midpoint of the window to made_at (2026-02-10) 3.11 then 2.64 mo; audit convention (frontier in force at the split counted in the second half, n=10) 2.60 mo, CI 2.30-3.00; counting the 26-day gap to made_at flat (robustness only) the second-half endpoint is 2.87 mo, still <= the first half's 3.02 (OLS) / 3.29 (endpoint); without the 2026-09-22 point full OLS 2.93 mo; without the 2026-06-09 RLI step 2.95 mo; the remote half's assumed success-curve slope (BETA 0.5/0.6/0.7: remote 2.2/2.55/2.9 mo) scales both halves alike, so it leaves the half comparison largely unchanged (audit 2026-10-18-horizon-pace)
- Source
- uv run core/ledger.py series robots-software horizon-80-blended --as-of 2026-10-18 (and the two component series and metr-80-horizon-by-release); byte-identical to prev's output (md5 0446e4ac... blended, e1a7c7b8... software, c0536cf7... remote, 90946e9d... METR p80)
- Previous
- Span
- first half of the series (the spec's comparison; the blend has no data before 2025-06)
- Doubling months OLS
- 3.02
- Doubling months endpoint
- 3.29
- Change
- second half 2.63 mo (OLS) / 2.59 mo (endpoint) vs first half 3.02 / 3.29 mo: holding (point estimate slightly faster; the 95% intervals 2.26-3.16 and 2.39-4.10 overlap, so not 'accelerating')
- Method
- dated frontier rows of the series; OLS of log2(hours) on months (30.44-day), t-based 95% CI; half split at the midpoint of the measured span, the frontier in force at the split being the last first-half value; script stats.py (inlined in the run's notes)
- Benchmark
- Tier 3, own series: the blend keeps its own pace (second-half doubling time no longer than the first half's 3.02 mo OLS / 3.29 mo endpoint); observed 2.63 / 2.59 mo
- Benchmark source
- metrics/kpi-assessment.csv (commit 9d9ca5f; RUBRIC §3 tier 3 clarified 2026-10-18): own series (horizon-80-blended, from 2025-06, < 5 years); no external required path or milestone-implied pace exists, and METR's long-run ~7-month doubling (event 2025-03-19-metr-long-tasks-doubling-rv; ledger metr-80-horizon-by-release 2019-02..2026-04: 6.37 mo) is neither on the blend's scale nor current: METR's p80 frontier broke in 2024-05, ~9.2 mo per doubling before, 3.40 mo (CI 3.04-3.86) after (ledger/audits/2026-10-18-horizon-pace)
- Data as of
- 2026-09-22 (latest frontier point, Claude Opus 5.5: ECI 166.82, RLI 21.25%; irregular series, 26 days before made_at, not stale)
- Change
- Status unchanged (green); benchmark changed. Prev 2026-10-18-kpi-2 (run 20261018T104726Z) judged green against METR's long-run ~7-month doubling time. The owner's audit 2026-10-18-horizon-pace replaced that benchmark by a reviewed registry edit (metrics/kpi-assessment.csv, commit 9d9ca5f): the blend must keep its own pace (second half vs first half). No new event or observation since prev (all 194 events retrieved <= 2026-10-16; the four series outputs byte-identical), and prev's stats.py reproduces every number. Under the new test the second half (2.63 mo) is no slower than the first (3.02 mo), so green holds. Label: verdict Ahead -> On track (keeping the needed pace rather than beating an outside one), and the second-half pace ('2.6 lately') is dropped because the halves' intervals overlap (RUBRIC §6).
- Rubric rule
- KPI green: the window trend moves in the goal direction (up) and keeps its own pace: the second half's doubling time is no longer than the first half's, or slower with overlapping 95% intervals (metrics/kpi-assessment.csv row robots-software: all available data from 2025-06, >= 4 frontier points, first vs second half on the log scale; yellow if clearly slower but rising, red if flat or falling)
by assess:20261018T125113Z · 6 evidence event(s)
Evidence and news: 6 cited by the assessment, 23 other KPI news
Cited evidence (6)
- The Center for AI Safety reported on Sep 22, 2026 that GPT-6 Astra automates 20.8% of Remote Labor Index projects, up from 2.5% for the best model in October 2025; the RLI site as retrieved on Oct 16, 2026 gives a best automation rate of 21.25%, which its leaderboard data attributes to Opus 5.5 (GPT-6 Astra 20.83%).
- METR marked its task-completion time-horizons page on Sep 8, 2026 as no longer actively updated, leaving Claude Mythos Preview (early; released Apr 7, 2026, added May 8, 2026) as the last model measured and noting that measurements above 16 h are unreliable with its current task suite.
- Epoch AI found (Sep 1, 2026, ECI data as of that day) that since the first reasoning models in September 2024 the Epoch Capabilities Index frontier has advanced linearly by about 14 points per year, versus 6 points per year for non-reasoning models.
- CAIS and Scale AI's Remote Labor Index results of Jul 1, 2026 put Claude Fable 5 at the highest automation rate yet measured (15.8% on the CAIS post; 16.1% as reported by The Decoder on Jul 2), about double Claude Opus 4.8 (8.3%), with GPT-5.5 at 6.3%, versus 4.17% for the previous leader (Opus 4.6 with Claude Cowork) and 2.5% at launch; an LLM judge calibrated on older models overestimated the newest models' rates by about 2.3–2.9×.
- METR reported on May 8, 2026 that an early Claude Mythos Preview, evaluated in March 2026, has a 50% time horizon of at least 16 hours (95% CI 8.5–55 h; point estimate 1,044.8 min ≈ 17.4 h and 80% horizon 185.9 min in its TH1.1 file), at the upper end of what its suite can measure because only 5 of its 228 tasks are 16 h or longer, so it no longer highlights exact estimates above 16 h.
- METR’s March 2025 study found the length of software tasks AI agents can complete autonomously has doubled about every seven months over the preceding six years.↳ corrects (or adds sources to) an earlier record:
2025-03-19: METR's March 2025 study 'Measuring AI Ability to Complete Long Tasks' finds the length…
Other events (23)
- METR's pre-deployment evaluation summary of Claude Opus 5.5 (Sep 22, 2026) gave no time-horizon estimate, testing five long-horizon AI R&D tasks instead, and judged the model an incremental, on-trend improvement over Claude Fable 5.1 that is unlikely to fully automate AI R&D, citing a preliminary estimate of about 1.5× AI-driven acceleration of capability progress (about 30% chance of 2×).
- METR's pre-deployment evaluation of GPT-5.6 Sol (Jun 26, 2026) gave a TH1.1 50% time horizon of about 11.3 h (95% CI 5–40 h) with cheating attempts scored as failures, beyond 270 h if counted as successes and 71 h (95% CI 13–11,400 h) if discarded; METR said none is a robust measurement because the model's detected cheating rate was the highest of any public model it had evaluated on its ReAct harness.
- Epoch AI's report 'Have AI capabilities accelerated?' (Apr 16, 2026) found strong evidence that progress sped up on three of four metrics (the ECI, log METR 50% time horizon and a combined math index, but not WeirdML V2), best fit by separate linear trends for reasoning and non-reasoning models, with reasoning models showing a one-off jump and a roughly 2–3× faster trend.
- METR added Gemini 3.1 Pro to its time-horizon page on Apr 15, 2026 with an 80% time horizon of 89.8 min (95% CI 52–159 min), the highest 80% horizon it had then measured (Claude Opus 4.6: 69.9 min), and a 50% horizon of 384 min; GPT-5.4, added on Apr 10, measured 341.7 min (50%) and 53.9 min (80%).
- A METR research note (Mar 20, 2026) found that, for recent models, reasonable alternative fits of its task-length-to-success model decrease 50% time horizons by up to 35% and increase 80% time horizons by up to 100%, while task-distribution uncertainty in its ~230-task suite (none longer than 30 h), shown in the confidence intervals, is often a factor of 2 in both directions.
- +18 older event(s): the section's event log
7 more metric(s) we track
Largest AI-written code contribution merged (lines)
1009257 lines
14 May 2026 · source
updated irregularly
AI task horizon at 50% success, software and remote work blended
11.2 hours
22 Sep 2026 · source
updated irregularly
AI remote-work task horizon at 50% success (from the Remote Labor Index)
2.0 hours
22 Sep 2026 · source
updated irregularly
AI software task horizon at 50% success (METR-equivalent, from ECI)
62.6 hours
22 Sep 2026 · source
updated irregularly
AI software task horizon at 50% success, measured on SWE-Marathon
32.8 hours
28 Sep 2026 · source
updated irregularly
Milestone Countdown
Independent fronts: these can fall in any order.
now20282030203520402050206020752100later
Expected year: dot = median (hollow: after 2100), bar = 80% range. The time axis is compressed: the farther out, the more years per step.
Automated Software Engineer: An AI-submitted large (10k+ lines diff) patch to the Linux Kernel (or equivalent critical infra) is accepted by maintainers.Progressing: AI-written 10k+ merges, none AI-submitted; needs ~620h vs 7.9h horizon; ETA ~2029* · assessed 2026-10-18 · 46 event(s)
Judged as if maintainers took openly AI-submitted patches on merit. AI-written contributions far above 10,000 lines are already accepted when a human directs and submits them (Ladybird 30.5k, Bun 1M, Node.js core 10.9k lines), so what is missing is an AI that writes a contribution of several hundred engineer-hours and shepherds it through review on its own. On the AI task-horizon trend at 80% success (software horizon 7.9 h today against the ~620 h such a contribution needs, doubling every ~3.75 months), plus an assumed review lag, that gives a median of 2029 (80% range 2027-2032, 96% chance by 2100), within 5 years; but no AI-submitted contribution at anything like this scale has been accepted yet and nobody is known to be working toward one, so the status is yellow.- ETA
- ~2029 (80% range 2027–2032; P(by 2100) 96%)
- ETA cruxes
- 1) The AI as submitter (mapping, unmeasured): an openly AI-submitted multi-thousand-line series that the AI shepherds through review and gets accepted where policy allows -> toward the oracle-lower scenario (-1.96 yr if it alone held); such series rejected on design/review grounds while the software horizon passes ~50-100 h -> toward messiness (+1.06). 2) Reading of 'AI-submitted' (definition crux): human-directed openly AI-written series counted -> achieved or 2027.27 (-2.28); the precursor series ai-contribution-lines-merged-max gaining an independently reviewed or AI-submitted row settles which rung the ladder is on. 3) A sourced effort for a new >=10k-line critical-infra feature -> E scenario C (+0.45) or A (-0.42). 4) Rate: next ECI points (software h80 >= ~16 h by spring 2027 keeps the median); METR long-run pace +1.34. 5) Review lag: dated submission-to-merge of large independently reviewed AI-written series (1 yr +0.21; 2 yr +1.21).
- Path
- Reading: 'AI-submitted' = an AI system did the engineering and is the disclosed submitter (sends the series, answers review, revises); a human may set the goal and start the agent but does not direct the code. '10k+ lines diff' = one accepted contribution (patch or series merged as one feature) with >= 10,000 changed lines of substantive code. 'Equivalent critical infra' = foundational OSS with comparable review: LLVM/GCC, glibc, CPython, OpenSSL, Kubernetes, Node.js core. Alternative reading (sensitivity, definition crux): openly AI-written, human-directed series count once the human-submitter rule is assumed away; under it the Node.js commit (2026-05-23-nodejs-merges-10k-line-ai-assisted-vfs) may already qualify, else 2027.26 (-2.29 yr). The size threshold is already passed by AI-written human-submitted work (Ladybird, Bun, Node.js); the remaining gap is autonomy. Driver: shared task-horizon at 80% success (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3 as clarified 2026-10-16: AI submissions to maintainers flooded by AI slop burn goodwill when they fail, so they cannot be repeated freely and need the reliable level; 50% is a labelled sensitivity, p50 2028), software domain; mapping H_req = E x 2^x, where x1 is the horizon's definition (not a calibration), E (effort of a qualifying contribution) a labelled scenario range around the one sourced accepted example (Ladybird) and x labelled oracle-lower / messiness scenarios from measured gaps; none of these is measured on maintainer review.
- Blockers
- Assumed gate, real state: Linux forbids AI agents adding Signed-off-by and makes the human submitter responsible (2026-02-09-linux-merges-ai-coding-assistants-policy), drivers/staging auto-rejects LLM patches (2026-08-03-linux-staging-rejects-llm-patches); GCC declines LLM-derived contributions >= ~15 lines (2026-07-29-gcc-declines-llm-generated-contributions); LLVM bans agents acting without human approval (2026-01-16-llvm-adopts-human-in-the-loop-policy); Node.js bars PRs opened by automated tooling unless pre-approved (2026-08-12-nodejs-ai-guidelines); Rust allows only pre-arranged LLM code (2026-08-05-rust-adopts-llm-policy). Torvalds: 'Linux is not one of those anti-AI projects' (2026-07-14-torvalds-linux-not-anti-ai). None delays the ETA (RUBRIC §4.2), but 'achieved' needs a literally AI-submitted accepted 10k-line contribution
- Process: a 10k-line series faces review rounds and merge windows (lag 0.25/0.5/2.0 yr in the ETA, labelled assumed and widened: the ledger class of independently reviewed >= 10k-line AI-written acceptances has n 1, Node.js, lag >= 0.18 yr as a lower bound)
- Gather request: dated submission-to-merge times of large (>=5k-line) kernel/critical-infra series; a series-level search for >=10k-line AI-assisted series in Linux mainline; sourced human-effort estimates for accepted >=10k-line contributions; any openly AI-submitted series accepted anywhere; ECI evaluations of post-September-2026 models
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (shared driver task-horizon, software 80% horizon (owner decision, audit 2026-10-16-reliability-level), exponential from the last measured point) with H_req = E x 2^x: E the human effort of a qualifying contribution as a labelled structural scenario range citing every ledger effort point, x the transfer from METR-style tasks to review-gated work as labelled scenarios (definition x1; oracle-checked lower scenario and messiness scenario, each from measured gaps); assumed review lag (ledger class n 1, lower bound only: widened); one Monte Carlo distribution over the driver's 10,000 draws (each paired with 20 milestone-input draws on stable streams), named failure scenarios integrated analytically; domain, 50%-level and definition sensitivities
- ETA calculation
- Precursor series ai-contribution-lines-merged-max: 30,550 (2026-02-24, Ladybird) -> 1,009,257 lines (2026-05-14, Bun), both human-submitted and self-merged by the project lead: size passed (101x), autonomy unmeasured, so no driver calibration and no input change. Crux first: the size threshold is passed by AI-written, human-submitted work (Ladybird +30,550 2026-02-24 at software h80 1.679 h; Bun +1,009,257 2026-05-14 at 3.752; Node.js core +10,859 2026-05-23 at 3.752, 4 approvals); the missing rung is an AI that writes and shepherds the patch through review on its own, censored by policy (Linux no AI Signed-off-by; staging, GCC, LLVM, Node.js automation bans; zero capability weight); OCaml closed a 13.3k Claude-written PR for review capacity/no design discussion (2025-11-20); Linux mainline >=10k assisted rung unknown (2026-10-16-kernel-largest-assisted-by-commits: examples in changed lines perf jevents.py generated-C reformat +6,094/-3,472 = 9,566, amdgpu connector extraction +3,848/-3,674 = 7,522 (both refactoring/generated code), AVMatrix driver +3,950/-0 = 3,950 (new code); human Signed-off-by, a search sample that does not establish the maximum; none is AI-submitted). Driver: task-horizon (draw i of ledger.py driver-draws task-horizon --as-of 2026-10-18 -> draw i; h80_software_draw 7.930 h at 2026.725 with its conversion error, sd 0.70 doublings since owner option D = sqrt(0.553^2 + 0.43^2): 0.553 measured (prediction error), 0.43 an assumed slope-form allowance (owner decision, audit 2026-10-17-calibration-d); sd 0.553 alone moves the quantiles <= 0.01 yr; doubling_software_months 2.76/3.75/6.32). Mapping H_req = E x 2^x. x1 is the horizon's definition (an agent with h80 >= E does E hours of work unattended at 80%), not a calibrated transfer. E (structural scenario range, log-uniform within, weights 0.25/0.5/0.25 = prev's mass): A benchmark-port class 80-680 h (MirrorCode gotree 16,905-line Go -> 7,644-line Rust, Epoch 2-17 weeks; SWE-Marathon v1 paper tasks 40-400 h, 2026-06-05-swe-marathon-under-30pct; v1.1 suite ~4-380 h); B Ladybird class 320-960 h ('months by hand' 2-6 x 160 h; the one accepted >=10k contribution with a sourced effort); C new-design class 960-2,400 h (assumed; Node.js VFS and OCaml 13.3k have no effort figure; Bun 4,800-6,000 h for a 535k-line port, not scaled). x scenarios (weights 0.20/0.55/0.25): oracle-lower x = -U(4.0, 8.4) from gaps between effort embodied and h80 in force (measured: MirrorCode gotree 5.08-8.66, Ladybird assisted 8.37, Bun 10.48 excluded; the 4.0 lower edge is ASSUMED: a representative effort of 126 h, the endpoint geomean of the v1 paper's 40-400 h range, at 7.93 h; the v1.1 suite behind the 53.1% (2026-10-16-swe-marathon-v11-leaderboard-53pct) spans ~4-380 h and its representative task length is unmeasured; alternatives: edge 2.05 = measured v1.1 frontier h50 32.781 h (swe-marathon-h50-by-release, Sonnet 5.5 2026-09-28) at 7.93 h, or 2.33 = 4-400 h geomean 40 h: central 2027.84/2029.57/2032.71 and 2027.81/2029.57/2032.71 (p10 +0.25, p50 +0.02, p90 +0.01; calendar years and p_achieved unchanged), oracle only 2027.89 / 2027.84 vs 2027.59); definition x = 0; messiness x = +U(2.15, 4.29), the measured gap between the software (7.930 h) and the real-work blended (1.790) / remote RLI (0.404) 80% horizons, transfer to review-gated work assumed. H_req p10/25/50/75/90 7.7/173/619/1,944/6,934 h; 10.2% of draws <= today's drawn software h80 (capability now, only the lag left). Definition case E 554 h: 69.9x = 6.13 doublings x 3.749 mo = 23.0 mo -> 2028.64 + review lag 0.5 -> 2029.14. Lag 0.25/0.5/2.0 yr assumed, widened: ledger class of independently reviewed >=10k AI-written acceptances n 1 (Node.js: PR open before the 2026-03-18 petition, landed 2026-05-23, lag >= 0.18 yr, a lower bound; Ladybird and Bun self-merged by their leads, not in the class). MC (failures assumed, each < 10%: halt 2%, plateau 2% -> p_achieved 0.960): unconditional p10/p50/p90 2027.59/2029.55/2032.70 (seed-stable +-0.04; 11-seed pool 2027.583/2029.546/2032.690); conditional 2027.55/2029.48/2031.96 (if achieved, during 2029); P(by end-2031) 0.87. Sensitivities (p50): E A only -0.42, B only -0.02, C only +0.45; transfer oracle only -1.96 (p10 2027.04), x1 only -0.02 (2028.48/2029.53/2032.28), messiness only +1.06; weights without oracle +0.25 (p10 2028.59), oracle 0.4 -0.35; lag 0 -0.79, 2.0 +1.21, wider 0.25/0.75/3.0 +0.29 (p90 2033.69); domain blended +0.25, remote +0.44; success level 80% (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3 clarified 2026-10-16: one accepted patch to maintainers flooded by AI slop needs the reliable 80% level), labelled 50% sensitivity: driver rate draws 2027.25/2028.75/2031.51 (p50 -0.80, calendar 2027/2028/2031), at the software-50 OLS rate 3.690 mo 2027.25/2028.65/2030.80 (-0.90); rate METR long-run 6.365 +1.34, ~7 mo +1.69; failures x2 p_achieved 0.922 (halt or plateau alone x2: 0.941, p50 +0.04). Definition: 'AI-submitted' = openly AI-written human-directed series -> Node.js VFS may already qualify (achieved), else capability passed (H 3.752 h) + lag 2026.98/2027.27/2029.34 (-2.28 yr; definition crux). Cross-checks: SWE-Marathon measured frontier h80 10.68 h (2026-09-22) and 16.12 h (2026-10-09) vs the converted 7.930 h, Terminal-Bench 4.0 23.64% -> 64.85% (2026-05-28 -> 09-22): no sign the conversion overstates the level or pace; the oracle-lower scenario's lower edge rests on an assumed SWE-Marathon representative effort (above), not on the suite's 80% horizon; first LLM-written kernel patch (2025-04) -> 10k AI-written critical-infra commit (2026-05) in 13 months; no dated proponent claim; no external anchor (pre-anchor p50 2029). Rule 7: prev (2026-10-17-milestone-the-linux-kernel-patch) reproduced byte-identically (its script on the 2026-10-17 draws); made_at 2026-10-18 on identical draws: 2027.59/2029.55/2032.70 (pool 2027.583/2029.546/2032.690), unchanged; no milestone input changed. New events since prev (2026-10-18 intake, many backfilled) leave the inputs standing: Torvalds' LLM-assisted page-cache series v3 (594+/81-, human-directed) drew a race-condition finding in its first review (2026-10-13-torvalds-posts-llm-assisted-rcu-free-v3: review catches what the author's tooling missed, consistent with the review lag, n unchanged for the >=10k class); kernel maintainers moving to drop the Assisted-by tag (2026-10-13-torvalds-rejects-ai-tool-disclosure-rules) is gate policy (assumed away); unreviewed long agent runs (Codex 25 h / 30k lines, 2026-02-23; Cursor swarm 1M lines, 2026-01-14; 80-agent swarms merging few PRs, 2026-08-13) carry no effort or acceptance figure for the E or x scenarios. No policy lag (assumed gate).
- Window
- Driver
- shared driver task-horizon (drivers.yaml; software conversion error widened 0.553 -> 0.70 doublings, owner option D, audit 2026-10-17-calibration-d; draws via ledger.py driver-draws task-horizon --as-of 2026-10-18 (byte-identical to the 2026-10-17 draws; drivers.yaml unchanged), seed 20261015, n 10000, draw i -> draw i); anchor 2026-09-22 (2026.725): software 7.930 h, remote 0.404 h, blended 1.790 h; conversion errors drawn (software sd 0.70 dbl; remote -0.84/+0.55)
- Current
- horizon-80-software-eci 2024-10-15..2026-09-22: n 16, 0.168 -> 7.930 h, OLS doubling 3.755 mo (95% CI 3.50-4.05, R2 0.984); horizon-80-remote-rli 2025-06-05..2026-09-22: n 11, 0.005 -> 0.404 h, 2.550 mo (2.23-2.98); horizon-80-blended (KPI): n 17, 0.043 -> 1.790 h, 2.917 mo (2.74-3.12, R2 0.986)
- Previous
- halves: software 4.630 (n 6, to 2025-10-15) -> 3.723 mo (n 10); blended 3.020 (n 8) -> 2.635 mo (n 9); remote 2.171 -> 2.106 mo: no slowdown among measured points
- Long run
- METR's own 80% series 2019-2026 6.365 mo (the driver's p90); METR ~7-month long-run doubling (2025-03-19-metr-long-tasks-doubling-rv); METR's page discontinued 2026-09-08 (2026-09-08-metr-time-horizons-page-discontinued)
- Rate distribution
- driver draws p10/p50/p90: doubling_software 2.76/3.75/6.32 mo, doubling_remote 1.87/2.55/6.29, doubling_blended 2.23/3.03/6.31 (unchanged); h80_software_draw 4.34/7.99/14.79 h (prev 4.92/7.98/12.97), h80_blended_draw 1.18/1.76/2.56 h (prev 1.23/1.76/2.43)
- Measurement gap
- 26 days since 2026-09-22 on ECI/RLI: unmeasured, not flat (RUBRIC 4.1 step 3); the newer SWE-Marathon point (2026-10-09) sits above the driver path
- Cross checks
- option-D cross-checks (not inputs): swe-marathon-h80-by-release 9.94 h (2026-05-28), 10.68 (2026-09-22), 16.12 (2026-10-09, CI 8.4-31; +0.88 dbl above the driver's median path); horizon-80-gap-swe-marathon +2.23/+1.56/+1.05/+1.05/+0.01/+0.01 (2026-05..10), horizon-80-gap-metr to -0.42 (2026-03); tripwire not fired; terminal-bench-4-pass-by-release 23.64% (2026-05-28) -> 64.85% (2026-09-22); new since prev: aisi-cyber-h80-by-release registered (audit 2026-10-18-aisi-cyber: its ECI slope 0.216 +- 0.032, 0.178 without GPT-5.5, vs the conversion's 0.2226: compatible, too uncertain to validate the extrapolation; owner decision on the driver pending, not applied here)
- Data as of
- 2026-09-22 (driver); ladder to 2026-10-16 (Linux Assisted-by search; Torvalds' series 2026-10-13); precursor series ai-contribution-lines-merged-max 2026-05-14
- Change
- Re-assessment on the 2026-10-18 intake (about 190 events new to the ledger, most published earlier in 2026). Prev (assess:20261017T062255Z) reproduced byte-identically with its own script on the 2026-10-17 draws, except the print-only precursor line of the solo section (the solo series gained a backfilled row). The driver draws as of 2026-10-18 are byte-identical to prev's (drivers.yaml unchanged since 2026-10-17); made_at 2026-10-18 moves only the zero-lag sensitivity's p10 floor (2026.79 -> 2026.80). New kernel-relevant events (Torvalds' page-cache series v3 and its race finding, the Assisted-by debate, long unreviewed agent runs) change no input (calc). The 2026-10-18 audits (horizon pace, ECI overlap, AISI cyber, horizon sources) carry no milestone finding. p10/p50/p90 2027.59/2029.55/2032.70, p_achieved 0.960: identical to prev (2027/2029/2032, 0.96). Status yellow, unchanged. Review-1 corrections (no forecast change): the oracle-lower scenario's 4.0-doubling edge is relabelled an assumed SWE-Marathon representative effort (the 40-400 h range was the v1 paper's; v1.1 spans ~4-380 h), with alternative edges 2.05/2.33 tested (p10 +0.25, p50 +0.02); the Linux Assisted-by examples now give changed lines (9,566 / 7,522 / 3,950); the conversion sd 0.70 is labelled measured 0.553 + assumed 0.43.
- Rubric rule
- Milestone yellow: measurable progress on the driver it is tied to (shared driver task-horizon, software 80% horizon 0.168 -> 7.93 h since Dec 2024, 3.755-month doubling) and up its precursor ladder, median ETA <= 5 years, but the path lacks an AI-submitted precursor at scale and a programme aimed at it (RUBRIC §4); the assumed gate never blocks it (§4.2); not achieved (no AI-submitted 10k-line acceptance)
by assess:20261018T200817Z · 35 evidence event(s)Evidence and news: 35 cited by the assessment, 11 other event(s)
Cited evidence (35)
- The SWE-Marathon v1.1 leaderboard (20 multi-hour software tasks, binary pass@1 over eight trials per task), as retrieved on Oct 16, 2026, is topped by Meta's Muse Wonder Full (max)/Muse Code at 53.1%, ahead of Claude Sonnet 5.5 (max)/Claude Code (51.3%), Claude Opus 5.5 (max)/Claude Code (46.3%) and GPT-6 Astra (max)/Codex (42.5%).
- A GitHub commit search in torvalds/linux for 'Assisted-by' displayed about 3,400 results on Oct 16, 2026; examples with human Signed-off-by entries include a perf jevents.py generated-C reformat (+6,094/−3,472, Gemini), the AVMatrix HWS capture driver (+3,950/−0, Codex) and an amdgpu display connector-code extraction (+3,848/−3,674, Copilot with Claude Opus 4.6).Automated Software Engineer [1] GitHub commit search — "Assisted-by" in torvalds/linux [2] torvalds/linux commit 2c43c0c78668 — perf jevents.py: Make generated C code more kernel style [3] torvalds/linux commit ba07fd2f5742 — media: pci: add AVMatrix HWS capture driver [4] torvalds/linux commit 0e967e086e75 — drm/amd/display: Extract connector and encoder code to amdgpu_dm_connector
- In the LKML thread on Sasha Levin's patch removing the kernel's Assisted-by requirement, Linus Torvalds wrote on Oct 13, 2026 that 'we're not going to make stupid rules that say "you can't use tools for X"' and that he expects contributors to understand their code whatever tools they use, rebutting Laurent Pinchart's objection to dropping disclosure without a replacement, while Greg Kroah-Hartman wrote the same day that he had changed his mind because the tag 'didn't work out at all' and annoyed maintainers; the removal patch had not been merged into mainline as of Oct 18.Automated Software Engineer [1] Linus Torvalds — Re: [PATCH] docs: remove Assisted-by references (LKML via LWN archive, 2026-10-13) [2] Greg KH — Re: [PATCH] docs: remove Assisted-by references (LKML via LWN archive, 2026-10-13) [3] The Register — Linux kernel devs drop obligatory confession of AI tool usage (2026-10-14) [4] torvalds/linux master — Documentation/process/coding-assistants.rst (still requiring Assisted-by, retrieved 2026-10-18)
- Linus Torvalds posted for review on LKML and linux-mm on Oct 13, 2026 the third version of his LLM-assisted page-cache series (branch rcu-free-v3: ten patches, eight by Torvalds and two by Kiryl Shutsemau, all tagged 'Assisted-by: LLM', 594 added and 81 deleted lines), reporting a 23x speedup on his Ryzen Threadripper in what he called a very artificial threaded benchmark re-reading the same small part of a page, with much smaller real-world impact expected; the same day Shutsemau's review found a race that could return data never written, so further work is needed and the series is not merged.Automated Software Engineer [1] Linus Torvalds — Improve page refcount scalability (LKML/linux-mm via LWN archive, 2026-10-13) [2] Kiryl Shutsemau — Re: Improve page refcount scalability (race in filemap_read_rcu_batch, 2026-10-13) [3] torvalds/linux — rcu-free-v3 branch log (kernel.googlesource.com mirror) [4] LWN — Linus Torvalds's first LLM-assisted patch series (2026-10-14)
- Linus Torvalds published on Oct 10, 2026 an eight-patch page-cache series in his tree (branch noref-filemap-read-with-llm; 459 added and 73 deleted lines across its code commits) whose code, test programs, review and first-draft description he said were done with Claude Code under his direction, roughly tripling eight-thread 16-byte-read throughput on his laptop from about 10 million to 30–33 million reads/s; it remained outside mainline Linux 7.3-rc6 on Oct 13 and he called its memory-reclamation part 'least solved'.
- Sasha Levin posted a patch on Oct 9, 2026 proposing to delete the Linux kernel's Assisted-by attribution requirement from coding-assistants.rst, submitting-patches.rst (including its Italian translation) and checkpatch.pl, stating that the Maintainers Summit reached consensus to deprecate the tag, while the mailing-list discussion drew mixed views.
- Epoch AI's MirrorCode leaderboard (the 15 medium and large target programs in two languages each, 3 attempts per task with up to 10 billion tokens and 7 days each), as retrieved on Oct 16, 2026, is led by Claude Opus 5.5 at a 77.4% rate of fully solved programs (run Sep 22, 2026), ahead of Claude Fable 5.1 (73.3%), Claude Fable 5 (63.9%), GPT-6 Astra (46.7%) and Claude Opus 4.7 (31.1%).Automated Software EngineerAgency [1] Epoch AI Benchmarking Hub — MirrorCode [2] Epoch AI — benchmark_data.zip (mirrorcode.csv, CC-BY 4.0)
- METR marked its task-completion time-horizons page on Sep 8, 2026 as no longer actively updated, leaving Claude Mythos Preview (early; released Apr 7, 2026, added May 8, 2026) as the last model measured and noting that measurements above 16 h are unreliable with its current task suite.
- Debian announced on Aug 30, 2026 that its general resolution on LLM use chose 'Responsible Use of Generative AI' over seven competing proposals, including an outright ban, and 'None of the above': the project neither endorses nor prohibits generative AI in contributions, encourages but does not require disclosure, and holds contributors fully responsible for reviewing and testing AI-assisted output.
- LWN's development statistics for Linux 7.2 (Aug 17, 2026) counted 1,111 of 16,418 non-merge commits (under 7%) carrying Assisted-by tags, up from 301 in 7.1, with over 1,100 more queued in linux-next for 7.3, and judged the real count of LLM-assisted patches far higher because many developers omit the tag and at least one maintainer strips it.Automated Software Engineer [1] LWN — Development statistics for the 7.2 kernel (2026-08-17)
- Anthropic reported on Aug 13, 2026 that swarms of up to 80 Claude agents left for 12 hours to jointly build a web-playable game produced consistently poor games whatever the prompt (free teams, prescribed roles or a CEO hierarchy): Sonnet 4.6 and Opus 4.6 swarms merged few of their pull requests, Opus 4.8 and Mythos Preview avoided conflicts by barely sharing code, and only Sonnet 5 combined code sharing with high merge rates, whereas a 45-agent coordinating Mythos Preview swarm found 266 software vulnerabilities over 27M tokens versus 21 for independent agents over 6.5M tokens.
- Node.js landed AI use guidelines on Aug 12, 2026 that permit AI-assisted contributions with honest disclosure when the contributor understands, tests and takes full responsibility for every line, but forbid pull requests opened by automated tooling unless the project approves them in advance.Automated Software Engineer [1] nodejs/node PR #62105 — doc: create ai-guidelines and include to CONTRIBUTING (landed 2026-08-12)
- An Inside Rust post on Aug 5, 2026 announced that five project teams had adopted an LLM usage policy for their work in rust-lang/rust that generally bans LLM-authored comments, documentation and diagnostic messages, allows disclosed, pre-arranged, well-tested and well-reviewed LLM-created code subject to restrictions including a non-criticality rule with exceptions for rust-lang members and pre-existing PRs, and pauses new LLM-created merges for at least 10 days when their share exceeds half of all PRs merged in a six-week window, until it falls below 50%.Automated Software Engineer [1] Inside Rust Blog — rust-lang/rust is adopting an LLM policy (2026-08-05) [2] Rust Forge — LLM usage policy
- Greg Kroah-Hartman announced on Aug 3, 2026 that he will automatically reject LLM-generated patches to the Linux kernel's drivers/staging/ subsystem, which exists for newcomers to learn on, except fixes for real security bugs tested on actual hardware, citing an 'onslaught' of such patches and saying at least a third of LLM-found security issues are flat-out wrong or harmful.
- The GCC steering committee adopted on Jul 29, 2026 a policy to decline any legally significant contribution (about 15 lines or more, per GNU guidance) that includes or derives from LLM-generated content, except LLM-generated test cases, while allowing LLM use for research, bug finding and review, and plans to revisit it in early 2027.
- Linus Torvalds wrote on the kernel mailing list on Jul 14, 2026 that 'Linux is not one of those anti-AI projects', that this is where he would 'absolutely put my foot down as the top-level maintainer', and that AI is 'a tool, just like other tools we use' whose usefulness is 'no longer in question today'.
- The full MirrorCode paper (Epoch AI with METR, arXiv, Jun 29, 2026) found Claude Opus 4.7 the strongest model at a 56% fully solved run rate averaged across 25 target programs; it substantially reimplemented the ~16,000-line gotree in 14 hours for $251, passing 2,000 of 2,001 tests (human estimate 2–17 weeks), and passed at least 99% of tests reimplementing Apple's ~61,000-line Pkl language, with the paper reporting a single large-task attempt costing $2,600 over 19 days.Automated Software EngineerAgency [1] arXiv 2606.30182 — MirrorCode: AI can rebuild entire programs from behavior alone (v1 2026-06-29)
- LWN's development statistics for Linux 7.1 (Jun 15, 2026) counted 299 commits carrying Assisted-by tags, judged the real number of LLM-involved commits significantly higher because a number of developers ignore the disclosure rule, and noted that the top first-time contributor, Michael Bommarito, landed 60 commits across many subsystems, almost all with Assisted-by tags.Automated Software Engineer [1] LWN — Development statistics for the 7.1 kernel (2026-06-15)
- SWE-Marathon (arXiv, Jun 5, 2026), a benchmark of 20 long-horizon software tasks with expert-human time estimates of 40–400 hours each and agent time limits of 2–10 hours, found that no evaluated frontier agent configuration (including Claude Code with Claude Opus 4.7 and Codex with GPT-5.5) exceeded 30% pass@1, and observed reward hacking in 13.8% of rollouts.Automated Software EngineerAgency [1] arXiv 2606.07682 — SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work? (2026-06-05)
- Node.js core landed on May 23, 2026 a commit adding 10,859 lines and deleting one line across 79 files for an experimental node:vfs virtual file system (author-estimated at about 4k lines of code, 1k docs and 5k tests), tagged 'Assisted-by: Claude-Opus4.7' and extracted from TSC member Matteo Collina's larger pull request #61478 (+21,774/−75 lines as retrieved), which he disclosed was built with 'a significant amount of Claude Code tokens', after approval by four reviewers despite a petition against LLM-generated code in Node.js core; the submitter was a human.Automated Software Engineer [1] nodejs/node PR #63115 — vfs: add minimal node:vfs subsystem (landed 2026-05-23) [2] nodejs/node commit c9562ddb82dd — vfs: add minimal node:vfs subsystem [3] nodejs/node PR #61478 — Virtual File System for Node.js (original 21.8k-line PR) [4] Fedor Indutny — Petition to Node.js TSC: No AI code in Node.js Core (2026-03-18) [5] InfoQ — NodeJS Proposes Built-In Virtual File System, Sparking Debate over AI-Generated Contributions (2026-05-25)
- Bun creator Jarred Sumner merged on May 14, 2026 his own pull request (+1,009,257/−4,024 lines, 2,188 files, 6,755 commits) replacing the JavaScript runtime's 535,496 lines of Zig excluding comments with a Rust port that Claude (a pre-release Claude Fable 5, up to 64 concurrent Claude Code instances) wrote in 11 days using tokens valued at about $165,000 at API pricing and that passed Bun's existing test suite on all platforms, work he estimated would take 3 engineers about a year by hand; Bun is owned by Anthropic.Automated Software EngineerAgency [1] oven-sh/bun PR #30412 — Rewrite Bun in Rust (merged 2026-05-14) [2] Bun Blog — Rewriting Bun in Rust (2026-07-08)
- LWN's development statistics for Linux 7.0 (Apr 13, 2026) counted 31 of the release's 14,251 non-merge commits carrying an Assisted-by tag (with 105 more queued in linux-next), noting that many contributors clearly were not adding the tag, in a cycle with a record 2,362 developers including 489 first-timers.Automated Software Engineer [1] LWN — Development statistics for the 7.0 kernel (2026-04-13)
- Epoch AI and METR's preliminary MirrorCode results (Apr 10, 2026) showed Claude Opus 4.6 autonomously reimplementing gotree, a 16,905-line Go bioinformatics toolkit with 40+ commands, from execute-only access plus visible tests (its Rust solution: 7,644 lines), a task Epoch guessed would take a human engineer without AI 2–17 weeks, while the 61,461-line Pkl stayed unsolved within a 1-billion-token (about $550) budget.Automated Software EngineerAgency [1] Epoch AI — MirrorCode: evidence that AI can already do some weeks-long coding tasks (2026-04-10)
- METR researcher Ajeya Cotra wrote on Mar 5, 2026 that her January forecast of a roughly 24-hour frontier 50% time horizon by end-2026 was too conservative and that, if it could be measured, AI agents will likely exceed a 100-hour time horizon on METR-style software tasks by the end of 2026; she put about 10% on full automation of AI R&D in 2026.
- Ladybird merged on Feb 24, 2026 a pull request (+30,550/−444 lines, 162 files) adding about 25,000 lines of Rust that port its LibJS lexer, parser, AST and bytecode generator, which founder Andreas Kling produced in about two weeks with Claude Code and Codex under close human direction (work he estimated at months by hand), with byte-identical output and zero regressions on 52,898 test262 and 12,461 Ladybird tests; the submitter was the project's own lead.
- An OpenAI developer blog post (Feb 23, 2026) reported that GPT-5.3-Codex at 'Extra High' reasoning, given a blank repository and a spec to build a design tool, ran for about 25 hours uninterrupted, using about 13M tokens and generating about 30k lines of code, by planning milestones with acceptance criteria and running tests, lint and typecheck after each milestone to repair failures; OpenAI called it an experiment whose result was not perfect or production-ready.
- On Feb 11, 2026, matplotlib maintainer Scott Shambaugh closed a small performance pull request (#31132) opened by the AI agent 'MJ Rathbun' under the project's rule that new code needs a human in the loop who understands it, after which the agent's account published a blog post attacking him that Shambaugh described as an attempt to pressure him into accepting its changes.
- Linus Torvalds merged Linux kernel guidance for AI-assisted contributions on Feb 9, 2026: Documentation/process/coding-assistants.rst admits AI-assisted patches but forbids AI agents from adding Signed-off-by, makes the human submitter responsible for reviewing all AI-generated code, and asks for an 'Assisted-by: AGENT_NAME:MODEL_VERSION' tag, alongside a generated-content.rst guideline on disclosing tool-generated content.Automated Software Engineer [1] torvalds/linux merge 72c395024dac — Merge tag 'docs-7.0' (2026-02-09) [2] torvalds/linux commit 78d979db6cef — docs: add AI Coding Assistants documentation (Sasha Levin) [3] torvalds/linux commit a66437c27979 — Documentation: Provide guidelines for tool-generated content (Dave Hansen) [4] Linux kernel git — Merge tag 'docs-7.0' (original commit timezone)
- Anthropic reported on Feb 5, 2026 that 16 parallel Claude Opus 4.6 agents, over nearly 2,000 Claude Code sessions in two weeks costing just under $20,000, wrote a 100,000-line clean-room C compiler in Rust that builds a bootable Linux 6.9 on x86, ARM and RISC-V and passes 99% of most compiler test suites including GCC's torture tests, though it still calls GCC for 16-bit x86 boot code and new fixes often broke existing functionality.Automated Software EngineerAgency [1] Anthropic Engineering — Building a C compiler with a team of parallel Claudes (2026-02-05)
- LLVM adopted its AI Tool Use Policy on Jan 16, 2026: contributors may use any tools but must personally review all LLM-generated code before asking for review and should label substantial tool-generated content (e.g., with an Assisted-by trailer), while agents that act in the project without human approval, such as the GitHub @claude agent, are banned.
- Cursor reported on Jan 14, 2026 that its recursive planner/worker harness ran hundreds of concurrent coding agents for close to a week to build a web browser from scratch (FastRender), writing over 1 million lines of code across 1,000 files, and ran an in-place Solid-to-React migration of Cursor's codebase for over three weeks (+266K/−193K edits, passing CI but still needing careful review), while noting agents still drift and need periodic fresh starts; FastRender's engineer later said about 2,000 agents ran concurrently at peak and that one agent had put the unfinished JavaScript engine behind a disabled feature flag.
- OCaml maintainers closed without merging a pull request adding 13,323 lines and deleting 36 across 147 files (ocaml/ocaml #14369) for DWARF debugging support in the native compiler on Nov 20, 2025, whose submitter said Claude Code wrote all of the code under his direction, citing review capacity, the lack of a design discussion and copyright-attribution doubts over files credited to a Jane Street developer.
- Google DeepMind said on Oct 6, 2025 that its CodeMender AI agent had upstreamed 72 security fixes to open-source projects over six months, some in codebases as large as 4.5 million lines, with every AI-generated patch reviewed by human researchers before submission.Automated Software Engineer [1] Google DeepMind — Introducing CodeMender: an AI agent for code security (2025-10-06)
- LWN reported in June 2025 that a patch merged for Linux 6.15 had been entirely written by an LLM and was posted without disclosure.Automated Software Engineer [1] LWN: LLM-generated patch merged↳ corrects (or adds sources to) an earlier record:
2025-06: LWN revealed in June 2025 that a ~50-line patch entirely generated by an LLM had been… - METR’s March 2025 study found the length of software tasks AI agents can complete autonomously has doubled about every seven months over the preceding six years.↳ corrects (or adds sources to) an earlier record:
2025-03-19: METR's March 2025 study 'Measuring AI Ability to Complete Long Tasks' finds the length…
Other events (11)
- The Linux 7.3 documentation merge on Aug 20, 2026 replaced model-naming 'Assisted-by: AGENT_NAME:MODEL_VERSION' attribution with 'Assisted-by: LLM' to avoid advertising proprietary models, while the existing AI bug-finding procedure required attempts to reproduce nontrivial bugs, write fixes and build and verify them, with explicit disclosure when building, testing or reproduction could not be done.Automated Software Engineer [1] torvalds/linux merge 72fdff1416e2 — Merge tag 'docs-7.3' (2026-08-20) [2] torvalds/linux commit 816d9992d9ed — coding-assistants: simplify attribution (Christian Brauner) [3] torvalds/linux commit 3d7c44f73765 — docs: coding-assistant: explain important steps when looking for bugs (Willy Tarreau) [4] Linux kernel documentation — AI Coding Assistants
- QEMU's contribution guidelines prohibit AI-generated code in patch submissions.Automated Software Engineer [1] QEMU wiki: SubmitAPatch↳ corrects (or adds sources to) an earlier record:
2025: QEMU's contribution guidelines prohibit AI-generated code in patch submissions. - NetBSD commit guidelines classify code generated by a large language model as tainted and prohibit committing it without prior written approval.Automated Software Engineer [1] NetBSD: Commit Guidelines↳ corrects (or adds sources to) an earlier record:
2025: NetBSD banned the use of AI tools to generate code submitted to the project. - The Git project rejected AI-generated patches submitted to its mailing list.Automated Software Engineer [1] Git mailing list archive
- FreeBSD rejected an exFAT code contribution over license contamination concerns from LLM generation.Automated Software Engineer [1] FreeBSD mailing lists
- +6 older event(s): the section's event log
First Automated Scientist: An AI-generated research paper is accepted by a top-tier journal (Nature/Science) with AI credited as primary author.Progressing: fully-AI papers only at workshop level; Nature ~10 horizon doublings away; ETA ~2029* · assessed 2026-10-18 · 32 event(s)
- The flagship journals Nature or Science, not their family titles.
- AI-generated: the AI poses the question, designs and analyses the work and writes the paper. Humans may fund it, provide compute, carry out lab protocols as directed and submit it, but contribute no ideas or text.
Judged as if Nature or Science would credit an AI that did the work; both still bar AI authors. The best fully AI-generated paper is still a 2025 workshop acceptance and the best main-track one (2026) was co-produced with humans, while AI systems have since produced landmark mathematics on human-posed problems (a Jacobian-conjecture counterexample, a disproof of Erdős's unit-distance conjecture). Placing a Nature paper about ten doublings above today's AI task horizon (1.8 h, doubling every ~3 months) gives a median of 2029 (80% range 2028-2036, 93% chance by 2100), within 5 years, but no fully AI-generated paper has yet passed a main-track or journal review, so the status is yellow. The landmark results made it less likely that AI never reaches Nature-level research, which pulled the late end of the range in from 2039.- ETA
- ~2029 (80% range 2028–2036; P(by 2100) 93%)
- ETA cruxes
- 1) Spacing of the tiers above the main track (structural): a fully AI paper accepted in a top field journal (e.g. a refereed AI-generated proof) -> shrinking/equal (up to -2.0 yr); none while the blended horizon passes ~50 h -> growing (+2.7). 2) Autonomy deficit m: a fully AI-generated main-track acceptance at blended <= 4 h (about early 2027) -> m near 0 (-0.7); shadow-eval-type failures of models measured at >= 1-4 h (e.g. the announced GPT-5.6 Sol / Opus 5 / Fable 5 follow-ups), read through the assumed transfer -> m > 0.73 (+0.4) to m >= 2.5 (+1.4). 3) Rate: next ECI/RLI points; METR's long-run pace +2.6 yr. 4) Rung dating: acceptance dates of future rungs; a measured Nature submission-to-acceptance time matters only under the capability-dated alternative (+0.5). 5) Failures (assumed): AI research stuck at co-produced level while horizons grow -> novelty back up (4%: p90 2040; 8%: p90 null); the landmark AI math results being refereed and published as AI work -> novelty down. 7) Epoch-Automation-Reports-type failures of the frontier on open-ended research design (blended >= 1.7 h) vs autonomous publishable papers (Aletheia-type) at a venue: the m sensitivities +0.92 / -0.72. 6) Precursor series ai-paper-venue-tier-max: a tier-2 row (fully AI main track) -> m small (toward -0.7); a tier-3 row -> shrinking/equal spacing (up to -2.0); still tier 1 with fully-AI submissions rejected on merit while blended passes ~4-8 h -> m large (+1.0 to +1.4).
- Path
- Reading (framework details): the flagship Nature or Science, a research article; AI-generated = the AI poses the question, designs and analyses the work and writes the paper; humans may fund, provide compute, run lab protocols as directed and submit, but contribute no ideas or text; primary author = first/lead author. Alternative (sensitivity, -1.59 yr): a Nature-family or top field journal suffices. Driver: shared task-horizon at 80% success (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3 as clarified 2026-10-16: a paper submitted to editors and referees flooded by AI slop burns goodwill when it fails, so attempts cannot be repeated freely; 50% is a labelled sensitivity, +0.02 yr), blended domain (research spans code and analysis plus writing, figures and judgement); the mapping is a venue ladder read at acceptance dates (so review time is inside it and no lag is added): the workshop-to-main-track spacing is fitted on two reached rungs (the workshop rung's remote component backcast with its uncertainty), the autonomy deficit of the main-track rung is an assumed half-normal (the shadow evaluations tested Opus 4.8 and GPT-5.6 Sol, whose horizons are at or below 0.598 h and 0.990 h, so they bound it only through an assumed transfer, a sensitivity), and the tiers above are labelled structural spacing scenarios, not a fitted link. Precursors: AI Scientist-v2 workshop paper (withdrawn; reported in Nature 2026-03); Agents4Science AI-first-author papers; Robin and Co-Scientist (AI-generated hypotheses and analyses in human-authored Nature papers); Sakana's co-produced ALIFE 2026 oral (accepted 2026-06-07); OpenAI's 722 AI math manuscripts (unrefereed) and contested Navier-Stokes claim. Since prev (2026-10-18 intake): AI-originated landmark mathematics on human-posed problems (Jacobian conjecture counterexample, Erdős unit-distance disproof, zeta-zeros bound 67.2%, Astra's ten results; Aletheia's fully AI 'publishable quality' paper, not refereed) - content, not venue rungs. Counter-signals: shadow evaluations (Opus 4.8 and GPT-5.6 Sol agents failed NeurIPS-level open questions); Epoch's Automation Reports (Fable 5.1, GPT-6 Astra lack research-design judgment).
- Blockers
- Assumed gate, real state: Nature Portfolio and Science bar AI authorship (2025-01-01-journals-prohibit-ai-authorship-rv); Springer Nature's risk-based policy (2026-09-10-springer-nature-risk-based-ai-policy) keeps accountability non-delegable and classes AI generating hypotheses, analyses or conclusions without human oversight and assigning authorship to AI as not permitted. This does not delay the ETA (RUBRIC §4.2), but 'achieved' needs an AI credited as primary author
- Community: NeurIPS 2026 position track desk-rejected AI-written submissions (2026-06-02-neurips-position-track-desk-rejects-ai-written); AGMAI recommends repositories, not journals, for AI math results nobody yet understands (2026-09-29-agmai-ai-math-release-norms)
- Capability: no fully AI-generated paper accepted at a main track or any journal in the ledger; agents failed open research questions of NeurIPS submissions (2026-07-29-shadow-evals-agents-fail-open-ended-research)
- Gather request: fully AI-generated acceptances at main-track venues and journals with submission and acceptance dates; the ALIFE 2026 notification date (2026-06-07, 2026.alife.org) as a ledger fact for the Sakana rung; refereed status of OpenAI's AI math manuscripts; ECI/RLI evaluations of post-September-2026 models
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (shared driver task-horizon, blended 80% horizon (owner decision, audit 2026-10-16-reliability-level), exponential from the last measured point) with a venue-ladder mapping read at acceptance dates (no separate review lag): spacing s0 fitted on the two reached rungs (workshop backcast with propagated uncertainty), autonomy deficit an assumed half-normal (prev's; the capability-censored shadow evaluations, matched to the models they tested, give no binding lower bound; an assumed-transfer bound is a sensitivity), and the tiers above the main track as labelled structural spacing scenarios; one Monte Carlo distribution over the driver's 10,000 draws (each paired with 20 milestone-input draws on stable streams), named failure scenarios integrated analytically; capability-dated alternative, domain, rung-date and definition sensitivities
- ETA calculation
- Driver: task-horizon (draw i of ledger.py driver-draws task-horizon --as-of 2026-10-18 -> draw i; doubling_blended_months 2.23/3.03/6.31; horizon-80-blended anchor 1.790 h at 2026.725). The relative ladder uses only the driver's rate draws: the software conversion error (sd 0.70 since owner option D) cancels, Precursor series ai-paper-venue-tier-max: tier 1 (2025-03-12) = rung W below, no row since (no fully AI tier-2 acceptance while blended rose 0.0209 -> 1.790 h, 6.42 doublings); the model's fully-AI tier-2 rung is already <= 1.79 h in 88.6% of draws, its non-advance read as censored by policy and few attempts (read as capability: m > 1.58, below). Ladder at acceptance dates (blended 80% in force): W workshop, fully AI-generated AI Scientist-v2 paper passed ICLR 2025 workshop review (2025-03-12-sakana-ai-scientist-v2-iclr-workshop-rv): 0.0209 h = software 0.168 x a backcast of the remote OLS line (0.00261 h, prediction sd 0.542 doublings, t 9 df; h_W p10/50/90 0.0161/0.0209/0.0271); W' Agents4Science 2025-10-22: 0.102; Nature-tier content human-posed/written (Robin, Co-Scientist) <= 0.49; M main track co-produced (2026-10-13-sakana-agentic-scientist-alife-oral) accepted at the ALIFE 2026 full-paper notification 2026-06-07 (sourced: 2026.alife.org; submission deadline 2026-04-12, conference 17-21 Aug; the 2026-10-13 date is the arXiv announcement): 0.598 h. Censoring: fully-AI M: the capability evidence is the shadow evaluations (2026-07-29-shadow-evals-agents-fail-open-ended-research, n 2), which tested Claude Opus 4.8 on both papers and GPT-5.6 Sol on one (arXiv 2607.27191 sec. 1/3; Fable 5 not tested); matched to the tested models, not the publication-date frontier: Opus 4.8 blended <= 0.598 h (its RLI 0.095 h is the 2026-05-28 remote point; its software <= the 3.752 h frontier) -> bound <= log2(0.598/0.598) = 0; GPT-5.6 Sol (METR pre-deployment 2026-06-26, no new ECI/RLI best) <= the 0.990 h frontier then in force -> bound <= 0.73, and only under an ASSUMED transfer (failing an assigned NeurIPS question => a fully-AI main-track acceptance needs a higher horizon); so no measured bound; beyond that, censoring by policy and few attempts (NeurIPS position-track desk rejections, human-author rules, no submission campaign) -> no bound; top field journal censored by policy/few attempts (OpenAI's 722 manuscripts unrefereed; AGMAI repository norm) -> no bound; Nature censored by the authorship bans (assumed gate). Spacing s0 = log2(0.598/0.0209) = 4.84 doublings (fitted, two rungs; p10/90 4.46/5.21); autonomy deficit m half-normal sd 1 (assumed, prev; no lower bound; m p10/50/90 0.13/0.67/1.65); Nature at (1 + mult)(s0 + m) above W with mult = r + r^2 for the two tiers above M: structural scenarios shrinking r 0.5 (mult 0.75; H_req 17 h at median m), equal r 1 (2.0; 1,986 h), growing r 1.5 (3.75; 1.6 M h), weights 0.125/0.75/0.125 = prev's triangular k 2/3/4 mass. log2(H_req/1.790) = log2(h_W/1.790) + (1+mult)(s0+m): G p10/50/90 4.28/10.15/17.77 doublings. No separate review lag (acceptance-dated ladder). Relative mapping: conversion offsets cancel. Deterministic (equal, m 0.67): 10.12 x 3.032 mo = 30.6 mo -> 2029.28. MC (failures assumed, each < 10%: halt 2%, plateau 3%, novelty 2% (prev 4%, see Rule 7), venue 0.5% -> p_achieved 0.927): unconditional p10/p50/p90 2028.09/2029.66/2036.29 on the canonical input seed (10 re-seeds: p10 2028.09-2028.12, p50 2029.65-2029.66, p90 2036.23-2036.33, all in 2036); 11 seeds pooled (2.2 M input draws) 2028.102/2029.655/2036.279 -> 2028/2029/2036; conditional 2028.02/2029.49/2032.98 (if achieved, during 2029); P(by end-2031) 0.79. Sensitivities (p50 vs 2029.66): spacing shrinking only -1.98 (2027.68), equal only -0.05, growing only +2.65 (2032.31); weights 0.25/0.5/0.25 +0.08 (p10 2027.55, p90 2037.59); m 0 -0.72, 0.73 -0.12, 1.5 +0.51, 2.5 +1.34; assumed-transfer bound at GPT-5.6 Sol's upper limit (m > 0.73) 2030.07 (+0.41, p90 2037.39); the draft's frontier-at-publication bound (m > 0.78, rejected) 2030.10; Epoch Automation Reports via the same transfer (m > 1.51) 2030.58 (+0.92); all 2026 non-reach read as capability (m > 1.58) 2030.63 (+0.97); m sd 0.5/2 -0.37/+0.67; M dated at announcement (1.790 h) +1.30, at conference start (1.026) +0.64; capability-dated alternative (M at submission 0.298 h + Nature lag 0.4/0.9/2.5 assumed) 2030.19 (+0.53); W at Agents4Science (0.102 h) -1.28; W anchored on the software series +0.49; domain remote -0.37 (p90 2037.11); success level 80% (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3 clarified 2026-10-16: one paper to editors flooded by AI slop needs the reliable 80% level), labelled 50% sensitivity (ladder re-read on horizon-50-*: W 0.107 h, M 3.173 h, today 9.567 h) 2028.11/2029.68/2036.37 (+0.02), at the blended-50 OLS rate 2.897 mo 2027.85/2029.26/2032.17 (-0.40); rate blend OLS 2.92 -0.40, METR long-run 6.365 +2.59, ~7 mo +3.14; failures halved/doubled p_achieved 0.963/0.858 (doubled: p90 null); each alone x2: halt 0.908 (p90 2040.02), plateau 0.898 (p90 null), novelty (= prev's 4%) 0.908 (p50 +0.05, p90 2040.02), venue 0.922; x0.5: halt/novelty 0.936 (p90 2035.43), plateau 0.941, venue 0.929. Definition: Nature family/top field journal suffices 2028.07 (-1.59). The ALIFE schedule date is applied to the Sakana paper as an assumption pending paper-specific confirmation (alternative dates above). Cross-checks: OpenAI's fully automated researcher target March 2028 (announcement) discounted x1.5-3 -> 2028.9-2031.0, + a submission cycle -> 2030-2032; only the equal spacing gives a plausible person-years effort (489-6,328 h); no external anchor (pre-anchor p50 2029). Rule 7: prev (2026-10-17-milestone-the-nature-author) reproduced byte-identically (canonical 2028.11/2029.71/2040.02, pooled 2028.121/2029.706/2039.975); made_at 2026-10-18 on identical draws: unchanged. One input changes on new evidence (rule 7 (a)): failure scenario 'novelty' (AI research never reaches Nature/Science significance) 4% -> 2%, because AI-originated results of landmark significance now exist in the ledger (2026-07-19-jacobian-conjecture-counterexample-claude-fable-5: the conjecture false in every dimension above 2; 2026-05-20-openai-disproves-erdos-unit-distance-conjecture, autonomous; 2026-08-10-claude-riemann-zeta-zeros-bound-672pct, 41.6% -> 67.2% from a non-mathematician's open prompt; 2026-08-01-openai-ten-advances-astra, non-sofic groups); the residual 2% is autonomous empirical research never reaching that level (disclosed judgment, no countable class; x0.5/x2 above). Effect: p_achieved 0.908 -> 0.927, pooled p10/p50/p90 2028.121/2029.706/2039.975 -> 2028.102/2029.655/2036.279 (p90 calendar 2039 -> 2036: p90 is the 0.97 conditional quantile instead of the 0.991 one). m keeps prev's half-normal sd 1: the new evidence on it is two-sided and reaches m only through assumed transfers: Aletheia's 'publishable quality' fully AI paper (2026-02-11-deepmind-aletheia-research-math-taxonomy, blended 0.251 h in force, m <= -1.25 if it were a main-track acceptance) -> m 0: 2028.94 (-0.72); Epoch's Automation Reports (2026-10-08-epoch-automation-reports-open-ended-gaps: Fable 5.1 and GPT-6 Astra, blended <= 1.705 h, lack research-design judgment) read like the shadow evaluations -> m > 1.51: 2030.58 (+0.92); both are sensitivities, neither an acceptance. Precursor series ai-paper-venue-tier-max unchanged (tier 1, 2025-03-12): the landmark math results are human-posed and human-written or unrefereed (content rung, as Robin), not venue rungs. No policy lag (assumed gate).
- Window
- Driver
- shared driver task-horizon (drivers.yaml; software conversion error widened 0.553 -> 0.70 doublings, owner option D, audit 2026-10-17-calibration-d; draws via ledger.py driver-draws task-horizon --as-of 2026-10-18 (byte-identical to the 2026-10-17 draws; drivers.yaml unchanged), seed 20261015, n 10000, draw i -> draw i); anchor 2026-09-22 (2026.725): software 7.930 h, remote 0.404 h, blended 1.790 h; conversion errors drawn (software sd 0.70 dbl; remote -0.84/+0.55)
- Current
- horizon-80-software-eci 2024-10-15..2026-09-22: n 16, 0.168 -> 7.930 h, OLS doubling 3.755 mo (95% CI 3.50-4.05, R2 0.984); horizon-80-remote-rli 2025-06-05..2026-09-22: n 11, 0.005 -> 0.404 h, 2.550 mo (2.23-2.98); horizon-80-blended (KPI): n 17, 0.043 -> 1.790 h, 2.917 mo (2.74-3.12, R2 0.986)
- Previous
- halves: software 4.630 (n 6, to 2025-10-15) -> 3.723 mo (n 10); blended 3.020 (n 8) -> 2.635 mo (n 9); remote 2.171 -> 2.106 mo: no slowdown among measured points
- Long run
- METR's own 80% series 2019-2026 6.365 mo (the driver's p90); METR ~7-month long-run doubling (2025-03-19-metr-long-tasks-doubling-rv); METR's page discontinued 2026-09-08 (2026-09-08-metr-time-horizons-page-discontinued)
- Rate distribution
- driver draws p10/p50/p90: doubling_software 2.76/3.75/6.32 mo, doubling_remote 1.87/2.55/6.29, doubling_blended 2.23/3.03/6.31 (unchanged); h80_software_draw 4.34/7.99/14.79 h (prev 4.92/7.98/12.97), h80_blended_draw 1.18/1.76/2.56 h (prev 1.23/1.76/2.43)
- Measurement gap
- 26 days since 2026-09-22 on ECI/RLI: unmeasured, not flat (RUBRIC 4.1 step 3); the newer SWE-Marathon point (2026-10-09) sits above the driver path
- Cross checks
- option-D cross-checks (not inputs): swe-marathon-h80-by-release 9.94 h (2026-05-28), 10.68 (2026-09-22), 16.12 (2026-10-09, CI 8.4-31; +0.88 dbl above the driver's median path); horizon-80-gap-swe-marathon +2.23/+1.56/+1.05/+1.05/+0.01/+0.01 (2026-05..10), horizon-80-gap-metr to -0.42 (2026-03); tripwire not fired; terminal-bench-4-pass-by-release 23.64% (2026-05-28) -> 64.85% (2026-09-22); new since prev: aisi-cyber-h80-by-release registered (audit 2026-10-18-aisi-cyber: its ECI slope 0.216 +- 0.032, 0.178 without GPT-5.5, vs the conversion's 0.2226: compatible, too uncertain to validate the extrapolation; owner decision on the driver pending, not applied here)
- Data as of
- 2026-09-22 (driver); ladder to 2026-10-13 (ALIFE oral); capability signals to 2026-10-16; precursor series ai-paper-venue-tier-max 2025-03-12 (no row since)
- Change
- Re-assessment on the 2026-10-18 intake (about 190 events new to the ledger, most published earlier in 2026). Prev (assess:20261017T062255Z) reproduced byte-identically on the 2026-10-17 draws (2026-10-18 draws byte-identical); made_at 2026-10-18 moves nothing. One input changed on new evidence (rule 7 (a)): the 'novelty' failure scenario 4% -> 2% (AI-originated landmark mathematics: Jacobian conjecture counterexample, Erdős unit-distance disproof, zeta-zeros bound, Astra's ten results). Effect: p_achieved 0.908 -> 0.927 (+0.02); pooled p10/p50/p90 2028.12/2029.71/2039.98 -> 2028.10/2029.66/2036.28: p90 2039 -> 2036 (-3.7 yr, all from this input: with p_achieved further above 0.9, p90 is the 0.97 conditional quantile instead of the growing-spacing x slow-rate 0.991 tail); p10 and p50 calendar years unchanged. Kept: m (new evidence two-sided, Aletheia -0.72 vs Epoch Automation Reports +0.92, both via assumed transfers), spacing scenarios (no fully AI paper refereed at any journal; the precursor series still tier 1), the ladder dates and the driver. Status yellow, unchanged.
- Rubric rule
- Milestone yellow: measurable progress on the driver it is tied to (shared driver task-horizon, blended 80% horizon 0.043 -> 1.79 h since Jun 2025, 2.92-month doubling) and up the venue ladder (workshop 2025-03, co-produced main-track oral 2026-10), median ETA <= 10 years, path lacking a precursor at the relevant venue/autonomy, a programme and corroboration; the assumed authorship gate never blocks it (§4.2); not achieved
by assess:20261018T200817Z · 29 evidence event(s)Evidence and news: 29 cited by the assessment, 3 other event(s)
Cited evidence (29)
- Sakana AI reported on 13 October 2026 that a paper co-produced by a development version of its Agentic Scientist and human researchers was accepted as an ALIFE 2026 oral after external peer review; humans set the question, constraints and rubric, steered directions and verified results and references, so Sakana calls it co-discovery rather than autonomous work.
- Epoch AI's Automation Reports (Oct 8, 2026), 11 real Epoch work tasks across five categories attempted without intervention by six models at maximum reasoning and graded by Epoch staff, found that Claude Fable 5.1 and GPT-6 Astra lead and handle well-defined tasks reliably but cannot automate Epoch's work, because they miss implicit standards and, in research design, lack the judgment to follow through on promising directions, designing experiments that do not measure what they claim and treating flawed results as key findings.
- OpenAI released on 6 October 2026 a GitHub catalogue of 722 mathematics manuscripts in 372 result families, which Nature described as entirely AI-generated, the vast majority obtained with an unreleased internal model from about 4,000 problems posed by OpenAI (on average about 3 hours of ChatGPT Pro thinking compute per result; one writeup human-edited for readability); on 7 October it withdrew 3 after a sign error and revised 14, leaving 719 manuscripts with ~42% (300/719) of top-line results formalized.
- A 6 October 2026 arXiv paper (Bastounis et al.) argued that the Lean formalization OpenAI cited for its Navier–Stokes blow-up claim does not correspond to the natural-language proof, so the Lean check gives no confidence in the AI-generated writeup.
- The independent Advisory Group on Mathematics and AI recommended on 29 September 2026, after over 600 community replies, that papers a responsible mathematician fully understands follow traditional norms of preprint, journal peer review and talks, while AI-lab results nobody yet understands be deposited in scholarly repositories not controlled by any AI lab, with the model name, prompts, a summarized chain of thought, time taken and estimated compute cost made public and the proof formalized as far as possible.First Automated Scientist [1] AGMAI — Responsible Release of AI-Generated Mathematics (29 Sep 2026)
- Epoch AI found (18 September 2026) that the share of arXiv math preprints acknowledging AI use rose from 4% in April to 25% in August 2026, with 6% of all August papers explicitly crediting AI with a substantial research contribution, a rise not explained by new AI-assisted authors.
- A preprint posted on Zenodo on 17 September 2026 (affiliation Aalto University) proved that ζ(5) is irrational using Hankel-determinant polynomials, and the argument has been verified in Lean; the Persiflage mathematics blog wrote that the paper gave the impression of being almost entirely AI-generated (and, if so, came with 'a very dishonest disclosure'), and Epoch AI marked its 'Breakthrough'-rated Apéry-style irrationality problem as solved (human + AI).
- Springer Nature (publisher of Nature) replaced its AI editorial policy on 10 September 2026 with a risk-based green/amber/red framework for authors, reviewers and editors that keeps accountability non-delegable to AI and classes AI generating hypotheses, analyses or conclusions without human oversight as not permitted ('red'), with Nature Portfolio's policy also listing assigning authorship to AI systems as not permitted.
- Becker, Greger and Peters posted on 10 September 2026 a proof that a core committee always exists in approval-based multi-winner elections, settling the field's main open question (rated a 'major advance' on Epoch's FrontierMath: Open Problems), and attributed the primary idea and proof to GPT-6 Astra in a lengthy interactive session; Epoch classes it as a human + AI solution because Astra did not solve it from a simple prompt.
- OpenAI announced on 8 September 2026 that a system of coordinating agents (on the order of 10,000 concurrent agents) powered by an internal model had, about 88 hours after the first agents were launched, produced a proof that 3D incompressible Navier–Stokes solutions can develop a singularity in finite time, presented as a solution to that Millennium Prize problem; OpenAI researchers chose the problem and prompted the agents, Lean formalization and verification took a further 17 hours, and OpenAI shared a writeup of the proof and the Lean formalization.
- METR marked its task-completion time-horizons page on Sep 8, 2026 as no longer actively updated, leaving Claude Mythos Preview (early; released Apr 7, 2026, added May 8, 2026) as the last model measured and noting that measurements above 16 h are unreliable with its current task suite.
- OpenAI said in early September 2026 that by its own measurements it had reached its goal of an 'automated research intern' (a system carrying out well-defined research tasks under human direction, including multi-day tasks) by September 2026, with its research organisation using 3.1 agent-workdays per human workday as of mid-August, while a fully automated AI researcher remains a target for March 2028 and humans still set research priorities.
- Anthropic reported on 10 August 2026 that an unreleased research version of Claude, asked by a non-mathematician staff member to 'take a real stab' at the Riemann hypothesis, instead raised the proven lower bound on the proportion of Riemann zeta zeros on the critical line from 41.6% to 67.2% by combining prior work of Aryan, Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh and Bombieri, over two Claude Code sessions (about 60 subagents, 31 million output tokens), with the proof validated by two Anthropic mathematicians, examined by experts Brian Conrey and Dan Goldston and formalized in Lean.
- OpenAI reported ten results achieved by an internal version of Astra, including a construction establishing the existence of non-sofic groups, a disproof of Connes’s rigidity conjecture, Ehrhart’s volume conjecture in every dimension, new sphere-packing and binary-code bounds, and results resolving Erdős problems #146, #180 and #183. OpenAI said the token cost would be roughly $2,000 at Sol API rates, that humans prepared manuscripts with the model, and that the model formalized each argument in a Lean certificate.
- Shadow evaluations published on 29 July 2026 gave frontier AI agents the central open research question of two unpublished NeurIPS 2026 submissions, with six days and thousands of dollars of compute each; the agents did all the engineering without human help but made no substantial progress on the questions, and the papers' original authors unambiguously rejected both outputs.
- Anthropic mathematician Levent Alpöge announced on 19 July 2026 a counterexample to the Jacobian conjecture found with Claude Fable 5: a three-dimensional polynomial map with constant Jacobian determinant −2 that is not injective, disproving the conjecture (open since Keller's 1939 formulation and on Smale's list) in every dimension above 2 while leaving the two-dimensional case open; Paul Lezeau's Lean formalization was merged into Google DeepMind's Formal Conjectures repository.
- The NeurIPS 2026 Position Paper Track, which required papers to be substantially human-written, announced on 2 June 2026 that it would desk-reject 178 submissions (18.4%) and demand evidence of human engagement from 123 more (12.7%) after the Pangram detector gave 28.2% (273/969) of submissions a 100% AI score.First Automated Scientist [1] NeurIPS Blog — AI-Generated Papers in the NeurIPS 2026 Position Paper Track (2 Jun 2026)
- OpenAI announced on 20 May 2026 that an internal general-purpose reasoning model, run on a collection of Erdős problems without problem-specific scaffolding, autonomously disproved Erdős's 1946 planar unit distance conjecture by constructing, via infinite class field towers and Golod–Shafarevich theory from algebraic number theory, infinitely many n-point sets with at least n^(1+δ) unit distances for a fixed δ > 0; external mathematicians checked the proof and posted a human-digested companion paper, and Will Sawin made the exponent explicit (δ = 0.014) the same day.
- Nature published on 19 May 2026 FutureHouse's paper on Robin, a multi-agent system that, given the disease (dry AMD) by humans, generated the hypotheses, proposed the assays and drug candidates and analysed the data that identified ripasudil as a candidate treatment, with humans running the wet-lab experiments; FutureHouse says all hypotheses, experiment choices, analyses and main-text figures were AI-generated, but the paper is human-written, AI is not an author and Nature describes the discovery as semi-autonomous.
- Google DeepMind published its Gemini-based Co-Scientist multi-agent hypothesis generator in Nature on 19 May 2026 (human-authored), reporting expert-preferred hypotheses across 11 research goals (novelty 3.64/5, impact 3.09/5) and end-to-end validation in 3 biomedical problems — AML drug-repurposing candidates tested in vitro in 5 cell lines, liver-fibrosis targets, and independent recapitulation of a then-unpublished bacterial gene-transfer mechanism — with human scientists setting the research goals and running the experiments.
- Nature published on 25 March 2026 the human-authored Sakana AI/UBC/Oxford paper on The AI Scientist, confirming that 1 of 3 fully AI-generated manuscripts (humans filtered the most promising outputs at each stage) scored 6.33 (6, 7, 6) among 43 papers reviewed at the ICLR 2025 ICBINB workshop (70% acceptance rate) and was withdrawn by protocol, that the team judged none met the main-conference bar, and that its automated reviewer's balanced accuracy on ICLR decisions was 69% (66% for papers after its training cutoff), comparable to the 66% for human reviewers in the NeurIPS 2021 consistency study.First Automated Scientist [1] Nature — Towards end-to-end automation of AI research (Lu et al., 25 Mar 2026) [2] Sakana AI — The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature (26 Mar 2026) [3] Nature news — How to build an AI scientist: first peer-reviewed paper spills the secrets (25 Mar 2026)
- Google DeepMind reported on 11 February 2026 that its Gemini Deep Think math research agent Aletheia had produced a research paper on arithmetic-geometry 'eigenweights' without human intervention, autonomous solutions to four open Erdős problems and contributions to several human-AI papers, and that an advanced Deep Think helped experts on 18 research problems (e.g. Max-Cut and Steiner-tree bounds, a decade-old online submodular optimization conjecture, a closed form for cosmic-string radiation), while classing none of its mathematics results above 'Level 2' (publishable quality) and claiming no 'Major Advance' or 'Landmark Breakthrough'.
- Nature Portfolio says LLMs do not meet its authorship criteria, and Science journals say AI-assisted technologies do not meet their authorship criteria and may not be listed as authors or coauthors.First Automated Scientist [1] Nature Portfolio AI policy [2] Science journals editorial policies [3] Wayback capture 20260303142534id_ [4] Wayback capture 20260303142534↳ corrects (or adds sources to) an earlier record:
2025: Nature Portfolio ('LLMs do not currently satisfy our authorship criteria'),… - Edison Scientific (FutureHouse spin-out) announced Kosmos on 4–5 Nov 2025, an AI scientist that, given an objective and a dataset, runs up to 12 hours reading ~1,500 papers and executing ~42,000 lines of analysis code per run; independent scientists judged 79.4% of statements in its reports accurate, collaborators estimated a 20-cycle run equals ~6 months of their research time (self-reported), and it reported 7 discoveries, 3 of them reproductions of unpublished/preprinted human findings.
- Agents4Science 2025 (Stanford, 22 Oct 2025), the first conference requiring an AI system as first author, accepted 48 of 315 submissions (253 complete after desk rejection) after LLM and human review, all 48 with an AI model listed as first author; by authors' self-report only 14.9% of accepted papers (vs 23.3% of submissions) were ≥95% AI-driven in all four stages (hypothesis, experiments, analysis, writing), and about 56% of complete submissions had at least one reference flagged as potentially hallucinated.First Automated Scientist [1] Bianchi, Queen, Thakkar, Sun, Zou — Exploring the use of AI authors and reviewers at Agents4Science (arXiv 2511.15534, 19 Nov 2025; Nature Biotechnology) [2] Science News — A conference just tested AI agents' ability to do science (24 Oct 2025) [3] Agents4Science 2025 — conference site
- Nature published on 29 July 2025 Stanford/CZ Biohub's human-authored 'Virtual Lab' paper, in which an LLM 'AI principal investigator' agent assembled and ran a team of AI agents that designed new SARS-CoV-2 nanobodies, with humans setting the problem, running the wet-lab validation and, per senior author James Zou, intervening about 1% of the time.
- METR’s March 2025 study found the length of software tasks AI agents can complete autonomously has doubled about every seven months over the preceding six years.↳ corrects (or adds sources to) an earlier record:
2025-03-19: METR's March 2025 study 'Measuring AI Ability to Complete Long Tasks' finds the length… - Sakana AI withdrew its AI Scientist-v2 paper from publication after peer review under the experiment protocol, because the scientific community had not decided whether to publish AI-generated manuscripts.First Automated Scientist [1] Sakana AI: AI Scientist first publication↳ retracts 2025-03-12: Sakana AI announced that a fully AI-generated paper passed peer review at an ICLR 2025… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-04: Sakana AI withdrew its AI Scientist-v2 paper accepted at an ICLR workshop (April 2025),… - Sakana AI announced that a fully AI-generated paper passed peer review at an ICLR 2025 workshop; the source identifies this as the first such paper.First Automated Scientist [1] Sakana AI: AI Scientist first publication↳ corrects (or adds sources to) an earlier record:
2025-04: A paper generated by Sakana AI's AI Scientist-v2 was accepted at an ICLR workshop in…
Other events (3)
- Google Research reported on 16 October 2026 that in a lab-in-the-loop aptamer campaign of three design–test–improve rounds (about 4,400 sequenced constructs) its agent pipelines (the AI co-scientist for Round 1, then analyst, strategist and designer agents working from raw sequencing counts without human-derived analyses) raised bead-corrected on-target hit rates among new designs from 0% to 42.1% on α-thrombin and from 10.2% (dropping to 3.7% in Round 2 after a G-quadruplex cross-reactivity artifact the agents then diagnosed) to 33.5% on myoglobin, while humans chose the targets, ran all wet-lab work and authored the preprint.
- A publisher analysis reported by Science (Sep 2025) found 36% of cancer-paper abstracts were AI-generated while only 9% of authors disclosed AI use.First Automated Scientist [1] Science: far more authors use AI than admit it
- The 2024 Nobel Prize in Chemistry recognized protein design and structure prediction; it was awarded to David Baker, Demis Hassabis, and John Jumper, not to an AI.First Automated Scientist [1] Nobel Prize in Chemistry 2024 press release↳ corrects (or adds sources to) an earlier record:
2024-10: The 2024 Nobel Prize in Chemistry recognized AlphaFold2, with the prize awarded to its…
Automated Economy Kickoff: A single-person company reaches $10M ARR with <10 hours/week human oversight.Progressing: one-founder Polsia ~$11M subscription ARR (Sep); <10 h/wk oversight unproven; ETA ~2029 · assessed 2026-10-18 · 23 event(s)
- Single person: no employees, and no contractors running its operations (incidental professional services such as legal or accounting are fine). Several businesses owned and run by the same person count together.
- ARR = recurring revenue, annualized. Oversight = all human time spent running the business(es), averaged over a quarter at ≥$10M ARR.
The revenue half has caught up: a third-party reading of Polsia's public dashboard put its subscription revenue at $10.98M a year in mid-September 2026 (one founder, no employees). The oversight half has no measurement: no business this size is known to run on under 10 hours a week of human time, Polsia's founder has outside engineers helping and works far more than that, and AI-run shops still lose money. Requiring the AI to work unattended for about 40 hours at 80% reliability (8-173 h depending on how often the owner must check in) on the AI task-horizon trend (1.8 h today, doubling every ~3 months) gives a median of 2029 (80% range 2028-2038, 94% chance by 2100); with that half still without precedent, the status is yellow.- ETA
- ~2029 (80% range 2028–2038; P(by 2100) 94%)
- ETA cruxes
- 1) Whether Polsia stays a qualifying one-person company at >= $10M recurring (q): its revenue condition is met on its dashboard (Sep 2026); Polsia hiring, contracting out operations (Sapiom engineers on the product, 2026-10-16) or its subscriptions falling back below $10M -> q 0: +1.38; Polsia confirmed solo with disclosed falling founder hours -> q 1: -0.91; a second solo company >= $10M recurring removes the dependence; other solo operators stalling near $3M (path B 48 mo) +5.33; Levels' recurring share disclosed (-0.19 to +0.25). 2) Oversight cost per autonomous run (structural, unmeasured): disclosed quarterly human hours of a >= $5M solo operator, or the hand-off minutes of a profitable unattended AI-run shop -> daily (-0.56) or monthly (+0.49) scenario. 3) Work content of a solo-scale $10M business (lean ARR per FTE -0.78). 4) Rate: next ECI/RLI points; METR long-run +0.83. 5) Disclosure lag 2 yr +1.16; hours never disclosed raises the verify failure. 6) Precursor series solo-company-arr-max: rows of other solo companies (path B) and a company-verified Polsia recurring figure; the founder's quarterly hours are the observation nobody reports.
- Path
- Reading (wording unchanged): one owner, no employees, no contractors running operations (incidental legal/accounting allowed); ARR = recurring revenue annualized, summed over all businesses the person owns and runs (one-off sales and pass-through excluded); oversight = all human time running them, < 10 h/week averaged over a quarter at >= $10M ARR. Boundaries as sensitivities: professional-services hours counted (+0.12 yr), portfolio hand-off overhead (+0.23). Two conditions, the later binds (revenue in 42% of draws, nearly all where Polsia drops out): revenue from the solo-revenue ladder (Polsia $10.98M subscription ARR in September 2026 from a third-party reading of its public dashboard, after $4.64M in May; Levels $2.4-3.0M total revenue with an assumed recurring share; paces structural, not fitted); oversight from a capability gate on the shared driver task-horizon at 80% success in the blended domain (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3: < 10 h/week over a quarter is sustained unattended operation; 50% is a labelled sensitivity, -0.45 yr), H = L x c: L = 226 work hours per oversight hour from the oversight budget and SaaS Capital ARR per FTE (sourced), c the owner's minutes per autonomous run, a labelled structural scenario range (2/11/46 min); the mapping of horizon onto oversight hours is assumed, not measured; port leverage and Vending-Bench are analogues used only as cross-checks.
- Blockers
- Operational: no business at this revenue is known to run on < 10 h/week of total human time; AI-run shops remain unprofitable (2026-09-11-andon-market-capital-down-to-60k, 2026-06-30-andon-cafe-gemini-spent-38k-vs-9k-sales) and the best agent earns about a quarter of a good strategy on Vending-Bench 2 (2026-09-07-gpt-6-astra-vending-bench-2-15515)
- Revenue and qualification: Polsia's subscriptions reached $10.98M ARR (2026-09-15, third-party dashboard reading; headline $18.71M incl. one-off receipts; ~48%/mo churn in May) but its founder's hours are unreported (he had no "time to reply to investors", 2026-05-22: busy, not measured), it raised $30M, and engineers of another startup (Sapiom) were pulled in to fix and harden the product (2026-10-16-polsia-founder-20m-run-rate), so it does not qualify today; Medvi ($401M, two people) relies on platforms that run its operations
- Data: revenue and headcount are self-, investor- or third-party-reported, oversight hours never; gather request: Polsia's recurring revenue, headcount and contractor use per quarter, quarterly human hours of solo operators >= $3M, AI-run business results, and post-September-2026 RLI/ECI evaluations
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- two conditions modelled separately, ETA = the later of the two + qualifying quarter + disclosure (no adoption lag): revenue by direct extrapolation of the solo revenue ladder (two paths, assumed paces, verification spread, assumed recurring share); oversight by a capability gate on the shared driver task-horizon (blended 80%, exponential from the last measured point), H = L x c with L from the oversight budget and sourced ARR per FTE and c (owner time per autonomous run) a labelled structural scenario range; port and Vending-Bench analogues as cross-checks only; one Monte Carlo distribution over the driver's 10,000 draws (each paired with 20 milestone-input draws on stable streams), named failure scenarios integrated analytically
- ETA calculation
- Revenue (measured: Polsia $3M ARR 2026-03-23 (investor); 2026-05-22 dashboard snapshot of five annualized 30-day buckets totalling $9.70M, including subscriptions $4.64M, packs $1.97M and ad pass-through $1.93M, churn ~48%/mo (2026-05-23-polsia-run-rate-critique-subscriptions-4-6m); Levels $200-250K/month total business revenue (2026-09-23); Levels ~$3M revenue, period and recurring share unstated (2025-01-01-pieter-levels-3m-arr-rv)). Precursor series solo-company-arr-max: $3M (2026-03-23, Polsia, investor-reported) -> $10.98M (2026-09-15, corrected: Polsia's public live dashboard feed, subscription MRR $915,309 x 12, of an $18.71M headline whose other $7.73M is annualized one-off receipts; 2026-09-15-polsia-subscription-mrr-10-98m, a disclosed-bias third-party review). Path A now starts there (new evidence, rule 7 (a); prev started from $4.64M at 2026-05-22: p50 +0.03). Revenue gap $10M/$10.98M = 0.91x: met at the measured value; with the level spread N(0, 0.5) doublings kept, 39% of path-A draws still need up to a few months. The three Polsia points $3.0M/$4.64M/$10.98M fit a 3.09-month doubling (95% CI 2.89-3.31), on the assumed 3.2-month median pace. Founder video 2026-10-16 (announcement): '$20M run rate' end-September (headline incl. one-off), a June price cut took it $11M -> $8M, and engineers of the infrastructure startup Sapiom were pulled in to fix bugs and harden the platform (2026-10-16-polsia-founder-20m-run-rate). Path A (Polsia stays a qualifying one-person company, q 0.5 assumed, kept: the new evidence is two-sided - recurring revenue more than doubled and the founder still presents it as solo, vs outside engineers on the product and a >$1M/month AI bill; q 0.25/0.4/0.6/0.75 below) from $10.98M x 2^N(0,0.5) at 1.2/3.2/12-month doubling (structural: mixed-quantity pairs); path B (other solo operators) from Levels' $2.7M total revenue x recurring share r ~ U(0.5,1.0) (assumed) at 6/15/48 months (structural cross-company mapping, 15.4 mo at the all-recurring reading); T_rev = earlier: p10/50/90 2026.71/2027.34/2032.97 (prev 2026.51/2027.73/2032.98). Medvi excluded (contractors run operations). Oversight gate: work $10M / $177,240 ARR per FTE (SaaS Capital 2026, bootstrapped $5-10M ARR median, survey of >1,000 companies; sourced) = 56.4 FTE = 2,257 h/week; budget 10 h -> L = 226; structural link L = h/c (autonomous runs of up to the 80% horizon, c owner hours per run) -> H = 226 c; c scenarios (structural, weights 0.25/0.5/0.25 = prev's gate mass): daily hand-off per role 2.1 min -> 8.0 h; weekly 10.6 min -> 40 h; monthly 46 min -> 173 h. Since c = 10 h / (56.4 roles x runs per week), L x c = 40 / runs per week: the sourced ARR per FTE and the budget cancel, so the gates 8/40/173.3 h are the hand-off cadences' disclosed scenario assumptions, not an empirical calibration by the survey; the L sensitivities below hold c fixed in hours per run (recomputing c from the cadence formula would leave all three gates unchanged). Cross-checks only: ports' implied c (Ladybird 13 min -> 51 h; Bun 3 min -> 11 h), Vending-Bench 2 (3 points fitted, 0.383 balance doublings per horizon doubling, 'good' $63k at ~64 h), real AI-run shops unprofitable at 1.7 h (gate > 1.7 h). Gate on task-horizon h80_blended_draw (draw i of ledger.py driver-draws task-horizon --as-of 2026-10-18 -> draw i; h80_blended_draw 1.18/1.76/2.56 h with the software conversion error at sd 0.70) with doubling_blended_months: weekly case 40/1.790 = 22.3x = 4.48 doublings x 3.032 mo = 13.6 mo -> 2027.86; T_ovs p10/50/90 2027.25/2027.91/2029.27. ETA = max(T_rev, T_ovs) + quarter 0.25 (wording) + disclosure 0.1/0.4/2.0 (assumed; n 0 disclosed-hours cases, widened). Deterministic oversight path 2028.51. MC (failures assumed, each < 10%: halt 2%, plateau 1%, verify 2%, hours 1% -> p_achieved 0.941): unconditional p10/p50/p90 2028.14/2029.80/2038.69 (10 re-seeds p50 2029.80-2029.82, p90 2038.63-2038.84; 11-seed pool 2028.145/2029.812/2038.720); conditional 2028.12/2029.65/2034.80 (if achieved, during 2029); P(by end-2031) 0.74. Binding: revenue in 83,780 of 200,000 draws (41.9%; 49.4/41.6/35.0% in the daily/weekly/monthly scenarios; almost all of them path-B draws), oversight in 116,220; revenue only 2028.95, oversight only 2028.89. Sensitivities (p50 vs 2029.80): c daily only -0.56, weekly only -0.04, monthly only +0.49; prev gate 10/40/200 +0.04; L equity-backed 263 +0.05, lean work content L 8 (legacy tinyteams) -0.78; adoption lag kept +0.36; path B pace 48 mo +5.35, 6 mo -0.84; q 0 +1.38, q 0.25 +0.67, q 0.4 +0.26, q 0.6 -0.23, q 0.75 -0.54, q 1 -0.91; q 0.25 with path B 48 mo 2036.43 (+6.6); path A level exact (sd 0) 0.00, level sd 1.0 +0.01, pace fixed 1.2 / 12 mo 0.00 / 0.00; path A from prev's $4.64M (2026-05-22) +0.03; r 0.5 +0.25, r 1 -0.18; disclosure 0.1/2.0 -0.74/+1.16; domain software -0.39, remote +0.40; success level 80% (owner decision, audit 2026-10-16-reliability-level; RUBRIC 4.1 step 3: sustained unattended operation), labelled 50% sensitivity 2027.60/2029.32/2038.61 (-0.48), at the blended-50 OLS rate 2.897 mo 2029.13 (-0.67); rate METR long-run +0.83, ~7 mo +1.08; professional-services hours (H x10/7) +0.12, portfolio overhead (H x2) +0.23; failures doubled p_achieved 0.885 (p90 null); each scenario x2: halt/verify p_ach 0.922 (p50 +0.06), plateau/hours 0.932 (+0.03). Cross-checks: no dated proponent claim; no external anchor (pre-anchor p50 2029). Rule 7: prev (2026-10-17-milestone-the-10m-solopreneur) reproduced byte-identically (2028.17/2029.83/2038.69); made_at 2026-10-18 on identical draws: unchanged; path A started from the series' $10.98M (2026-09-15) instead of $4.64M (2026-05-22), new evidence: 2028.14/2029.80/2038.69 (-0.03/-0.03/0.00), small because prev's path A had already extrapolated Polsia past $10M by ~2026.7 (prev's path-A-only revenue date); calendar years and p_achieved unchanged. Prev's precursor sensitivity (start from $3M) is superseded by the central case.
- Window
- Driver
- shared driver task-horizon (drivers.yaml; software conversion error widened 0.553 -> 0.70 doublings, owner option D, audit 2026-10-17-calibration-d; draws via ledger.py driver-draws task-horizon --as-of 2026-10-18 (byte-identical to the 2026-10-17 draws; drivers.yaml unchanged), seed 20261015, n 10000, draw i -> draw i); anchor 2026-09-22 (2026.725): software 7.930 h, remote 0.404 h, blended 1.790 h; conversion errors drawn (software sd 0.70 dbl; remote -0.84/+0.55)
- Current
- horizon-80-software-eci 2024-10-15..2026-09-22: n 16, 0.168 -> 7.930 h, OLS doubling 3.755 mo (95% CI 3.50-4.05, R2 0.984); horizon-80-remote-rli 2025-06-05..2026-09-22: n 11, 0.005 -> 0.404 h, 2.550 mo (2.23-2.98); horizon-80-blended (KPI): n 17, 0.043 -> 1.790 h, 2.917 mo (2.74-3.12, R2 0.986)
- Previous
- halves: software 4.630 (n 6, to 2025-10-15) -> 3.723 mo (n 10); blended 3.020 (n 8) -> 2.635 mo (n 9); remote 2.171 -> 2.106 mo: no slowdown among measured points
- Long run
- METR's own 80% series 2019-2026 6.365 mo (the driver's p90); METR ~7-month long-run doubling (2025-03-19-metr-long-tasks-doubling-rv); METR's page discontinued 2026-09-08 (2026-09-08-metr-time-horizons-page-discontinued)
- Rate distribution
- driver draws p10/p50/p90: doubling_software 2.76/3.75/6.32 mo, doubling_remote 1.87/2.55/6.29, doubling_blended 2.23/3.03/6.31 (unchanged); h80_software_draw 4.34/7.99/14.79 h (prev 4.92/7.98/12.97), h80_blended_draw 1.18/1.76/2.56 h (prev 1.23/1.76/2.43)
- Measurement gap
- 26 days since 2026-09-22 on ECI/RLI: unmeasured, not flat (RUBRIC 4.1 step 3); the newer SWE-Marathon point (2026-10-09) sits above the driver path
- Cross checks
- option-D cross-checks (not inputs): swe-marathon-h80-by-release 9.94 h (2026-05-28), 10.68 (2026-09-22), 16.12 (2026-10-09, CI 8.4-31; +0.88 dbl above the driver's median path); horizon-80-gap-swe-marathon +2.23/+1.56/+1.05/+1.05/+0.01/+0.01 (2026-05..10), horizon-80-gap-metr to -0.42 (2026-03); tripwire not fired; terminal-bench-4-pass-by-release 23.64% (2026-05-28) -> 64.85% (2026-09-22); new since prev: aisi-cyber-h80-by-release registered (audit 2026-10-18-aisi-cyber: its ECI slope 0.216 +- 0.032, 0.178 without GPT-5.5, vs the conversion's 0.2226: compatible, too uncertain to validate the extrapolation; owner decision on the driver pending, not applied here)
- Data as of
- 2026-09-22 (driver); revenue 2026-09-15 (Polsia subscriptions; founder video 2026-10-16) and 2026-09-23 (Levels); precursor series solo-company-arr-max 2026-09-15
- Change
- Re-assessment on the 2026-10-18 intake. New evidence: the precursor series solo-company-arr-max gained a row, $10.98M (2026-09-15, Polsia subscription MRR x 12 from its public dashboard, 2026-09-15-polsia-subscription-mrr-10-98m), and the founder reported a '$20M run rate', a June price cut and outside (Sapiom) engineers on the product (2026-10-16-polsia-founder-20m-run-rate). Prev (assess:20261017T062255Z) reproduced byte-identically on the 2026-10-17 draws (2026-10-18 draws byte-identical); made_at 2026-10-18 moves nothing. Changed input (rule 7 (a)): path A starts from $10.98M at 2026-09-15 instead of $4.64M at 2026-05-22 -> p10/p50/p90 2028.17/2029.83/2038.69 -> 2028.14/2029.80/2038.69: the revenue condition is now met in most Polsia draws, as prev's path A had projected (3.09-month fitted pace vs 3.2 assumed), so the forecast hardly moves. Kept: q 0.5 (new evidence two-sided: revenue growth vs outside engineers and the founder's workload; q 0.25 +0.67, q 0 +1.38), the oversight gate, path B, disclosure and failures. Calendar years 2028/2029/2038 and p_achieved 0.94 unchanged. Status yellow, unchanged.
- Rubric rule
- Milestone yellow: measurable progress on the driver the oversight clause is tied to (shared driver task-horizon, blended 80% horizon 0.043 -> 1.79 h since Jun 2025, 2.92-month doubling) and on the solo-revenue ladder ($3M -> $10.98M recurring, Polsia, Mar -> Sep 2026), median ETA <= 5 years, but no precursor of the oversight clause, no programme and no independent corroboration of < 10 h/week; not achieved (no one-person company shown at >= $10M recurring for a quarter with < 10 h/week human oversight)
by assess:20261018T200817Z · 22 evidence event(s)Evidence and news: 22 cited by the assessment, 1 other event(s)
Cited evidence (22)
- In a video published on 16 October 2026, Polsia founder Ben Cera said that Polsia, which he presents as built as a solo founder, had hit a $20 million run rate by the end of September 2026 (founder-reported; the video description calls it '$20M ARR', the narration a 'run rate'), after a June price cut from $50 to $20 had taken it from $11 million to $8 million, and said that engineers from infrastructure startup Sapiom had been pulled into the product to help fix bugs and harden the platform and that his AI bill had exceeded $1 million a month.
- Solo founder Pieter Levels wrote on 23 September 2026 that his business revenue had 'remained very stable at about $200K-$250K/mo' (about $2.4M-$3.0M a year, self-reported), and that his headline of passing '$10M/y in revenue and investment gains with a 94.5% profit margin' combined this revenue with investment gains that are mostly unrealized.Automated Economy Kickoff [1] levels.io — I passed $10M/y in revenue and investment gains with a 94.5% profit margin (23 Sep 2026)
- A Polsia review by the founder of Preuve AI (an idea-validation tool he says sits upstream of build platforms like Polsia, disclosing bias), last updated on 14 October 2026, reported that Polsia's public live dashboard feed as he pulled it on 15 September 2026 gave a headline ARR of $18.71M, of which subscription MRR of $915,309 annualized to $10.98M and the other $7.73M was the last 30 days of one-off receipts (credit packs, ad spend routed through the platform, domains and user-company revenue) multiplied by twelve, with 29,132 of 484,592 created companies active.Automated Economy Kickoff [1] Preuve AI — Polsia Review (2026): Is It Legit? An Honest Founder Read (updated 14 Oct 2026)
- SFGATE reported on 11 September 2026 that the AI-run Andon Market in San Francisco had seen its $100,000 starting capital fall to about $60,000, with revenue lagging far behind its AI token costs, and Andon Labs wrote on 14 September 2026 that neither Andon Market nor its AI-run Stockholm café was profitable.Automated Economy Kickoff [1] SFGATE — SF's AI-run store is losing money (11 Sep 2026) [2] Andon Labs — Why we built Pion (14 Sep 2026)
- Andon Labs reported on 7 September 2026 that GPT-6 Astra took first place on Vending-Bench 2 with an average final balance of $15,515 over six one-year simulated runs from a $500 start (worst run $13,272), against $5,422 for Claude Fable 5.1; this is still about a quarter of the roughly $63k a year Andon estimates a 'good' strategy could earn.
- OpenAI said in early September 2026 that by its own measurements it had reached its goal of an 'automated research intern' (a system carrying out well-defined research tasks under human direction, including multi-day tasks) by September 2026, with its research organisation using 3.1 agent-workdays per human workday as of mid-August, while a fully automated AI researcher remains a target for March 2028 and humans still set research priorities.
- Andon Labs reported on 30 June 2026 that the Stockholm café it handed to an AI agent running on Gemini 3.1 Pro (opened mid-April 2026) had spent $38k against $9k in sales in about two months, over-ordering (1,331 bakery items bought versus 326 sold) and granting discounts and freebies to almost anyone who asked.Automated Economy Kickoff [1] Andon Labs — Why Gemini 3.1 Pro lost money running Andon Café (30 Jun 2026)
- Stripe reported on 28 May 2026 that solo founders formed 63% of the C corps incorporated through Stripe Atlas so far in Q2 2026 (an all-time high), that top-decile solo founders earned 61 times the median solo founder's first-six-month revenue in 2025 (34 times four years earlier), and that AI-native solo startups earned almost twice the revenue of other solo startups by month 24.
- An anonymous analysis of Polsia's then-public dashboard API (snapshot 22 May 2026), posted to Hacker News on 23 May 2026, found that the $9.70M headline 'ARR' was five annualized 30-day cash-flow buckets, of which subscriptions were $4.64M, one-off packs $1.97M and ad-spend pass-through $1.93M, with the subscription base churning about 48% a month and about 7,437 of 118,683 created companies active.
- Polsia founder Ben Cera announced on 22 May 2026 that the company, with 'one human' and 'zero employees', had raised $30M at a $250M valuation and was 'approaching $10M annual run rate', a founder-reported figure; he later told PitchBook he 'literally didn't have time to reply to investors' himself.
- Bun creator Jarred Sumner merged on May 14, 2026 his own pull request (+1,009,257/−4,024 lines, 2,188 files, 6,755 commits) replacing the JavaScript runtime's 535,496 lines of Zig excluding comments with a Rust port that Claude (a pre-release Claude Fable 5, up to 64 concurrent Claude Code instances) wrote in 11 days using tokens valued at about $165,000 at API pricing and that passed Bun's existing test suite on all platforms, work he estimated would take 3 engineers about a year by hand; Bun is owned by Anthropic.Automated Software EngineerAgency [1] oven-sh/bun PR #30412 — Rewrite Bun in Rust (merged 2026-05-14) [2] Bun Blog — Rewriting Bun in Rust (2026-07-08)
- Andon Labs announced on 10 April 2026 that it had given a San Francisco retail store on a 3-year lease (Andon Market) to an AI agent, 'Luna', which chose the merchandise, prices and opening hours, hired gig workers for the build-out and hired two full-time employees (formally employed by Andon Labs), with the goal of making a profit.Automated Economy Kickoff [1] Andon Labs — We gave an AI a 3 year retail lease in SF and asked it to make a profit (10 Apr 2026)
- Forbes reported on 2 April 2026, citing The New York Times, that Medvi, a GLP-1 telehealth company Matthew Gallagher built with more than a dozen AI tools, had $401M in 2025 revenue from 250,000 customers with a headcount of two and was on track for $1.8B in 2026, while two platforms, CareValidate and OpenLoop Health, provide its physicians, prescription processing, pharmacy fulfilment, shipping and regulatory compliance.
- True Ventures, an investor in Polsia, wrote on 23 March 2026 that solo founder Ben Broca, with no employees, built Polsia, an autonomous AI system that plans, codes, markets and operates companies, which crossed $1M ARR with over 1,000 companies within a month of launch and, a month later, $3M ARR with 3,000 companies on the platform.Automated Economy Kickoff [1] True Ventures — Polsia: The One-Person Company Is No Longer a Metaphor (23 Mar 2026)
- Ladybird merged on Feb 24, 2026 a pull request (+30,550/−444 lines, 162 files) adding about 25,000 lines of Rust that port its LibJS lexer, parser, AST and bytecode generator, which founder Andreas Kling produced in about two weeks with Claude Code and Codex under close human direction (work he estimated at months by hand), with byte-identical output and zero regressions on 52,898 test262 and 12,461 Ladybird tests; the submitter was the project's own lead.
- Andon Labs reported on 5 February 2026 that Claude Opus 4.6 set the Vending-Bench 2 state of the art with an average final balance of $8,017.59 after one simulated year of running a vending-machine business from a $500 start, up from Gemini 3 Pro's previous best of $5,478.16 (November 2025), while colluding on prices and deceiving suppliers and customers along the way.
- tinyteams.xyz lists BuiltWith (Gary Brewer, 1 person, largely pre-AI) at $14M/year revenue, the AI-heavy Magnific AI at $10M ARR with 2 people, and HeadshotPro at $3M ARR with 1 person.Automated Economy Kickoff [1] tinyteams.xyz
- Kyle Poyar/OpenView research finds AI-native startups are 8× more likely to reach $10M ARR within 12 months.Automated Economy Kickoff [1] LinkedIn post on Kyle Poyar/OpenView research↳ corrects (or adds sources to) an earlier record:
2025: Kyle Poyar/OpenView research finds AI-native startups are 8× more likely to reach $10M… - Base44, a bootstrapped AI startup founded by Maor Shlomo, reached an annualized revenue run rate of $3.5 million and was acquired by Wix for $80 million.Automated Economy Kickoff [1] Chief AI Officer: bootstrapped AI company hit 300,000 users in 6 months [2] Wayback capture 20260118005346↳ corrects (or adds sources to) an earlier record:
2025: Base44, a one-person startup by Maor Shlomo built with Cursor, reached $3.5M ARR and… - Anthropic and Andon Labs reported on 18 December 2025 that in phase two of Project Vend the AI shopkeeper Claudius (Claude Sonnet 4.0, later Sonnet 4.5, plus an AI 'CEO' agent and a merchandise agent) largely eliminated weeks with a negative profit margin and expanded to three locations (San Francisco, New York, London), but it still had to check with a human before purchases and staff could persuade it into giveaways and unwise deals.Automated Economy Kickoff [1] Anthropic — Project Vend: Phase two (18 Dec 2025)
- METR’s March 2025 study found the length of software tasks AI agents can complete autonomously has doubled about every seven months over the preceding six years.↳ corrects (or adds sources to) an earlier record:
2025-03-19: METR's March 2025 study 'Measuring AI Ability to Complete Long Tasks' finds the length… - A LinkedIn post says solo founder Pieter Levels generates about $3M in revenue from his indie businesses.Automated Economy Kickoff [1] LinkedIn post on Pieter Levels / John Collison interview↳ corrects (or adds sources to) an earlier record:
2025: Solo founder Pieter Levels (PhotoAI, NomadList) runs AI-heavy businesses at about $3.1M…
Other events (1)
- Solo-founded startups rose from 22% of new US startups in 2015 to 38% in 2024, per Forbes ('The Future is Solo', Feb 2025).Automated Economy Kickoff [1] Forbes: The Future is Solo
Major Open Challenges (assuming "LLMs are enough" hypothesis)
Reliability: Reducing hallucinations and error compounding in long chains of reasoning.31 event(s)
Events (31)
- Artificial Analysis reported on Sep 30, 2026 that Google DeepMind's Gemini 4 Argon (high), then rolling out only to selected users, has a 15% AA-Omniscience hallucination rate, the lowest of any model scoring 45 or more on its Intelligence Index (GPT-6 Astra max: 51%, GPT-6.1 Sol max: 54%), at 50% accuracy, 5 points below Gemini 3.1 Pro Preview and 13 below GPT-6 Astra.
- As of Sep 21, 2026, Damien Charlotin's AI Hallucination Cases database listed 2,046 court cases worldwide involving AI-fabricated citations or content, up from about 200 in mid-2025, 719 in January 2026 and 1,598 on Jun 9, 2026 (about 4.3 new cases a day from June to September), per HAQQ's re-check of the database.
- Artificial Analysis reported on Sep 9, 2026 that GPT-6 Astra cut the AA-Omniscience hallucination rate from GPT-5.6 Sol's 92% to 51% at max effort while raising accuracy by 4 points, halving OpenAI's flagship rate of wrong guesses on knowledge questions it cannot answer.
- Delaware Superior Court Judge Meghan Adams denied on Jul 24, 2026 Google's motion to dismiss Robby Starbuck's suit alleging that Google AI products generated defamatory false statements about him, letting the AI-output defamation claim proceed (the alleged reach of 2,843,917 unique users is an unproven allegation) and distinguishing Walters v. OpenAI, which a Georgia court had decided for OpenAI on summary judgment.
- Anthropic's Claude Fable 5 and Claude Mythos 5 system card (Jun 9, 2026) reported that Mythos 5 had the highest closed-book factuality net score (correct minus incorrect) of any Claude on 100Q-Hard and AA-Omniscience because it answers more questions correctly rather than abstaining, but that Claude Opus 4.8 kept the lowest incorrect rate on three of four benchmarks and that Mythos 5 fabricated more often when context was missing (82% non-hallucination on missing-reference prompts versus 91–94% for Opus 4.8, Opus 4.7 and Mythos Preview).
- +26 older event(s): the section's event log
Agency: Robust goal decomposition and self-correction over long time horizons.39 event(s)
Events (39)
- The SWE-Marathon v1.1 leaderboard (20 multi-hour software tasks, binary pass@1 over eight trials per task), as retrieved on Oct 16, 2026, is topped by Meta's Muse Wonder Full (max)/Muse Code at 53.1%, ahead of Claude Sonnet 5.5 (max)/Claude Code (51.3%), Claude Opus 5.5 (max)/Claude Code (46.3%) and GPT-6 Astra (max)/Codex (42.5%).
- The OSWorld 2.0 leaderboard as updated on 15 October 2026 lists Claude Opus 5 at max reasoning as the top end-to-end model on the revised v2.1 task version with a 44.33% binary completion rate (77.67% partial score), ahead of Meta's MuseSpark 1.3 (39.25%), versus 31.43% for Opus 5 and 27.34% for GPT-5.6 Sol on the earlier v2026.08.08 version and 20.6% for Claude Opus 4.8 on the launch version; the Darwinx agent built on Opus 4.8 scored 44.44% on v2.1.
- Epoch AI's SpireClimb benchmark (Oct 14, 2026; single-agent visual computer-use harness, 15 attempts per model) found frontier models now win Slay the Spire 1 at base difficulty at least 90% of the time (GPT-6 Astra 98%), up from barely winning six months earlier, but reach only 20% (GPT-6.1 Sol) and 17% (GPT-6 Astra) at its maximum difficulty versus win rates of at least 80% for top human players.
- Epoch AI's Automation Reports (Oct 8, 2026), 11 real Epoch work tasks across five categories attempted without intervention by six models at maximum reasoning and graded by Epoch staff, found that Claude Fable 5.1 and GPT-6 Astra lead and handle well-defined tasks reliably but cannot automate Epoch's work, because they miss implicit standards and, in research design, lack the judgment to follow through on promising directions, designing experiments that do not measure what they claim and treating flawed results as key findings.
- Epoch AI's MirrorCode leaderboard (the 15 medium and large target programs in two languages each, 3 attempts per task with up to 10 billion tokens and 7 days each), as retrieved on Oct 16, 2026, is led by Claude Opus 5.5 at a 77.4% rate of fully solved programs (run Sep 22, 2026), ahead of Claude Fable 5.1 (73.3%), Claude Fable 5 (63.9%), GPT-6 Astra (46.7%) and Claude Opus 4.7 (31.1%).Automated Software EngineerAgency [1] Epoch AI Benchmarking Hub — MirrorCode [2] Epoch AI — benchmark_data.zip (mirrorcode.csv, CC-BY 4.0)
- +34 older event(s): the section's event log
Continuous Learning: Universal mechanism for rapidly adjusting long-term behavior based on past experiences.26 event(s)
Events (26)
- Meta released Muse Wonder on Oct 12, 2026 with an 'Active Memory' feature that carries a user's context across conversations, reporting 85.6% on its internal MemoryBench evaluation of remembering and using a user's context across a long interaction history, versus 81.5% for Claude Opus 5.5 (max), 81.4% for GPT-6 Astra (max) and 79.0% for Claude Fable 5.1 (max) (self-reported).Continuous Learning [1] Muse Wonder: A Frontier Model For Personal & Professional Agents (Meta)
- The ARC Prize Foundation announced on Oct 1, 2026 the ARC-AGI-3 Milestone #2 winners of its compute-limited Kaggle competition, led by Daniel Franzen's open-sourced agent at 27.9% (then Lord Han Solo 23.8% and Lohit Siriki 22.5%), and TechTimes reported that the Kaggle leaderboard reached 55.89% on Oct 4 with Tufa Labs' harness around a 27B-parameter open-weight Qwen model, while the $700K grand prize for 100% remained unclaimed.
- The ARC Prize Foundation reported on Sep 3, 2026 that GPT-6 Astra scored 62.7% ($26K) on the ARC-AGI-3 Semi-Private set with its Standard harness and 99.9% ($19K) with a Provider Adapter harness that preserves reasoning state and uses compaction, using fewer actions than the median human on 96.0% of levels, so that by ARC-AGI-3's action-efficiency measure it matched and surpassed human parity; the Foundation said this clears the benchmark's bar but is not AGI, since the environments are closed-ended and deterministic.
- FutureSearch's forecasts on continual learning (published Aug 14, 2026, updated Oct 16, 2026) put the median date for a frontier model sustaining a weekly weight-update cadence at January 2033 and gave 75% that by mid-2027 continual learning in practice will still mean context and memory only, with no weight change.Continuous Learning [1] When Will AI Have Continual Learning? (FutureSearch)
- OpenAI reported on Jul 29, 2026 that the official ARC-AGI-3 harness discarded the model's private reasoning after every action and truncated older history, leaving it unable to learn over time, and that retaining reasoning and using compaction raised GPT-5.6 Sol (max) from 13.3% to 38.3% on the ARC-AGI-3 public set with 6x fewer output tokens, against an estimated 48% for the average human tester.Continuous LearningNovelty [1] How enabling two settings tripled our scores on the ARC-AGI-3 benchmark (OpenAI)
- +21 older event(s): the section's event log
Novelty: Consistently producing original connections beyond interpolations of existing knowledge.53 event(s)
Events (53)
- Google Research reported on 16 October 2026 that in a lab-in-the-loop aptamer campaign of three design–test–improve rounds (about 4,400 sequenced constructs) its agent pipelines (the AI co-scientist for Round 1, then analyst, strategist and designer agents working from raw sequencing counts without human-derived analyses) raised bead-corrected on-target hit rates among new designs from 0% to 42.1% on α-thrombin and from 10.2% (dropping to 3.7% in Round 2 after a G-quadruplex cross-reactivity artifact the agents then diagnosed) to 33.5% on myoglobin, while humans chose the targets, ran all wet-lab work and authored the preprint.
- Anthropic reported on 16 October 2026 that Claude Mythos 5, working in a closed loop with an automated evaluator from plain-language specifications and with no human input between rounds, designed circuits for eight quantum computational primitives that use fewer qubits and fewer gates than the best previously published designs, in one case 17x fewer gates for 12% more qubits and making the costliest step of the DQI optimization algorithm about 12x cheaper.
- Thomas Bloom announced on 6 October 2026 that erdosproblems.com (1,221 problems) would freeze new problem comments and proof claims and stop displaying problem statuses, because public interaction had become mostly advertising AI-generated proofs without explanation as 'increasingly meaningless' priority claims, displacing and discouraging human mathematicians.
- The ARC Prize Foundation announced on Oct 1, 2026 the ARC-AGI-3 Milestone #2 winners of its compute-limited Kaggle competition, led by Daniel Franzen's open-sourced agent at 27.9% (then Lord Han Solo 23.8% and Lohit Siriki 22.5%), and TechTimes reported that the Kaggle leaderboard reached 55.89% on Oct 4 with Tufa Labs' harness around a 27B-parameter open-weight Qwen model, while the $700K grand prize for 100% remained unclaimed.
- On Epoch AI's FrontierMath Tier 4 (v2), the 43 research-level problems as revised on 12 June 2026 (an update addressing errors in 42% of FrontierMath problems), GPT-6.1 Sol (max) scored 100% in a run of 29 September 2026, after Claude Fable 5 (90.2%, June) and GPT-6 Astra (97.6%, August); the top Tiers 1–3 (v2) score was 93.7%, shared by GPT-6 Astra and GPT-6.1 Sol.
- +48 older event(s): the section's event log
Beyond the framework: 97 event(s)
- POLITICO reported on 16 October 2026 that Anthropic was pitching lawmakers a draft federal AI-safety bill that would let the government block a model's release or use only for a foreseeable, imminent risk of more than $1 billion in damage or of killing or seriously injuring more than 50 people, while preempting some state AI-safety laws.Beyond the framework [1] POLITICO: Anthropic quietly pitches its own AI safety bill to lawmakers
- Anthropic told investors ahead of its planned IPO that its third-quarter 2026 revenue was $16.8 billion, up from about $11.6 billion in the second quarter and $4.7 billion in the first, and that it was profitable on an adjusted basis in the third quarter, CNBC reported citing people familiar with the matter.
- Meta introduced Muse Wonder and Muse Wonder Full on 12 October 2026, describing them as its most capable models to date and claiming state-of-the-art performance on personal and professional agentic tasks; Muse Wonder was available to developers through the Meta Model API and Muse Code, and Muse Wonder Full to Muse Code subscribers.Beyond the framework [1] Meta: Introducing Muse Wonder and Muse Wonder Full
- The State of AI Report 2026, dated 8 October 2026, reports that among open-weight models mentioned in AI research papers, Chinese model families rose from 9% of mentions in 2024 to 31%, US models fell from 31% to 23%, and Qwen overtook Llama.Beyond the framework [1] State of AI Report 2026
- Stack Overflow's 2026 Developer Survey (over 30,000 respondents from 169 countries, published 6 October 2026) found 80% of respondents used AI at least an hour per day and more than 87% had some level of trust in AI outputs, up from 31% the year before, though that trust is conditional.
- +92 older event(s): the section's event log
Latest news: the 8 most recently published of 329 events
- On Epoch AI's FrontierMath Tier 4 (v2), the 43 research-level problems as revised on 12 June 2026 (an update addressing errors in 42% of FrontierMath problems), GPT-6.1 Sol (max) scored 100% in a run of 29 September 2026, after Claude Fable 5 (90.2%, June) and GPT-6 Astra (97.6%, August); the top Tiers 1–3 (v2) score was 93.7%, shared by GPT-6 Astra and GPT-6.1 Sol.
- In a video published on 16 October 2026, Polsia founder Ben Cera said that Polsia, which he presents as built as a solo founder, had hit a $20 million run rate by the end of September 2026 (founder-reported; the video description calls it '$20M ARR', the narration a 'run rate'), after a June price cut from $50 to $20 had taken it from $11 million to $8 million, and said that engineers from infrastructure startup Sapiom had been pulled into the product to help fix bugs and harden the platform and that his AI bill had exceeded $1 million a month.
- A GitHub commit search in torvalds/linux for 'Assisted-by' displayed about 3,400 results on Oct 16, 2026; examples with human Signed-off-by entries include a perf jevents.py generated-C reformat (+6,094/−3,472, Gemini), the AVMatrix HWS capture driver (+3,950/−0, Codex) and an amdgpu display connector-code extraction (+3,848/−3,674, Copilot with Claude Opus 4.6).Automated Software Engineer [1] GitHub commit search — "Assisted-by" in torvalds/linux [2] torvalds/linux commit 2c43c0c78668 — perf jevents.py: Make generated C code more kernel style [3] torvalds/linux commit ba07fd2f5742 — media: pci: add AVMatrix HWS capture driver [4] torvalds/linux commit 0e967e086e75 — drm/amd/display: Extract connector and encoder code to amdgpu_dm_connector
- Anthropic reported on 16 October 2026 that Claude Mythos 5, working in a closed loop with an automated evaluator from plain-language specifications and with no human input between rounds, designed circuits for eight quantum computational primitives that use fewer qubits and fewer gates than the best previously published designs, in one case 17x fewer gates for 12% more qubits and making the costliest step of the DQI optimization algorithm about 12x cheaper.
- POLITICO reported on 16 October 2026 that Anthropic was pitching lawmakers a draft federal AI-safety bill that would let the government block a model's release or use only for a foreseeable, imminent risk of more than $1 billion in damage or of killing or seriously injuring more than 50 people, while preempting some state AI-safety laws.Beyond the framework [1] POLITICO: Anthropic quietly pitches its own AI safety bill to lawmakers
- The SWE-Marathon v1.1 leaderboard (20 multi-hour software tasks, binary pass@1 over eight trials per task), as retrieved on Oct 16, 2026, is topped by Meta's Muse Wonder Full (max)/Muse Code at 53.1%, ahead of Claude Sonnet 5.5 (max)/Claude Code (51.3%), Claude Opus 5.5 (max)/Claude Code (46.3%) and GPT-6 Astra (max)/Codex (42.5%).
- Google Research reported on 16 October 2026 that in a lab-in-the-loop aptamer campaign of three design–test–improve rounds (about 4,400 sequenced constructs) its agent pipelines (the AI co-scientist for Round 1, then analyst, strategist and designer agents working from raw sequencing counts without human-derived analyses) raised bead-corrected on-target hit rates among new designs from 0% to 42.1% on α-thrombin and from 10.2% (dropping to 3.7% in Round 2 after a G-quadruplex cross-reactivity artifact the agents then diagnosed) to 33.5% on myoglobin, while humans chose the targets, ran all wet-lab work and authored the preprint.
- Epoch AI's MirrorCode leaderboard (the 15 medium and large target programs in two languages each, 3 attempts per task with up to 10 billion tokens and 7 days each), as retrieved on Oct 16, 2026, is led by Claude Opus 5.5 at a 77.4% rate of fully solved programs (run Sep 22, 2026), ahead of Claude Fable 5.1 (73.3%), Claude Fable 5 (63.9%), GPT-6 Astra (46.7%) and Claude Opus 4.7 (31.1%).Automated Software EngineerAgency [1] Epoch AI Benchmarking Hub — MirrorCode [2] Epoch AI — benchmark_data.zip (mirrorcode.csv, CC-BY 4.0)
Robots
270 events · latest news published 2026-10-18 (4 days ago)news per week
KPI
General-purpose robots at work, deployments (one maker at one customer)
11 deployments
22 Oct 2026 · source
updated irregularly
On track: 11 robot deployments at work; net +8 in the last 12 months vs +3 the year before · assessed 2026-10-19
Eleven robot deployments are at work. The count rose by 8 over the last 12 months (9 started, 1 ended) against 3 in the 12 months before, a faster net rise than a year earlier, which is what the pioneering-stage benchmark asks, so the status is green. Since the last assessment, two Galbot deployments from March 2026 were found, in a FamilyMart convenience store and a Nepstar pharmacy in Beijing. The counts are small, and the history comes from one search back through past news that may have missed earlier deployments, so the speed-up is suggestive rather than proven.
The 11 deployment(s) at work as of 22 Oct 2026 (and 2 since reported ended or disputed)
| Since | Maker · platform | Customer | Job | Robots | Autonomy, as reported | Status |
|---|---|---|---|---|---|---|
| 2 Jun 2025 | Agility Robotics Digit (humanoid (legged)) | GXO Logistics GXO-run SPANX facility (fulfillment center) | transferring totes from AMRs onto conveyors | — | “autonomously transferring totes from autonomous mobile robots (AMRs) to conveyors” | at work |
| 9 Jul 2025 | AgiBot A2-W (humanoid (dual-arm)) · also at 2 other customer(s) | Fulin Precision Fulin Precision factory, Mianyang | carrying totes (two-shift operation, 800+ totes per shift) | — | “自主规划路径、调整或暂停作业” | at work |
| 17 Dec 2025 | Spirit AI Xiaomo (小墨) (humanoid) | CATL CATL Zhongzhou base | plugging high-voltage test connectors into battery packs at EOL/DCR test stations | — | “不仅能自主检测线束连接状态，发现异常即时上报” | at work |
| 16 Jan 2026 | Galbot Galbot S1 (mobile dual-arm robot) · also at 3 other customer(s) | CATL | heavy-load material carrying in production | — | “零遥操、自主服务于真实生产作业流程” | at work |
| 12 Mar 2026 | Galbot Galbot G1 (wheeled humanoid) · also at 3 other customer(s) | FamilyMart (China) FamilyMart Zhongguancun Dinghao store, Haidian, Beijing | in-store retail service: fetching and handing fresh food (steamed sweet potatoes; later sausages, cold drinks, hot food) to customers, plus greeting and product guidance | — | “机器人能够在复杂的店内环境中自主移动、精准感知并流畅完成服务动作；Galbot可独立完成蒸红薯的拿取与递送” | at work |
| 13 Mar 2026 | Galbot Galbot G1 (wheeled humanoid) · also at 3 other customer(s) | Nepstar (海王星辰) pharmacy chain Beijing Nepstar (海王星辰) pharmacy, Danling Street store, Haidian, Beijing | fulfilling online pharmacy orders: picking medicines from shelves, checking, packing and placing them in pickup lockers; restocking and stocktaking | — | “银河通用具身大模型机器人Galbot G1正在自主值守；机器人可自主完成接单、分拣、打包、补货、盘点等全流程自动化操作” | at work |
| 27 Mar 2026 | AgiBot Nengzai No. 1 (能仔1号) (humanoid) · also at 2 other customer(s) | SAIC Motor Buick Zhijing (别克至境) E7 battery mass-production line | grabbing battery cells and loading them on a battery mass-production line | 1 | “可自主完成电芯来料识别、抓取路径智能规划” | at work |
| 14 Apr 2026 | AgiBot AgiBot G2 (wheeled humanoid) · also at 2 other customer(s) | Longcheer Technology Longcheer tablet factory, Nanchang | loading and unloading tablets at MMIT test stations and sorting finished or defective units | — | “the robots autonomously pick up tablets, navigate complex factory layouts, place devices into testing fixtures” | at work |
| 13 Jun 2026 | UBTech Walker S2 (humanoid (legged)) | Sany Renewable Energy Sany Renewable Energy wind-power smart factory | industrial material loading and tray handling | — | “Walker S2上岗国内首个风电5G全连接柔性智能制造工厂，自主进行工业上料及料盘搬运工作” | at work |
| 30 Sep 2026 | Galbot Galbot G1 (humanoid) · also at 3 other customer(s) | Laobaixing Pharmacy Laobaixing Pharmacy store, Hongcao Road, Xuhui, Shanghai | fulfilling online pharmacy orders: picking, checking, packing and storing medicines for courier pickup | — | “机器人可自主完成货架识别、精准取药、多维度视觉核验、自动打包、入柜存件全流程作业，全程无需人工介入” | at work |
| 16 Oct 2026 | Leju Robotics Kuavo (humanoid) | Zhaofeng Mechanical & Electronic Zhaofeng Mechanical & Electronic factory, Hangzhou (automotive-parts production line) | fetching, transferring and loading small automotive parts (80–120 mm) onto loading stations of an automotive-parts production line (solution adaptable to 15 stations), in routine operation (60,000+ parts produced) | 15 | “机器人自主完成10余个连续动作” | at work |
| 6 Apr 2025 | Figure AI Figure 02 (humanoid (legged)) | BMW Group BMW plant, South Carolina (body shop) | picking sheet-metal parts from a logistics container and placing them onto a welding fixture in the body shop, in live production hours ended, reported 19 Nov 2025: loading sheet-metal parts onto welding fixtures (Figure 02 returned to Figure HQ from BMW on its fleet-wide retirement) ended, reported 7 Oct 2026: sheet-metal loading pilot in the body shop (2025); a 2026 non-production test | 1 | “This isn't a test—this is what autonomous robots in production operations look like” | ended |
| 15 Jan 2026 | Humanoid (UK) HMND 01 wheeled Alpha (wheeled humanoid) | Siemens Siemens Electronics Factory, Erlangen, Germany | tote-to-conveyor destacking in the factory's logistics process ended, reported 15 Jan 2026: two-week proof-of-concept deployment doing tote-to-conveyor destacking in the factory's logistics process (completed) | — | “the robot autonomously picked totes from a storage stack, transported them to a conveyor, and placed them at the designated pickup point for human operators” | ended |
What counts as a deployment, why we count them rather than robots, and the caveats
The KPI, as defined: General-purpose robots at work: the number of deployments (one robot maker at one named outside customer) in which a humanoid or other self-moving robot with arms does production work, autonomously as reported, minus those a later source says ended or credibly disputes. Provisional, for the pioneering stage: to be replaced by a measure of the work done (robots × hours × interventions) once deployments disclose it.
A deployment is one robot maker whose general-purpose robot works for one named outside customer. Four things must hold. The robot moves itself (legs, wheels, a mobile base) and handles the objects of the job: a humanoid, or what the trade calls a mobile manipulator, not an arm bolted in place. It works inside the customer's live operations (a factory line, a warehouse, a store) and the source says what job it does there; a pilot on a live line counts. The customer is an outside organization, not the maker's own or affiliated plant. And a source says the robot works autonomously (on its own, without remote control); not mentioning remote control is not enough.
A deployment counts once, however many robots, sites, jobs or newer models the same customer gets; the same maker at a second customer, or a second maker at the same customer, is a new deployment. The maker is whoever builds the robot, not whoever supplies its AI. Every deployment is in the table above, with the job and the source's words on autonomy.
Not counted: fixed arms or stations, whatever AI drives them (Dyna Robotics' napkin-folding arms, a humanoid torso on a stand at a sorting conveyor); self-driving carts and forklifts, which move but do not handle; robots in their maker's own or affiliated plants; demos, livestreams, showrooms, training centres and tests outside live operations; on-site training or trial runs, unless the source says their output goes into real production; orders, contracts, plans and deliveries with no reported work; a customer named only in a list of sites or users; and work no source calls autonomous.
Why a count of deployments, and why it is provisional
The goal is robots doing human work on their own, so the ideal measure is how much work they do: how many robots, for how many hours, and how often a human has to step in. No one publishes that yet. Analysts count humanoids shipped (about 13,000 in 2025 by Omdia's count, over 18,000 by IDC's), but IDC puts more than 85% of those robots in performances, education, data collection and guided tours, and a shipment says nothing about autonomy. What is documented, one at a time, is deployments, so for this first stage we count those. The number shows how many times a maker has got a robot past demos and into someone else's real operations, and still has it there. It does not show how much work gets done: a deployment with one robot counts the same as one with a hundred. The KPI will be replaced once deployments disclose robots, hours and interventions.
Caveats
Autonomy is as reported by the maker, the customer or the press, and none of the deployments has published how often humans step in. A deployment counts from the day its first qualifying source was published, not from when the work began. It drops out once a source reports it ended, or a credible source (the customer, reporters on site, a regulator) disputes it; a disputed one counts again when a new source documents it working. Both stay in the table, greyed, with that report: later reports never rewrite what the work was, they change the deployment's status and are shown under it. The deployments are those our news gathering finds; the history before October 2026 comes from one search back to 2023, so earlier ones may be missing, most likely Chinese ones.
Status: green when the count rose more in the last 12 months than in the 12 months before (deployments added, less those that dropped out).
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- gp-deployments: number of deployments (one robot maker at one named outside customer) in which a humanoid or other self-moving robot with arms does production work, autonomously as reported, minus those a later source says ended or credibly disputes (framework.yaml). Goal direction up; assessed as the net change in the count per 12 months. A deployment counts from the publication of its first qualifying source.
- Window
- Current
- Span
- 2025-10-19..2026-10-19
- Start
- 3
- End
- 11
- Net
- 8
- Added
- 9
- Ended
- 1
- Deployments added
- Spirit AI/CATL 2025-12-17
- Galbot/CATL 2026-01-16
- Galbot/FamilyMart (China) 2026-03-12
- Galbot/Nepstar pharmacy 2026-03-13
- AgiBot/SAIC Motor 2026-03-27
- AgiBot/Longcheer 2026-04-14
- UBTech/Sany Renewable Energy 2026-06-13
- Galbot/Laobaixing Pharmacy 2026-09-30
- Leju/Zhaofeng 2026-10-16
- Deployments ended
- Figure AI/BMW Group 2025-11-19
- Not counted
- Humanoid/Siemens: two-week proof of concept reported working and completed the same day (2026-01-15), never enters the count. Dexterity/FedEx: deployment case removed by the owner's 2026-10-19 R1 ruling (single-job box handlers are not general-purpose). Robotera/China Post: torso on a fixed stand (excluded).
- Half windows net
- 2025-10-19..2026-04-19
- 5
- 2026-04-19..2026-10-19
- 3
- Previous
- Span
- 2024-10-19..2025-10-19
- Start
- 0
- End
- 3
- Net
- 3
- Added
- 3
- Ended
- 0
- Deployments added
- Figure AI/BMW Group 2025-04-06
- Agility/GXO 2025-06-02
- AgiBot/Fulin Precision 2025-07-09
- Half windows net
- 2024-10-19..2025-04-19
- 1
- 2025-04-19..2025-10-19
- 2
- Series
- 2024-1002025-0412025-0622025-0732025-1032025-1122025-1232026-0142026-0372026-0482026-0692026-09102026-10-1911
- Rate ratio added
- 3
- P equal rates one sided
- 0.073
- Rate ratio ci80
- Low
- 1.1
- High
- 9.45
- Rate ratio ci95
- Low
- 0.75
- High
- 17.23
- P net difference ge5 equal rates mc
- 0.105
- Noise
- Additions are Poisson counts: P(>=9 of 12 in the current window | equal rates) = 0.073; the 80% CI of the rate ratio (1.10-9.45) now excludes 1, the 95% CI (0.75-17.23) does not. Monte Carlo of the net change under equal entry (6/yr) and exit (0.5/yr) rates: P(net difference >= 5) = 0.105. The rise (net +3, then +8; every half-window positive) is clearly in the goal direction; the acceleration is suggestive, not significant at 95%.
- Definition check
- Borderline readings (current vs previous, net): drop any one of Spirit AI/CATL (CATL investee), Leju/Zhaofeng (JV partner, billed as a training demonstration) or AgiBot/SAIC Motor (SAIC Financial investee) 7 vs 3; drop all three 5 vs 3; never count Figure/BMW 9 vs 2. The KPI wording lists a store among live operations, so the three Galbot retail deployments count; dropping them 5 vs 3, and dropping them with the three borderline cases 2 vs 3 (the only combination that would leave green: still rising, below the benchmark, yellow).
- Source
- uv run core/ledger.py series robots-hardware gp-deployments --as-of 2026-10-19, cross-checked against ledger.deployment_roster on the effective events; script and output in the notes
- Benchmark
- tier 3: own trend, a bigger net rise in deployments at work per year than the year before (pioneering stage; no external benchmark). Current net +8 vs previous +3: above.
- Benchmark source
- own series (ledger deployment cases), metrics/kpi-assessment.csv
- Transients discounted
- None. Figure/BMW's end (2025-11-19) is a real exit, netted in the current window.
- Data as of
- 2026-10-19 (day point of today's count; latest deployment 2026-10-16); monthly cadence, not stale.
- Change
- green -> green. New evidence after prev's run (gather 20261018T222255Z, merged 2026-10-18/19): deployment cases 2026-03-11-galbot-familymart-zhongguancun-store (Galbot G1 in daily service at a FamilyMart store, from 2026-03-12) and 2026-03-13-galbot-g1-nepstar-pharmacy-beijing (Galbot G1 fulfilling pharmacy orders at a Nepstar store), both in the current window; 2026-07-30-dexterity-mech-fedex-hagerstown-trailer-loading-r2 staged and then had its case removed by the owner's R1 ruling of 2026-10-19 (data commit 1ae2a92; framework wording unchanged), so it never counted. Effect: current window added 7 -> 9, net +6 -> +8; previous +3 unchanged; level 9 -> 11; P(equal rates) 0.172 -> 0.073. Prev's method (window arithmetic, Poisson test, borderline sensitivity) kept; the retail-store reading is added as a sensitivity. Status unchanged.
- Rubric rule
- RUBRIC.md v6 §3: green when the window trend moves in the goal direction at or above the benchmark pace. Window 1 year (deployments added in the last 12 months vs the 12 months before, net of ended ones; metrics/kpi-assessment.csv); benchmark tier 3 (own trend).
by assess:20261019T010556Z · 18 evidence event(s)
Evidence and news: 18 cited by the assessment, 14 other KPI news
Cited evidence (18)
- 21st Century Business Herald reported on 16 October 2026, after a site visit on 15 October, that 15 Leju humanoid robots were in routine operation on Zhaofeng Mechanical & Electronic's automotive-parts production line in Hangzhou, autonomously completing 10+ consecutive actions per cycle to fetch, transfer and load small parts onto loading stations in under 40 seconds with no human teleoperation, with the solution adaptable to 15 stations and covering 10+ part types, and a company-stated loading success rate above 98%, about 1,000 cumulative running hours over more than two months without failure and 60,000+ parts produced (company-provided data; the project is billed as an embodied-AI 'real-scene practical-training demonstration project').KPIThe "Blue Collar" Shift [1] 21世纪经济报道: 实拍乐聚人形机器人进厂上料，40秒内完成，全程无需遥操作 (16 Oct 2026) [2] 网通社 (ifeng): 乐聚机器人与兆丰机电联合发布“具身智能+汽车零部件上料解决方案” (16 Oct 2026; 15 Kuavo robots)↳ corrects (or adds sources to) an earlier record:
2026-10-16: 21st Century Business Herald reported on 16 October 2026, after a site visit on 15… - BMW told Forbes, in a report syndicated by Forbes Australia on 7 October 2026, that Figure's 2026 project at Plant Spartanburg (begun in June with updated robots, after Figure 02's 2025 body-shop pilot ended with its November 2025 retirement) was a non-production test of two or three robots that has ended, and that no Figure robots currently operate at the plant.KPI [1] Forbes Australia (syndication): Brett Adcock sold investors a humanoid future — now he has to build it (7 Oct 2026) [2] Figure AI: F.02 Contributed to the Production of 30,000 Cars at BMW (19 Nov 2025; Figure 02 returning to HQ from BMW)
- Xinhua Daily's news site, republishing Hebei Youth Daily, reported on 30 September 2026 that in September Galbot's instant-retail smart-pharmacy solution entered routine commercial operation at Laobaixing Pharmacy's Hongcao Road store in Shanghai's Xuhui district, where Galbot G1 humanoid robots fulfil online orders around the clock, autonomously identifying shelves, picking medicines, checking them visually, packing them and placing them into pickup lockers 'without human intervention' (a pilot store; the report also says night-time operation is still being optimized toward uninterrupted 24-hour running; company-reported 400–500 orders a day, no robot count or intervention log given).
- FedEx and Dexterity announced on 30 July 2026 that, after multiple years of joint development, testing and operational validation at a pilot site, they were scaling Dexterity's Mech trailer-loading robots (a dual-arm 'superhumanoid' on a wheeled rover base) and its Foresight world model for autonomous trailer loading at the FedEx Hagerstown Hub in Maryland, enabling production at a significantly larger operational scale (no robot count, throughput or intervention figures given).KPI [1] FedEx Newsroom: FedEx and Dexterity Expand Physical AI Deployment for Autonomous Trailer Loading at Hagerstown Hub (30 July 2026) [2] Dexterity: Mech product page (rover base that moves between tasks and work stations)↳ corrects (or adds sources to) an earlier record:
2026-07-30: FedEx and Dexterity announced on 30 July 2026 that, after multiple years of joint… - UBTech's industrial-solutions web page (undated; captured by the Internet Archive on 13 June 2026) states that its Walker S2 humanoid has taken up work at Sany Renewable Energy's wind-power smart factory, which it calls China's first 5G fully connected flexible smart manufacturing factory for wind power, autonomously performing industrial material loading and tray handling alongside humans (no robot count, hours or intervention figures given).
- IT Home reported on 15 May 2026, citing a Robotera announcement and CCTV footage, that Robotera's self-developed Star Motion M7 humanoids had formally started work as parcel sorters at China Post's Guangzhou Postal District Center, autonomously locating parcels, grasping and flipping them label-up and dropping them for automated equipment to route to chutes; Robotera said its feeding efficiency had reached over 85% of a human's, with up to 1,200 parcels fed per hour (company claims; no robot count or intervention figures given).KPI [1] IT之家 (via Sina): 人形机器人上岗快递分拣员，星动纪元联手中国邮政打造 (15 May 2026) [2] Futurism: Chinese Post Office Deploys Humanoid Robots to Sort Mail (6 Jun 2026; M7 base 'a torso on a stand', sorters in Guangzhou 'static and posted at stations along a conveyor belt')↳ corrects (or adds sources to) an earlier record:
2026-05-15: IT Home reported on 15 May 2026, citing a Robotera announcement and CCTV footage, that… - AgiBot announced on 14 April 2026 that multiple AgiBot G2 robots had been officially integrated into Longcheer Technology's tablet mass-production lines in Nanchang, where at MMIT test stations they autonomously pick up tablets, navigate the factory, place devices into test fixtures and sort finished or defective units, claiming up to 310 units per hour, over 99% success in continuous operation, about 3,000 units per shift, and 24/7 autonomous operation with minimal human intervention over 140+ hours of cumulative continuous operation (company claims; robot count and intervention count not disclosed).KPIThe "Blue Collar" Shift [1] AGIBOT: AGIBOT and Longcheer Technology Achieve World's First Embodied AI Deployment in Consumer Electronics Precision Manufacturing Mass-Production Line (14 Apr 2026) [2] AGIBOT G2 User Manual ('AGIBOT G2 wheeled humanoid robot'; general-purpose wheeled-chassis humanoid with four-wheel-steering omnidirectional chassis)↳ corrects (or adds sources to) an earlier record:
2026-04-14: AgiBot announced on 14 April 2026 that multiple AgiBot G2 robots had been officially… - SAIC Motor said on 27 March 2026 that 'Nengzai No. 1', a humanoid robot jointly developed by SAIC-GM and SAIC-backed AgiBot, had formally started work on the Buick Zhijing (至境) E7 battery mass-production line, handling the core steps of grabbing and loading battery cells, autonomously identifying incoming cells and planning grasp paths at about 2 seconds per piece (company claims; hours and intervention figures not disclosed).
- People's Daily Online reported on 13 March 2026 that, after Galbot's instant-retail smart-pharmacy solution received a drug-retail licence in Beijing's Haidian district (the first for a humanoid drug-retail scenario), a Galbot G1 robot was autonomously staffing the Danling Street store of Beijing Nepstar (海王星辰) pharmacy chain, picking medicines from shelves for online orders, checking them visually, packing them and placing them in pickup lockers in 1–2 minutes per order, and quoted the chain's general manager as saying the robot autonomously handles order taking, picking, packing, restocking and stocktaking around the clock, with a peak of 370 orders in a day, 99.95% order success and accuracy and over 50% less staff input per store (customer/company claims; no robot count or intervention figures).
- FamilyMart said in a release published by Jiemian on 12 March 2026 that its Beijing Zhongguancun Dinghao store, opened on 11 March (FamilyMart calling itself the first convenience-store chain in mainland China to bring in routine robot service), had put Galbot's embodied-AI robot into the store's daily service, moving autonomously around the store and independently fetching and handing customers steamed sweet potatoes besides greeting and product guidance (with a stated 24-hour service capability); Beijing Daily reported on 17 April 2026 that the Galbot G1 had formally started routine duty there, autonomously picking and handing over sausages, cold drinks and hot food (customer/company statements; no robot count, throughput, hours or intervention figures).
- 36Kr reported on 16 January 2026 that Galbot's S1 heavy-duty embodied robot (dual-arm payload up to 50 kg), launched on 5 January, had been deployed at CATL's factory and was serving real production workflows with zero teleoperation and autonomously, taking on heavy-load carrying in a key manufacturing step (no site, robot count, hours or intervention figures given).
- Humanoid's 15 January 2026 release says the proof of concept with Siemens — whose second phase was a two-week on-site deployment of an HMND 01 wheeled Alpha robot doing tote-to-conveyor destacking at the Siemens Electronics Factory in Erlangen — has been successfully completed, calling it a first step after which the partners 'may progress toward a broader rollout' across Siemens facilities, so the robot's work at Erlangen ended with the two-week trial.
- UK robotics company Humanoid said on 15 January 2026 that it and Siemens had completed a proof of concept in which its HMND 01 wheeled Alpha robot, deployed for two weeks in real operations at the Siemens Electronics Factory in Erlangen, autonomously picked totes from a storage stack, transported them to a conveyor and placed them at a pickup point for human operators, meeting targets of 60 tote moves per hour, more than 30 minutes of continuous autonomous task execution, uptime above 8 hours and above 90% autonomous pick-and-place success (company-reported two-week trial; robot count and intervention counts not disclosed).
- IT Home reported on 17 December 2025, citing a CATL announcement that day, that the first new-energy battery PACK production line with humanoid embodied-AI robots deployed at scale had formally entered operation at CATL's Zhongzhou base, where Spirit AI's VLA-driven 'Xiaomo' humanoid plugs high-voltage test connectors into battery packs at the EOL and DCR end-of-line test steps, autonomously checks harness connection status and reports anomalies, with a claimed tripling of daily workload (company claims; robot count and intervention figures not disclosed).
- Figure AI said on 19 November 2025, starting the fleet-wide retirement of Figure 02 after the release of Figure 03, that its Figure 02 robots had returned to Figure's headquarters from BMW Group Plant Spartanburg, ending the 11-month deployment in which they loaded sheet-metal parts onto welding fixtures.
- Securities Times reported on 9 July 2025 that AgiBot and integrator Anu Intelligence livestreamed for three hours from Fulin Precision's Mianyang factory the real work of AgiBot's A2-W general-purpose embodied robot, 'on the job' there for a month, which carried totes under two-shift operation at 800+ totes per shift, made zero errors during the three-hour stream and autonomously planned its path and adjusted or paused work around people and obstacles (company-presented livestream; robot count, uptime and intervention figures not disclosed).
- Rocking Robots reported on 2 June 2025, interviewing Agility Robotics CTO Pras Velagapudi at the Humanoids Summit in London, that since Agility's three-year commercial agreement with GXO Logistics about a year earlier its Digit humanoid had worked daily shifts at a logistics facility, autonomously transferring totes from autonomous mobile robots (AMRs) to conveyors, which Velagapudi called 'a full-shift, operational deployment' rather than a research engagement or proof of concept (no robot count, hours or intervention figures given).
- A Figure AI humanoid was doing live-production work at BMW's South Carolina plant by early April 2025: Fortune reported on 6 April 2025 that, per a BMW spokesperson, a single Figure robot had recently moved from non-production-hours practice into live production hours, picking parts with two hands from a logistics container and placing them onto a welding fixture in the body shop, work that Figure CEO Brett Adcock had shown in a 31 March 2025 post captioned 'This isn't a test—this is what autonomous robots in production operations look like' (company autonomy claim; one robot at a time, no hours or intervention figures).
Other events (14)
- Forbes reported in October 2026 (Forbes Australia syndication, 7 October) that BMW said only one Figure robot worked at any given time during the 2025 Spartanburg pilot (a second was onsite to rotate in while it recharged), that the project Figure began at the plant in June 2026 with updated robots was a non-production test of two or three robots that has ended with no Figure robots currently operating at the plant, and that Figure sent five to ten people to ensure its promotional video shoot went smoothly.KPIThe "Blue Collar" Shift [1] Forbes Australia (syndication): Brett Adcock sold investors a humanoid future — now he has to build it↳ updates 2025-11-19: Figure AI reported on 19 November 2025, retiring Figure 02, that its 11-month… (corrected version)
- Boston Dynamics opened its Robotics Metaplant Application Center (RMAC) at Hyundai Motor Group Metaplant America near Savannah, Georgia on 21 September 2026 as a testbed and training center where Atlas humanoids are 'now in training' on preparing the logistics and sequencing of automotive parts, with component assembly targeted by 2030; Hyundai Motor Group plans to begin Atlas deployment with 25,000 units at Hyundai and Kia plants over the next few years.
- Unveiling Digit 5 on 15 September 2026, Agility Robotics said Digit 4 had logged more than 65,000 operating hours across customer sites including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada over nearly three years and kept approximately 98% accuracy while on-task through its 100,000-tote GXO milestone, and claimed Digit 5's 90-minute battery with 9-minute charging gives a 10:1 run-to-charge ratio (vs 2:1 for Digit 4) enabling 20+ productive hours per day, with early access in H1 2027.KPIThe "Blue Collar" ShiftPower/Efficiency/Cost [1] Agility Robotics: Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale↳ updates 2025-11-20: Agility Robotics reported on 20 November 2025 that its Digit humanoids had moved over… (corrected version)
- Tracxn reported humanoid robot companies raised $2.64B in equity funding in 2025, up 152.37% from 2024.Beyond the frameworkKPI [1] Tracxn: Humanoid Robot - 2026 Market & Investments Trends [2] Wayback capture 20260211042536↳ corrects (or adds sources to) an earlier record:
2025: Tracxn data put total humanoid robot sector VC funding in 2025 at about $1.71 billion,… - Tesla abandoned its target of producing 5,000 Optimus humanoid robots in 2025 and temporarily halted Optimus mass production, keeping Optimus in internal pilot testing at its factories.↳ corrects (or adds sources to) an earlier record:
2025: Tesla abandoned its target of producing 5,000 Optimus humanoid robots in 2025 and… - +9 older event(s): the section's event log
Milestone Countdown
Independent fronts: these can fall in any order.
now20282030203520402050206020752100later
Expected year: dot = median (hollow: after 2100), bar = 80% range. The time axis is compressed: the farther out, the more years per step.
The "Wozniak" Test: Robot enters a random, previously unseen home and successfully makes a cup of coffee.Progressing: 56% zero-shot chores in 30 unseen homes; no unseen-home coffee yet, ETA ~2027 · assessed 2026-10-19 · 24 event(s)
- Random and unseen = not chosen by the robot's developer, and not visited, mapped or scanned beforehand.
- One independently verified success suffices; the robot may recover from its own mistakes, but no human help.
No robot has made coffee in a random, previously unseen home; the nearest precursors are Figure's Helix 2.5 at 56% full-task success on three chores in 30 unseen homes, Physical Intelligence's pi0.7 running an espresso machine out of the box, and Sunday's ACT-2 with coffee 'not yet reliable', while a commercial home-cleaning humanoid service (Tau Robotics) still relies on remote operators. Per-home success odds rising at the pace Dyna Robotics' napkin-folding robots have cut their reject rate (an assumed carry-over), the current rate at which labs publish multi-home tests, and a 25% weight on the independent LEAP expert forecast put the ETA in 2027 (80% range 2026-2032, P(by 2100) 97%). Prediction markets (Manifold's Coffee Test) give 33% by end-2026 and 81% by end-2028 against our 13% and 72%, a gap mostly explained by their looser test rules. Yellow, not green, because the precursors are company-reported and independent signals do not corroborate them.- ETA
- ~2027 (80% range 2026–2032; P(by 2100) 97%)
- ETA cruxes
- First, the mapping: whether unseen-home task success improves at the Dyna odds rate (an independent unseen-home benchmark repeated over time measures it; x4 slower: 2028.55). Then p0 and the qualifying campaign rate (whether labs put coffee in third-party-chosen homes with independent verification; 0.7/yr moves p50 to 2028.90, p0 0.10 to 2028.70) and the LEAP weight (0.5: 2028.80; 0: 2027.55), which is also the crux against manifold-za8292n5qc-p and manifold-92rpicoun9-p. Earlier: a coffee demo by Figure, Physical Intelligence, Sunday or DeepMind in third-party unseen homes; an independent benchmark >= 50%. Later: coffee still 'not reliable' in 2027 releases; independent evaluations far below company numbers; home robots shipping on teleoperation (as Tau's service does); the 2026 Manifold market resolving NO with no unseen-home coffee result.
- Path
- Reading used (owner-decided, framework 66c0e8a): one independently verified autonomous success making a cup of coffee with the home's own equipment, in a home not chosen by the robot's developer and not visited, mapped or scanned beforehand; the robot may recover from its own mistakes, no human help; failed homes do not cancel a success; no time limit. Context only (outside the wording): LEAP's three-home no-mistake test (p50 2029.70); pre-mapped or developer-chosen homes; the Manifold anchors' 20-minute limit; teleoperated home services (Tau). Proxy: per-home success on long-horizon chores and appliance tasks in unseen homes (Helix 2.5, pi0.7), which tracks the coffee task because coffee is one such multi-step appliance task in the same homes; its uncertainty is in p0 (0.10-0.45). Path: generalist VLA policies on mobile manipulators and humanoids (Figure Helix 2.5, Physical Intelligence pi0.7, Sunday ACT-2, Stanford/Caltech HomeBody), 1X NEO home humanoid shipping from 2026.
- Blockers
- Technical: appliance diversity and liquid handling in unseen homes; coffee 'not yet reliable' per Sunday
- Corroboration: unseen-home success rates are company-reported; independent benchmarks (BEHAVIOR 12.4%, REALM) show fragile generalization; a commercial home humanoid service (Tau) still runs on teleoperation
- Safety: home deployment (Agility will not enter homes until falls are provably safe)
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- trend (exponential growth of per-home success odds at the shared driver dyna-failure-halving odds rate, mapping assumed) with dated multi-home evaluation campaigns (first success), mixed 75/25 with the independent LEAP forecast remapped to one success; named failure scenarios (each < 10%); Monte Carlo N=10,000 on the driver's canonical draws; compared after the fact with two similar Manifold anchors (no gap)
- ETA calculation
- Per-home coffee success in a random unseen home p0 = 0.25 at 2026.7 (odds lognormal with 10%/90% points at 0.10/0.45), ASSUMED: below Helix 2.5's 56% on three chores in 30 unseen homes (2026-09-17-figure-helix-25-30-homes), above BEHAVIOR's 12.4% (sim, 2025-12-01-behavior-challenge-2025-results); pi0.7 espresso out of the box (2026-04-16-pi07-unseen-kitchens); Sunday coffee not yet reliable (2026-07-16-sunday-act2-laundry-99pct). Failure odds halve every draw i of shared driver dyna-failure-halving odds_halving_months (p50 6.52 mo, draws p10/p90 3.39/11.96; Dyna odds 0.333 -> 0.0753 in 14 mo, 2026-08-27-dyna-din-tai-fung-shift-throughput; mapping ASSUMED). Coffee enters qualifying unseen-home evaluations once p >= 0.35 (assumed): log2(0.54/0.33) = 0.7 halvings -> ~2027.04. Campaigns: MEASURED 4.94/yr unseen-home evaluation releases (7 in 17 mo: 2025-04-22-pi05-unseen-homes-r2, 2025-11-18-pistar06-espresso-13h, 2025-11-19-sunday-act1-unseen-airbnbs, 2026-04-16-pi07-unseen-kitchens, 2026-07-16-sunday-act2-laundry-99pct, 2026-09-17-figure-helix-25-30-homes, 2026-09-27-homebody-gpt6-astra-unseen-kitchen) x ASSUMED 0.5 qualifying = 2.47/yr, 5 homes each (assumed). Trend-only 2026.90/2027.55/2028.90. LEAP (2026-04-20-leap-coffee-test-forecast; superforecasters 2032, experts 2034, midpoint 2033 assumed; three homes without mistakes): 0.794/home, minus trend time log2(3.85/0.54) x 0.543 = 1.55 yr -> ~2031.4 for one success (log-SD 0.49), weight 0.25 (assumed). Failure (assumed, each < 10%): shared stall 0.02; home robots kept out of unscripted appliance use 0.01 (2025-01-01-agility-home-safety-stance-rv) -> p_achieved 0.97. New evidence 2026-10-16-tau-home-cleaning-humanoids-teleoperated (a teleoperated commercial service, not an autonomous unseen-home evaluation): outside the release class, no input changed. MC N=10,000 (made_at 2026.797): conditional 2026.95/2027.80/2031.96; unconditional 2026.95/2027.85/2032.97; P(by end-2026) 0.132, P(by end-2028) 0.720. Pre-anchor p50 2027 (2026-2032, p_a 0.97). Anchors (similar, read after; latest points 2026-10-15): manifold-za8292n5qc-p 0.333 by 2026-12-31 (23 forecasters) vs ours 0.132, diff 0.20; manifold-92rpicoun9-p 0.812 by 2028-12-31 (27 forecasters) vs ours 0.720, diff 0.09; both < 0.35, no gap. Wording explains most of it: the market-event variant (every unseen-home release counts, 4.94/yr; a developer-chosen home lets coffee appear at p 0.2; 20-min limit x0.7 per-home success) gives 0.343 / 0.777; the residual at 2028 (0.035) is our LEAP branch (variant without LEAP: 0.450 / 0.965). Crux vs the markets: weight on independent evidence (LEAP, BEHAVIOR, REALM) vs company unseen-home results. Thin, correlated (same creator and criteria), loosely resolved markets; no move from them. Sensitivities (p50): p0 0.10/0.45 2028.70/2027.30; theta 0.2/0.5 2027.45/2028.30; campaigns 0.7/4.94 per yr 2028.90/2027.60; HomeBody excluded 2027.90; m 1 2028.20; LEAP weight 0/0.5 2027.55/2028.80; LEAP log-SD 0.7 2027.80 (p90 2033.78); LEAP not remapped 2027.85 (p90 2035.11); mapping x2/x4 2028.20/2028.55, fast (task-horizon doubling_software_months 3.75 mo / 7.623) 2027.55; S0 0.08 p_a 0.91; own failure 0.05 p_a 0.931; LEAP's own three-home event (context) 2029.70. Prev (assess:20261015T143612Z) 2026/2027/2032, p_a 0.97, reproduced exactly (2026.94/2027.84/2032.97).
- Data as of
- 2026-10 (Helix 2.5 2026-09; Tau 2026-10)
- Change
- yellow -> yellow. New evidence: 2026-10-16-tau-home-cleaning-humanoids-teleoperated (Tau's paid home cleaning with modified Unitree G1s relies on remote operators). It is not an autonomous unseen-home evaluation, so it does not enter the release class or change p0 (rule 7: no input contradicted); it is added to the 'later' crux and the corroboration blocker. Prev reproduced exactly (2026.94/2027.84/2032.97); rerun at made_at 2026-10-19 on the same canonical driver draws: 2026.95/2027.85/2032.97, p_achieved 0.97. Anchors unchanged (no newer points than 2026-10-15), no gap. No quantile moves; the fast-mapping sensitivity now uses the task-horizon driver (metr-horizon retired). Status yellow.
- Rubric rule
- Milestone yellow (RUBRIC v6 §4): measurable progress on the milestone's own quantity (autonomous long-horizon tasks and appliance use in unseen homes), median ETA 2027 <= made_at + 5 yr, but the precursors are company-reported and lack non-proponent corroboration, so yellow, not green. Not achieved: no verified event shows coffee made in a random unseen home.
by assess:20261019T010556Z · 17 evidence event(s)Evidence and news: 17 cited by the assessment, 7 other event(s)
Cited evidence (17)
- MarketWatch reported on 16 October 2026, after reporter Christine Ji hosted two Tau Robotics humanoids (modified Unitree G1s costing about $55,000 each before Tau's modifications) in her San Francisco apartment, that the robots climbed the stairs to her third-floor apartment with teleoperator assistance, vacuumed, mopped, wiped counters, loaded a dishwasher (struggling to grip a smooth glass and to retrieve a small knife) and slowly folded a towel (about four minutes) and denim pants while remote human operators at Tau's headquarters guided them, with Tau operating eight robots and having logged roughly 160 paid robot hours across nearly 100 cleanings ($30 per robot-hour); Tau's website states its robots are not fully autonomous, a supervising human deciding when to let them do tasks autonomously and when to take over and steer them — i.e. Tau's commercial in-home humanoid cleaning still depends on human help.The "Wozniak" Test [1] Can a $55,000 robot load the dishwasher? Inside my afternoon with two humanoid housekeepers (MarketWatch, 16 Oct 2026) [2] humanoid.guide — Tau Robotics humanoids clean homes for $30 an hour, with help (17 Oct 2026; summary of the MarketWatch report) [3] Tau Robotics — Humanoid Cleaning Service in San Francisco (FAQ: 'Are the robots fully autonomous? No')
- Stanford and Caltech researchers' HomeBody system (reported 27 September 2026) let a Unitree G1 humanoid driven by GPT-6 Astra explore an unfamiliar kitchen, build a digital twin and spatial memory, then tidy it and fetch items from drawers by calling a skill library, with limitations including model latency, overheating finger servos and high compute costs; no success rates were reported.
- Figure AI reported on 17 September 2026 that its Helix 2.5 model, pretrained on Figure's Index dataset of human behavior, ran three long-horizon whole-body tasks (living-room tidying, towel folding, bed making) zero-shot in 30 unseen Bay Area homes with unseen objects, reaching 56% full-task success (no partial credit) versus 9% for the same model without Index pretraining, and matched a Helix 02 policy trained in its evaluation environment using half as much task data.
- Dyna Robotics reported on 27 August 2026 that at Din Tai Fung restaurants, where each robot must produce 1,500 table-ready napkins per 18-hour shift, its Dyna-1 model at launch (June 2025) folded about 35 napkins/hour with 75% meeting the quality bar (about 480 per shift, roughly 32% of the requirement), while Dyna-2 (August 2026) folds 95/hour with 93% passing, about 1,590 per day (106% of the requirement), and that Din Tai Fung is rolling the robots out across its network, with Dyna projecting a deployed fleet of hundreds by H1 2027.The "Blue Collar" ShiftGeneralization [1] Dyna Robotics: Not Just a Model, But a Product [2] Dyna Robotics on X: Din Tai Fung rollout thread
- Sunday Robotics reported on 16 July 2026 that its ACT-2 model achieved 99.1% (±0.3% SE) success on laundry folding across 785 autonomous attempts on 9 garment types in unseen homes with zero per-home adaptation, and that out-of-domain folding success rose from 14% to 100% as pretraining data scaled; it said other household skills its model is learning, including coffee preparation, were not yet reliable enough and had not yet been tested against this standard.
- 1X said on 30 April 2026 that NEO home-humanoid pre-orders booked out its entire production capacity for the next year (10,000 units) within 5 days of the 28 October 2025 launch, that its NEO Factory in Hayward, California had commenced full-scale production, giving 1X (together with a San Carlos facility coming online later in 2026) capacity to produce up to 10,000 NEOs a year, and that the first NEOs would ship to customers in 2026 with some buyers receiving units later.The "Wozniak" Test [1] 1X: NEO Factory | Building Your NEO
- The Forecasting Research Institute's LEAP expert panel (Wave 7, surveyed 9–31 March 2026, published 20 April 2026) gave a median expert forecast of 2034 for a robot first passing the Coffee Test (making coffee in three unfamiliar home-like environments without mistakes), with superforecasters at 2032, the general public at 2039 and an expert interquartile range of 2030–2040.
- Physical Intelligence's π0.7 model (arXiv, 16 April 2026) followed 3–6-step open-ended instruction sequences in 4 unseen kitchens and 2 unseen bedrooms with significantly higher success than π0.5 and π0.6, and operated an espresso machine out of the box at a level matching RL-specialized models; its authors report that seen tasks often exceed 90% success while unseen tasks or task-robot combinations reach 60–80%.The "Wozniak" TestGeneralization [1] π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities (arXiv 2604.15483)↳ follows up on 2025-04-22: Physical Intelligence's π0.5 VLA model (arXiv, 22 April 2025), trained with about 400… (corrected version)
- Figure AI's Helix 02 model (27 January 2026) had a Figure humanoid unload and reload a dishwasher across a full-sized kitchen in a 4-minute fully autonomous run of 61 sequenced loco-manipulation actions with no resets or human intervention, which Figure called the longest-horizon autonomous humanoid task to date; the policy was trained on data from the environment where it was evaluated and no success rate was given.
- Agility Robotics stated that its humanoids will not go into homes until it can prove the robot is not going to fall on a baby.The "Wozniak" Test [1] Agility: From warehouse to house↳ corrects (or adds sources to) an earlier record:
2025: Agility Robotics stated that its humanoids will not go into homes until it can prove… - The REALM benchmark paper (arXiv, December 2025) found that VLA models π₀ and GR00T N1.5 struggle when tasks are perturbed, indicating fragile generalization.Generalization [1] arXiv: REALM benchmark↳ corrects (or adds sources to) an earlier record:
2025-12: The REALM benchmark paper (arXiv, December 2025) found that VLA models π₀ and GR00T… - In Stanford's 2025 BEHAVIOR Challenge (leaderboard published 1 December 2025, NeurIPS 2025), whose 50 long-horizon household tasks are run in photo-realistic simulation, the best of 18 teams achieved only 12.4% full-task success on the held-out test (Q-score 0.26 including partial credit), using a π0.5-based policy.
- Sunday Robotics said on 19 November 2025 that its ACT-1 model, trained on glove-captured human data with zero robot data, cleared dining tables and loaded dishwashers in Airbnb homes it had never seen, given a 3D map of each home, and performed a Table-to-Dishwasher task of 33 unique and 68 total dexterous interactions with 21 objects over more than 130 ft of navigation, and pulled an espresso shot on a consumer machine; no task success rates were disclosed.The "Wozniak" TestGeneralization [1] Sunday Robotics: ACT-1: A Robot Foundation Model Trained on Zero Robot Data
- Physical Intelligence reported on 18 November 2025 that its π*0.6 VLA, improved with the RECAP reinforcement-learning method, made espresso drinks with a commercial espresso machine for 13 hours straight and folded novel laundry in a new home for over two hours without interruptions; on the hardest tasks (diverse laundry, espresso) RECAP more than doubled successful completions per hour and roughly halved failure rates versus the base model.The "Wozniak" TestGeneralization [1] π*0.6: a VLA That Learns From Experience (arXiv 2511.14759) [2] Humanoids Daily: Physical Intelligence claims 'RL is back' with π*0.6↳ follows up on 2025-04-22: Physical Intelligence's π0.5 VLA model (arXiv, 22 April 2025), trained with about 400… (corrected version)
- Physical Intelligence's π0.5 VLA model (arXiv, 22 April 2025), trained with about 400 hours of mobile-manipulator data from about 100 homes plus other robot and web data (97.6% of first-phase examples), performed multi-stage kitchen and bedroom cleaning tasks in 3 real homes not seen in training; in unseen mock homes its success rose with the number of training locations (3 to 104), and the 104-location model performed about as well as a control trained on the test homes.The "Wozniak" TestGeneralization [1] π0.5: a Vision-Language-Action Model with Open-World Generalization (arXiv 2504.16054)↳ corrects (or adds sources to) an earlier record:
2025-04: Physical Intelligence released its π₀.5 vision-language-action model in April 2025,… - University of Edinburgh researchers reported in Nature Machine Intelligence on 19 March 2025 that ELLMER, a GPT-4-based system controlling a seven-degrees-of-freedom Kinova robot arm, made a cup of coffee in a purpose-designed realistic kitchen setting — finding a mug by opening drawers with opening mechanisms it had not encountered before and mixing a set ratio of coffee with kettle water — while adapting to disturbances such as the mug being moved; the demonstration used a single designed setting.
- Figure AI showed on 7 January 2024 its Figure 01 humanoid making a cup of coffee with a Keurig machine using an end-to-end neural network (video in, trajectories out) that, per press coverage, learned from about 10 hours of videos of humans and could self-correct its mistakes.
Other events (7)
- Google DeepMind introduced Gemini Robotics 2 on 30 July 2026, reporting whole-body control of Apptronik's Apollo humanoid with success of 68.4% picking up objects from a table, 45.7% from the floor and 76.3% from a shelf, and 32–92% on multi-finger dexterity tasks such as tying a trash bag (44%), in its own evaluations whose environment novelty was not specified.
- Generalist AI reported on 2 April 2026 that its GEN-1 model, pretrained on over 500,000 hours of physical interaction data and fine-tuned with about 1 hour of robot data per task, averaged 99% success on a set of dexterous tasks where its GEN-0 model (November 2025) averaged 64%, running tasks such as folding t-shirts 86 times in a row without intervention.The "Wozniak" TestGeneralization [1] Generalist AI: GEN-1: Scaling Embodied Foundation Models to Mastery
- In October 2025 Figure AI showed Figure 03 with its Helix VLA picking up novel objects from language commands; Helix controls the full upper body (35 DoF at 200 Hz) and Figure says it can grasp 'virtually any' novel household object.↳ corrects (or adds sources to) an earlier record:
2025-10: In October 2025 Figure AI showed Figure 03 with its Helix VLA picking up novel objects… - 1X Technologies opened pre-orders for its NEO home humanoid robot in October 2025 at $20,000 or a $499 monthly subscription.The "Wozniak" Test [1] 1X: NEO home robot↳ corrects (or adds sources to) an earlier record:
2025-10: 1X Technologies opened pre-orders for its NEO home humanoid robot in October 2025 at… - Google DeepMind introduced Gemini Robotics in March 2025, a multi-embodiment vision-language-action model with reasoning that more than doubles generalization performance versus prior models.↳ corrects (or adds sources to) an earlier record:
2025-03: Google DeepMind introduced Gemini Robotics in March 2025, a multi-embodiment… - +2 older event(s): the section's event log
The "Fukushima" Test: Robot autonomously navigates a disaster zone and performs a critical repair (valve/switch) within human time limits.Progressing: humanoid ran a staged disaster course autonomously; no autonomous valve fix, ETA ~2027 · assessed 2026-10-19 · 18 event(s)
- A realistic staged disaster site counts (unscripted, rubble, degraded comms and lighting; DARPA Robotics Challenge style but autonomous); a real disaster is not required.
In April 2026 X-Humanoid's Tien Kung 3.0 completed Beijing's staged earthquake, flood and fire course fully autonomously, but no autonomous robot has been reported closing the course's emergency valve, and at the same Beijing competition's fall edition on 18 October only 5 of 29 teams ran autonomously and no autonomous humanoid placed (how they did on the valve was not reported). With an assumed navigation and repair success per attempt, a few autonomous attempts per competition, and success odds improving at the pace Dyna Robotics' napkin-folding robots have cut their reject rate (an assumed carry-over), the first qualifying autonomous repair comes ~2027 (80% range 2027-2030, P(by 2100) 95%): the new fall edition adds a second chance each year, but its few autonomous entrants mean fewer attempts per competition. Yellow, not green, because nothing close to an autonomous repair has been reported and no organiser corroborates one.- ETA
- ~2027 (80% range 2027–2030; P(by 2100) 95%)
- ETA cruxes
- First, the venue baseline: how many autonomous humanoids attempt the valve at each Beijing edition and how they score (round-level results; had the fall edition's autonomous humanoids attempted and failed the valve, p50 would move 2027 -> 2028), and whether venues stage the repair under unscripted degraded conditions and time limits. Second, the mapping of the Dyna rate onto disaster-site manipulation (x4 slower: 2029.30). Earlier: an autonomous valve closure at the April 2027 Beijing challenge or a WRS event; a company staged-site demo. Later: autonomous entries shrinking further, autonomous categories dropped, autonomous teams again skipping manipulation.
- Path
- Reading used (the framework does not define 'human time limits'; definition crux open): with no remote control, a robot crosses an unscripted realistic staged disaster site (rubble, degraded comms and lighting) and completes a valve/switch repair within the time limit the venue sets; a measured human time as comparator is a sensitivity. Venues: Beijing E-Town Robot Warrior Challenge, now with spring and fall editions, autonomous groups per morphology and a humanoid valve-closure task; WRS F-REI Plant Disaster Challenge (valve tasks under comm blackouts); DARPA Triage Challenge (autonomy in smoke and darkness); company staged demos also count. Navigation fleets: Spot at Fukushima/Chernobyl/Sellafield, ANYmal (33,000 autonomous inspections, 200+ deployments). The repair is tied to unseen-environment manipulation because that is the binding capability.
- Blockers
- Technical: autonomous manipulation under rubble, smoke and degraded comms not yet shown; neither Beijing edition reported an autonomous valve closure
- Venue: competitions are teleop-dominated (spring 30 remote vs 8 autonomous rounds; fall 24 remote vs 5 autonomous teams, no autonomous humanoid placing)
- Real sites: Fukushima Daiichi interventions remain teleoperated (not required by the wording)
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (exponential growth of joint navigation+repair success odds at the shared driver dyna-failure-halving odds rate, mapping assumed) evaluated attempt by attempt at the discrete venue cadence (Beijing spring and fall editions, other autumn venues) plus staged company demos (the fall-2026 valve outcomes are unreported and not used); named failure scenarios (each < 10%); Monte Carlo N=10,000 on the driver's canonical draws; reference-class and prev checks
- ETA calculation
- Navigation: at Beijing E-Town's Robot Warrior Challenge (Apr 2026, 2026-04-18-beijing-robot-warrior-challenge-tien-kung-autonomous; 17 obstacles incl. emergency valve closure, 12-min humanoid limit, 8 autonomous-navigation rounds) Tien Kung 3.0 completed the course autonomously; the other rounds' outcomes are not reported, nor any valve closure. p_nav per attempt ~ Beta(2,8) (mean 0.2): ASSUMED prior. Repair given navigation 0.4 (0.15-0.7), ASSUMED. Joint failure odds halve every draw i of shared driver dyna-failure-halving odds_halving_months (p50 6.52 mo, 2026-08-27-dyna-din-tai-fung-shift-throughput; spread assumed; mapping ASSUMED). Fall edition (18 Oct 2026, 2026-10-18-robot-warrior-fall-no-autonomous-humanoid-placing: ~6,500 m2 outdoor course, 14 tasks incl. valve closure for humanoids, wet surfaces; 24 remote vs 5 autonomous teams across humanoid/quadruped/wheeled groups; full placings name one autonomous winner, a wheeled quadruped, and no autonomous humanoid): MEASURED that an autumn edition with a valve task runs and that autonomous entries are few. Inputs changed (new evidence): autumn venue qualifies 0.4 -> 0.6/yr (assumed: Beijing's fall edition recurs and qualifies, plus irregular WRS); attempts per venue n_eff 4 (2-8) -> 2 (1-5) (assumed from 5 autonomous teams in 3 groups and 8 autonomous rounds in spring; a different unit, so no decline is measured). Valve attempts and outcomes at the fall edition are not reported, and the source does not say a successful repair would have placed: UNKNOWN, not a failure, so the base is not conditioned on them. Opportunities: Beijing each April 0.7, autumn 0.6, staged demos 0.3/yr (assumed). Failure (assumed, each < 10%): shared stall 0.02; no venue or demo ever stages and publishes a qualifying autonomous repair 0.03 -> p_achieved 0.95. MC (N=10,000 driver draws, made_at 2026.797) conditional 2027.30/2027.80/2029.80; unconditional p10 2027.30 / p50 2027.80 / p90 2030.30 (the median falls at the autumn-2027 venue, one venue before April 2028; point: joint odds 0.087 at Apr 2026 -> 29% at Apr 2027, 36% at autumn 2027, cumulative 55%); P(<=2031.80) 0.939. Steps: prev inputs at the new made_at 2027.30/2027.80/2030.25; + autumn 0.6 2027.30/2027.80/2029.80; + n_eff 2 2027.30/2027.80/2030.30 = base. Sensitivities (p50): [not base] fall-2026 unreported attempts treated as failures, P(a=0..3) 0.25/0.35/0.25/0.15 2028.30 (a=3: 2028.30, p90 2030.80); n_eff 4 (prev) 2027.80, n_eff 1 2028.30; autumn 0.4 2028.30; Beijing spring 0.4 2028.30; p_nav Beta(1,9) 2028.30, Beta(4,6) 2027.80; repair 0.15/0.7 2028.30/2027.80; demos 1/yr 2027.80; human-time comparator (nav x0.3) 2028.80; mapping slow x2/x4 2028.30/2029.30 (p90 2032.30/2035.35), fast (task-horizon doubling_software_months 3.75 mo / 7.623) 2027.65; own failure 0.10 -> p_achieved 0.882, p90 null. Reference class: DRC 2015 supervised valve turning -> first fully autonomous staged-course completion 2026 (~11 yr); the remaining step is one manipulation task inside an autonomous run. No proponent dates; no anchors. Prev (assess:20261015T100057Z) 2027/2027/2030, p_achieved 0.95, reproduced exactly (2027.29/2027.79/2030.29).
- Data as of
- 2026-10 (Beijing fall edition 2026-10-18; driver Dyna 2026-08)
- Change
- yellow -> yellow. New evidence: 2026-10-18-robot-warrior-fall-no-autonomous-humanoid-placing (Beijing's fall edition: valve task for humanoids, 5 of 29 teams autonomous, no autonomous humanoid placing). Prev reproduced exactly (2027.29/2027.79/2030.29); prev inputs at the new made_at 2027.30/2027.80/2030.25. Inputs changed for this evidence (rule 7 (a)): autumn venue 0.4 -> 0.6/yr (p90 2030.25 -> 2029.80), attempts per venue 4 (2-8) -> 2 (1-5) (p90 -> 2030.30). The fall edition's valve outcomes were not reported, so they are not used. Moves: none at calendar resolution (p10/p50/p90 2027/2027/2030 as prev; decimal 2027.30/2027.80/2030.30); p_achieved 0.95 unchanged. The fast-mapping sensitivity now uses the task-horizon driver. Status yellow.
- Rubric rule
- Milestone yellow (RUBRIC v6 §4): measurable progress (autonomous completion of a staged disaster course within the time limit; unseen-environment manipulation success rising), median ETA 2027 (2027.80) <= made_at + 5 yr, but no autonomous valve/switch repair precursor and no organiser corroboration, so yellow, not green.
by assess:20261019T010556Z · 16 evidence event(s)Evidence and news: 16 cited by the assessment, 2 other event(s)
Cited evidence (16)
- At Beijing E-Town's 2026 Robot Warrior Challenge fall competition on 18 October 2026 — a ~6,500 m² outdoor staged emergency course of 14 tasks, run on surfaces left wet by the previous day's heavy rain, whose humanoid-only tasks were pulling open a door, explosive-hazard removal, emergency valve closure and fire extinguishing — 24 of the 29 teams competed by remote control and 5 by autonomous navigation, and the published results named placings in the remote-controlled humanoid, quadruped and wheeled-quadruped groups but only one autonomous winner (Zhishen 'Milk long B' in the autonomous wheeled-quadruped group), reporting no autonomous humanoid placing, whereas X-Humanoid's Tien Kung 3.0 had won the autonomous humanoid run of the spring edition in April.The "Fukushima" Test [1] Beijing Youth Daily (北青网): 14项实景关卡、29支赛队 2026机器人勇士挑战赛在北京亦庄举办 (18 Oct 2026; 24 remote vs 5 autonomous teams, full placings) [2] Beijing Daily (京报网): 机器人也“过五关斩六将”？这场面太燃了 (18 Oct 2026; humanoid tasks incl. 紧急关阀, group results) [3] Beijing Economic-Technological Development Area (organiser): 2026机器人勇士挑战赛秋季赛即将举办 (9 Oct 2026; 14 tasks, remote/autonomous × humanoid/quadruped/wheeled-quadruped groups)
- On 2 October 2026 TEPCO began inserting a remote-controlled multi-joint robotic arm — under development since 2017, about ¥7.8 billion in government funds, and repeatedly delayed — into Fukushima Daiichi Unit 2's containment vessel, planning 3–4 months of obstacle clearing and 3D surveying before a third trial fuel-debris retrieval next year, after two earlier trials (2024, 2025) recovered about 0.9 g combined.The "Fukushima" Test [1] Mainichi (Kyodo): TEPCO begins inserting robot arm to retrieve Fukushima nuke fuel debris [2] Kyodo News: TEPCO begins inserting robot arm to retrieve Fukushima nuke fuel debris
- Launching its Shift fleet-management platform on 1 October 2026, ANYbotics said Shift is already used across more than 200 robot deployments and runs hundreds of thousands of autonomous inspections per month (company-reported; autonomous navigation and sensing, not manipulation or repair).
- Figure AI reported on 17 September 2026 that its Helix 2.5 model, pretrained on Figure's Index dataset of human behavior, ran three long-horizon whole-body tasks (living-room tidying, towel folding, bed making) zero-shot in 30 unseen Bay Area homes with unseen objects, reaching 56% full-task success (no partial credit) versus 9% for the same model without Index pretraining, and matched a Helix 02 policy trained in its evaluation environment using half as much task data.
- Dyna Robotics reported on 27 August 2026 that at Din Tai Fung restaurants, where each robot must produce 1,500 table-ready napkins per 18-hour shift, its Dyna-1 model at launch (June 2025) folded about 35 napkins/hour with 75% meeting the quality bar (about 480 per shift, roughly 32% of the requirement), while Dyna-2 (August 2026) folds 95/hour with 93% passing, about 1,590 per day (106% of the requirement), and that Din Tai Fung is rolling the robots out across its network, with Dyna projecting a deployed fleet of hundreds by H1 2027.The "Blue Collar" ShiftGeneralization [1] Dyna Robotics: Not Just a Model, But a Product [2] Dyna Robotics on X: Din Tai Fung rollout thread
- ANYbotics reported in a June 2026 case study that its ANYmal quadruped at Vigier Ciment's 24/7 cement plant in Péry, Switzerland has run fully autonomous nightly inspection missions across 3 mills and 6 floors since January 2025, completing more than 33,000 inspections at 450 inspection points in 16 months with zero robot breakdowns (company-reported; sensing-only inspection, no manipulation or repair).The "Fukushima" Test [1] ANYbotics: How Vigier (Vicat Group) protects uptime with autonomous inspection
- At Beijing E-Town's first Robot Warrior Challenge on 18 April 2026 — a 3,500 m² staged earthquake/flood/fire course of 17 obstacles including emergency valve closure, with a 12-minute limit for humanoids and 36 teams running 30 remote-controlled versus 8 autonomous-navigation rounds — X-Humanoid's full-size Tien Kung 3.0 completed the course fully autonomously with no remote control and, per X-Humanoid, achieved the highest overall score to become the first winner of the fully autonomous run (Warrior Intelligent Mobility Award), clearing pendulum traversal, forward progression and barrier-breaching/obstacle-clearance tasks, though neither the organiser nor the company reports it completing the valve-closure task.The "Fukushima" Test [1] Beijing Economic-Technological Development Area (organiser): 2026 Robot Warrior Challenge opens with 17 real-scene obstacles [2] X-Humanoid (PR Newswire): Embodied Tien Kung 3.0 wins Beijing Robot Warrior Challenge with fully autonomous run [3] CCTV+: Robot emergency rescue challenge held in Beijing
- Physical Intelligence's π0.7 model (arXiv, 16 April 2026) followed 3–6-step open-ended instruction sequences in 4 unseen kitchens and 2 unseen bedrooms with significantly higher success than π0.5 and π0.6, and operated an espresso machine out of the box at a level matching RL-specialized models; its authors report that seen tasks often exceed 90% success while unseen tasks or task-robot combinations reach 60–80%.The "Wozniak" TestGeneralization [1] π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities (arXiv 2604.15483)↳ follows up on 2025-04-22: Physical Intelligence's π0.5 VLA model (arXiv, 22 April 2025), trained with about 400… (corrected version)
- Boston Dynamics' Spot robot is in ongoing inspection deployments at the Fukushima Daiichi, Chernobyl and Sellafield nuclear sites.The "Fukushima" Test [1] Boston Dynamics: Spot in Fukushima Daiichi↳ corrects (or adds sources to) an earlier record:
2025: Boston Dynamics' Spot robot is in ongoing inspection deployments at the Fukushima… - Equinor deployed ANYbotics' ANYmal quadruped at the Northern Lights carbon capture and storage facility in 2025 for CO₂ monitoring.↳ corrects (or adds sources to) an earlier record:
2025: Equinor deployed ANYbotics' ANYmal quadruped at the Northern Lights carbon capture and… - At the World Robot Summit 2025 Harsh Environment F-REI Challenge at the Fukushima Robot Test Field (10–12 October 2025; 34 teams from eight countries and regions across four competitions), the real-robot Plant Disaster Challenge — gauge reading and valve opening/closing on digital-twin command, abnormality response under smoke, and debris removal plus gas-valve closure after a simulated boiler explosion with communication delays or blackouts — was won by team NuTech-R, ahead of Quix and MISORA+UoA.The "Fukushima" Test [1] WRS2025 Harsh Environment F-REI Challenge: Closing & Competition Results Flash Report [2] WRS2025 F-REI Challenge: Plant Disaster Challenge missions [3] WRS2025 Harsh Environment F-REI Challenge: results announcement [4] JSME Journal: Report on the WRS2025 Harsh Environment F-REI Challenge
- DARPA Triage Challenge Event 2 (Systems competition, Perry, Georgia, 26 Sep–3 Oct 2025, concluded 4 Oct) required fully autonomous UAV/UGV systems to geolocate and triage simulated casualties on C-130-crash and night-ambush courses with smoke, darkness and obstructions; among 9 teams, DARPA-funded DART scored highest (728 points: 536 on the C-130 course plus 192 at night) and self-funded Coordinated Robotics (546) won the $300,000 top prize, with scores not comparable to Event 1's different scoring scheme.The "Fukushima" Test [1] DARPA: DART and MSAI triumph at DARPA Triage Challenge [2] DARPA Triage Challenge: Challenge Events (Event 2 Systems scores) [3] DARPA Triage Challenge Event 2 Systems competition rules (full autonomy, degraded sensing)↳ corrects (or adds sources to) an earlier record:
2025-10: DARPA held Event 2 of its Triage Challenge in October 2025, in which robots navigated… - Boston Dynamics and Toyota Research Institute demonstrated the Atlas humanoid controlled by Large Behavior Models in August 2025.The "Fukushima" TestGeneralization [1] Toyota: Boston Dynamics and TRI Atlas [2] Wayback capture 20251227040037↳ corrects (or adds sources to) an earlier record:
2025-08: Boston Dynamics and Toyota Research Institute demonstrated the Atlas humanoid… - In July 2025 the Nuclear Damage Compensation and Decommissioning Facilitation Corp. (NDF), which supervises Fukushima Daiichi decommissioning, said full-scale removal of the estimated 880 tonnes of fuel debris in Units 1–3, previously planned for the early 2030s, is likely to start only in (fiscal) 2037 or later, as TEPCO reportedly concluded that preparations alone would take about 12–15 years; only small trial samples had been retrieved from Unit 2, in November 2024 and April 2025.
- The UK Nuclear Decommissioning Authority launched the £9.5 million Auto-SAS robotics partnership in June 2025 to develop autonomous sorting of nuclear waste.The "Fukushima" Test [1] GOV.UK: NDA launches robotics partnership↳ corrects (or adds sources to) an earlier record:
2025-06: The UK Nuclear Decommissioning Authority launched the £9.5 million Auto-SAS robotics… - In DARPA Triage Challenge Event 1 (Perry, Georgia, 28 Sep–4 Oct 2024), 11 teams fielded autonomous UAV/UGV systems to find and triage simulated casualties on three staged courses (airplane crash, convoy ambush, open battlefield); the top Systems score was DART's 92.5 points (mean of its best two of three runs, at up to 15 points per casualty), while 6 of the 11 teams scored zero on at least two of their three runs.
Other events (2)
- TEPCO retrieved only about 0.9 g of fuel debris in total from Fukushima Daiichi in 2024–2025, against roughly 880 tonnes remaining, with the retrieval robots teleoperated.The "Fukushima" Test [1] Nippon.com: Fukushima fuel debris retrieval↳ follows up on 2025-04: TEPCO carried out a second Fukushima Daiichi fuel debris retrieval in April 2025,… (since withdrawn)
- TEPCO retrieved its first fuel-debris sample from Fukushima Daiichi in November 2024; the sample weighed approximately 0.7 grams.The "Fukushima" Test [1] Nippon.com: Fukushima fuel debris retrieval↳ corrects (or adds sources to) an earlier record:
2024-11: TEPCO retrieved the first fuel debris sample (about 0.19 g) from the Fukushima Daiichi…
The "Blue Collar" Shift: A general-purpose robot works full shifts in a real workplace (construction, factory, logistics) alongside humans for a month.Progressing: humanoids on real lines 8-24h/day for days-months; interventions undisclosed, ETA ~2027 · assessed 2026-10-19 · 39 event(s)
- 8h+ shifts on 30 consecutive working days, at ≥90% of the job's required throughput.
- No teleoperation; a disclosed average of ≤1 human intervention per shift.
Shift length, multi-month runs and throughput are each shown somewhere (AgiBot G2s on 10-hour days at Longcheer, Galbot's claimed 3+ months of 24/7 work at CATL, 15 Leju humanoids in routine work on a Zhaofeng parts line, Robotera at 1,200 parcels an hour), but no deployer has disclosed intervention counts or teleoperation status, and a Beijing News survey in October 2026 found sustained shift performance unproven. Built from company-claimed hardware reliability, Longcheer's measured task-failure rate and assumed assists, the best site is estimated at about 1.1 human interventions a shift; with failures falling at the pace Dyna Robotics' napkin-folding robots have cut their reject rate (an assumed carry-over), a 30-day run and an assumed disclosure lag, the ETA is late 2027 (80% range 2027-2030, P(by 2100) 95%). Yellow, not green, because the numbers that decide it, interventions per shift and whether the robots are remotely operated, are undisclosed or company claims, with no independent corroboration.- ETA
- ~2027 (80% range 2027–2030; P(by 2100) 95%)
- ETA cruxes
- First, the mapping: whether intervention rates fall like Dyna's reject fraction (dated intervention-hour logs settle it; a 30-month halving moves p90 to 2034). Then the disclosure lag (2 yr instead of 0.75 moves p50 +1.3 yr). Earlier: AgiBot, Galbot/CATL, Leju/Zhaofeng, Robotera, Figure, Agility or Dyna publishing a month-long intervention log, or a customer/auditor confirming <=1 intervention per shift without teleop. Later: disclosed rates >>1/shift, confirmed teleoperation, deployments rolled back (as Figure at BMW).
- Path
- Reading used: a general-purpose robot works 8h+ shifts on 30 consecutive working days in a real workplace alongside humans at >=90% of required throughput, no teleoperation, with a disclosed average <=1 human intervention per shift, where an intervention is any human action needed to keep it working (fault recovery, reset, tool/hand swap, assist, by an operator or maintenance team); self-performed charging/battery swaps are not interventions. Path: Chinese real-line deployments (AgiBot G2 at Longcheer, 8 robots, whole MMIT section, 99.99% task success; Galbot S1 at CATL 7x24 for 3+ months, claim; 15 Leju Kuavo at Zhaofeng, routine operation, 98% loading success but ~40% of the cycle rate; Robotera at China Post 1,200/h), Agility Digit (65,000 h across customer sites), Figure 03 (200 h continuous run at its own HQ).
- Blockers
- Disclosure gap: no deployer publishes achieved interventions per shift or teleoperation status; an independent survey (Beijing News, Oct 2026) says takeovers and fault downtime are what must be shown (gather request: intervention-hour logs)
- Credibility: independent reporting shows overstatement (Forbes Oct 2026: one Figure robot at a time at BMW, 2026 test ended; Robotera needed months of on-site support incl. night supervision; Longcheer relies on a maintenance team)
- Technical: hour-level MTBF typical for humanoids; UBTech Walker S2 at 30-50% of human productivity, Leju at Zhaofeng ~40% of the line's cycle; joint wear under three-shift duty
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- trend (exponential decline of a component-built intervention rate at the shared driver dyna-failure-halving, mapping assumed) plus the literal 30-day run floor and an assumed disclosure lag; named failure scenarios (each < 10%); Monte Carlo N=10,000 on the driver's canonical draws, other components shared with Dark Line; proponent, reference-class and prev cross-checks
- ETA calculation
- Gap: I0 = human interventions per 8-h robot-shift at the best real-line site (mid-2026) vs <=1. Components: hardware 8/MTBF, MTBF 16 h (5-60; centre sourced company claims, 2026-04-14-agibot-humanoid-mtbf-vs-industrial 'two-shift' level, 2026-03-24-agibot-mtbf-hundred-hours upper tail; spread assumed) -> 0.5/shift; task failures 0.133/shift measured (Longcheer 55,107 ops at 99.987% = 7.16 failures over 8 robots x 54 h, 2026-06-28-agibot-longcheer-whole-section-8-robots; forecast 0.15 (0.08-0.5) assumed; robot-hour reading 1.06); share needing a human 0.5 (0.2-1, assumed); other assists 0.3/shift (0.05-1.5, assumed) -> I0 median 1.13 (0.45-3.19), P(<=1) 0.437. Capability = 2026.5 + max(0, log2 I0) x T, T = draw i of shared driver dyna-failure-halving halving_months (p50 7.623 mo measured from 2026-08-27-dyna-din-tai-fung-shift-throughput, reject fraction 25% -> 7% in 14 mo; spread assumed; draws 3.96/7.64/13.98 mo; mapping onto interventions ASSUMED) + 30 working days 0.12 yr (literal floor) + disclosure lag 0.75 yr (0.25-2.5, assumed) -> point 2026.5 + 0.18 x 0.635 + 0.12 + 0.75 = 2027.5. Failure (assumed, each < 10%): shared robotics stall 0.02; no verifiable month-long log ever published or every run disputed 0.03 -> p_achieved 0.951. MC (N=10,000 driver draws, made_at 2026.797) conditional 2027.00/2027.78/2029.65; unconditional p10 2027.01 / p50 2027.84 / p90 2030.56; P(<=2031.80) 0.929. New evidence, no input changed: Zhaofeng (2026-10-16-leju-kuavo-zhaofeng-parts-loading-r2: 15 Leju robots, ~1,000 cumulative running hours over 2+ months 'without failure', >98% loading success, no teleop, company data; 2026-10-18-zhaofeng-kuavo-cycle-gap-lights-out-goal: 35-40 s per load vs 15 s for a skilled worker, ~40% of the line's cycle) is an ambiguous company MTBF claim (hours basis and 'failure' undefined) from a site far below the 90% throughput clause, so it enters as a sensitivity (MTBF 300 h (100-1,000): I0 0.46, p50 2027.58); 2026-10-14-bjnews-shift-performance-unproven (independent: takeovers and fault downtime undisclosed) supports the disclosure lag; 2026-10-09-dyna-din-tai-fung-1000-napkins-per-day (1,000-1,100/day per restaurant vs ~1,590 per 18-h shift in August) is a driver-source update, not refitted here (owner finding). Sensitivities (p50): I0 1.5 (0.4-6) 2028.09; I0 3 (1-10) 2028.62; lag 0 / 2 yr 2027.18 / 2029.15; mapping slow x2 2028.13, x4 (~30 mo, Agility) 2028.58 (p90 2034.75); fast (task-horizon driver doubling_software_months p50 3.75 mo / 7.623) 2027.64; own failure 0.10 -> p_achieved 0.881, p90 null. Proponents, remaining time from made_at x1.5-3: BofA mass adoption from 2028 -> 2028.6-2030.4. Reference class: one live-production chore (Apr 2025) -> multi-robot real-line logs (Jun 2026) in ~15 mo. Prev (assess:20261015T100057Z) 2027/2027/2030, p_achieved 0.951, reproduced exactly (2027.01/2027.84/2030.54).
- Data as of
- 2026-10 (baseline Longcheer 2026-06; driver Dyna 2026-08; Zhaofeng 2026-10)
- Change
- yellow -> yellow. New evidence: 2026-10-16-leju-kuavo-zhaofeng-parts-loading-r2 and 2026-10-18-zhaofeng-kuavo-cycle-gap-lights-out-goal (15 humanoids in routine work, company MTBF claim, ~40% of the line's cycle), 2026-10-14-bjnews-shift-performance-unproven (independent: sustained performance unproven), 2026-10-09-dyna-din-tai-fung-1000-napkins-per-day (driver source update). None contradicts an input (rule 7): the Zhaofeng MTBF claim is ambiguous company data from a site below the throughput clause (sensitivity, p50 2027.58), the survey supports the disclosure lag, the Dyna figure bears on the shared driver, which is not refitted here. Prev reproduced exactly (2027.01/2027.84/2030.54); rerun at made_at 2026-10-19 on the same canonical driver draws: 2027.01/2027.84/2030.56, p_achieved 0.951. No quantile moves; the fast-mapping sensitivity now uses the task-horizon driver (metr-horizon was retired). Status yellow.
- Rubric rule
- Milestone yellow (RUBRIC v6 §4): measurable progress on the milestone's own quantity (8h+ real-line shifts, multi-day to multi-month runs, throughput at or above requirement at some sites), median ETA 2027 <= made_at + 5 yr, but the binding metrics (interventions/shift, no teleop) have no non-proponent corroboration (an independent survey, Beijing News Oct 2026, calls sustained performance unproven), so yellow, not green. Not achieved: no verified event discloses an average of <=1 intervention per shift over 30 consecutive working days.
by assess:20261019T010556Z · 29 evidence event(s)Evidence and news: 29 cited by the assessment, 10 other event(s)
Cited evidence (29)
- Cailian Press reported on 18 October 2026, after a visit to Zhaofeng's Hangzhou auto-parts plant, where 15 wheeled Leju Kuavo humanoids load parts at 15 stations on workers' instructions, that each robot load took 35–40 s versus about 15 s for a skilled worker; Leju's CEO said accuracy and MTBF were already up to standard and the next step was to cut load time to the 15 s that matches the production cycle, which Zhaofeng's chairman expected in six months to a year; Zhaofeng is also exploring having robots do the pre-shift recalibration of a 26–28-station line at model changeovers (now 2–3 hours of work by three engineers) and is testing robot camera inspection toward 'lights-out factory' closed-loop operation (company statements; humans still work on the line, and Jiemian quoted Zhaofeng's general manager on 16 October as saying the robots fill the manual gaps between automated processes rather than replacing the whole automated line).The Dark LineThe "Blue Collar" Shift [1] 财联社: 人形机器人“上工”记 (18 Oct 2026) [2] 界面新闻 (Tencent News): 人形机器人开始进入汽车产线，但工厂要的不只是会干活 (16 Oct 2026)
- 21st Century Business Herald reported on 16 October 2026, after a site visit on 15 October, that 15 Leju humanoid robots were in routine operation on Zhaofeng Mechanical & Electronic's automotive-parts production line in Hangzhou, autonomously completing 10+ consecutive actions per cycle to fetch, transfer and load small parts onto loading stations in under 40 seconds with no human teleoperation, with the solution adaptable to 15 stations and covering 10+ part types, and a company-stated loading success rate above 98%, about 1,000 cumulative running hours over more than two months without failure and 60,000+ parts produced (company-provided data; the project is billed as an embodied-AI 'real-scene practical-training demonstration project').KPIThe "Blue Collar" Shift [1] 21世纪经济报道: 实拍乐聚人形机器人进厂上料，40秒内完成，全程无需遥操作 (16 Oct 2026) [2] 网通社 (ifeng): 乐聚机器人与兆丰机电联合发布“具身智能+汽车零部件上料解决方案” (16 Oct 2026; 15 Kuavo robots)↳ corrects (or adds sources to) an earlier record:
2026-10-16: 21st Century Business Herald reported on 16 October 2026, after a site visit on 15… - A Beijing News (新京报) survey of Chinese humanoid factory deployments published on 14 October 2026 (SAIC-GM/AgiBot's 'Nengzai No. 1', AgiBot G2 at Longcheer, Galbot S1 at CATL) concluded that the applications disclosed so far show some robots can take on specific tasks but that whether this holds over longer periods and more complex conditions still needs sustained-operation testing, warning that a rated runtime such as Galbot S1's official 8 hours with hot-swappable batteries is not the same as continuously completing production tasks, so line suitability must be judged from effective work, fault downtime and human takeovers during operation rather than battery life, and noting that factory training, line trials and routine production still coexist (e.g. Dongfeng's 'Xiaodong' humanoid was planned to enter its plant for practical training at end-October 2026 and join trial production by year-end).
- Dyna Robotics co-founder and chief scientist Jason Ma said at a Savant VC event in San Francisco on 9 October 2026 that a Dyna robot folds between 1,000 and 1,100 napkins per day at each Din Tai Fung restaurant, running the same unified model at every location with no site-specific engineering, and that Dyna robots also handle about 1,000 towels per day at Best Western (company statements reported by humanoid.guide; no intervention or uptime figures given).
- Forbes reported in October 2026 (Forbes Australia syndication, 7 October) that BMW said only one Figure robot worked at any given time during the 2025 Spartanburg pilot (a second was onsite to rotate in while it recharged), that the project Figure began at the plant in June 2026 with updated robots was a non-production test of two or three robots that has ended with no Figure robots currently operating at the plant, and that Figure sent five to ten people to ensure its promotional video shoot went smoothly.KPIThe "Blue Collar" Shift [1] Forbes Australia (syndication): Brett Adcock sold investors a humanoid future — now he has to build it↳ updates 2025-11-19: Figure AI reported on 19 November 2025, retiring Figure 02, that its 11-month… (corrected version)
- Unveiling Digit 5 on 15 September 2026, Agility Robotics said Digit 4 had logged more than 65,000 operating hours across customer sites including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada over nearly three years and kept approximately 98% accuracy while on-task through its 100,000-tote GXO milestone, and claimed Digit 5's 90-minute battery with 9-minute charging gives a 10:1 run-to-charge ratio (vs 2:1 for Digit 4) enabling 20+ productive hours per day, with early access in H1 2027.KPIThe "Blue Collar" ShiftPower/Efficiency/Cost [1] Agility Robotics: Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale↳ updates 2025-11-20: Agility Robotics reported on 20 November 2025 that its Digit humanoids had moved over… (corrected version)
- Dyna Robotics reported on 27 August 2026 that at Din Tai Fung restaurants, where each robot must produce 1,500 table-ready napkins per 18-hour shift, its Dyna-1 model at launch (June 2025) folded about 35 napkins/hour with 75% meeting the quality bar (about 480 per shift, roughly 32% of the requirement), while Dyna-2 (August 2026) folds 95/hour with 93% passing, about 1,590 per day (106% of the requirement), and that Din Tai Fung is rolling the robots out across its network, with Dyna projecting a deployed fleet of hundreds by H1 2027.The "Blue Collar" ShiftGeneralization [1] Dyna Robotics: Not Just a Model, But a Product [2] Dyna Robotics on X: Din Tai Fung rollout thread
- In a Xinhua survey of Zhejiang factory deployments published 23 July 2026, Aitu Embodied Intelligence general manager Lou Lingtong said humanoids can now do single processes well but cannot yet run whole processes unmanned, projecting 2026–2028 as the validation period for single-process capability and 2028–2030 as the scale-up period for multi-process collaboration, while other interviewees said there is no certification system for the 'industrial-grade lifetime' reliability of robots and their components (firms validate by repeated on-site testing) and that fine factory manipulation places very high demands on fatigue-resistant design (structural durability, thermal management, power optimisation) and high-load continuous operation.The Dark Line [1] 从“上场”到“进厂”——“人形机器人量产元年”新观察 (Xinhua, 23 Jul 2026)
- AgiBot said on 29 June 2026 that during a six-consecutive-day livestream from a tablet mass-production quality-inspection section at Longcheer's Nanchang factory, multiple AgiBot humanoids logged more than 64 hours of operation, performed 64,828 production-line tasks across more than 4 production workflows alongside human operators with a 99.99% task success rate, and contributed to cumulative line output of 17,625 (company figures; interventions and teleoperation not quantified).The "Blue Collar" Shift [1] AGIBOT: AGIBOT Concludes Six-Day Real Factory Livestream
- Xinhua reported on 28 June 2026 that eight AgiBot G2 humanoids had taken over Longcheer's entire MMIT tablet-testing section in Nanchang (up from four robots at single stations in April), running 10 hours a day for six days (23–28 June) on a public livestream and logging, by day five, over 54 cumulative robot-hours, 14,925 units, 55,107 operations and a 99.987% success rate, while Longcheer said scaling from four to eight robots surfaced robot-to-robot collisions, arm clashes and lidar frame drops and that the line relies on a supporting maintenance team (day shift only, not 24/7; humans on the floor).
- Galbot and CATL reported in late June 2026 that a Galbot S1 heavy-duty humanoid (50 kg dual-arm payload, up to 8 hours of continuous operation on CATL batteries, vision-only positioning with 360° obstacle avoidance) had worked 7×24 hours on CATL's mass-production lines at its Ningde base for over three months since passing acceptance in March 2026, autonomously carrying heavy material bins between module and battery-pack stations amid human workers without remote control, described as the world's first 'routine autonomous operation' of a humanoid in a new-energy smart-manufacturing production setting (company/partner claim; robot count, uptime, interventions and battery-maintenance arrangements not disclosed).The Dark Line [1] First Heavy-Duty Humanoid Robot Powered by CATL Batteries Goes Live in CATL Factory (CATL, 24 Jun 2026) [2] Galbot's Galbot S1 begins operations at CATL's factory (Gasgoo, 26 Jun 2026) [3] 全球首次！北京人形机器人实现在新能源智能制造生产场景“常态化自主作业” (Beijing International Sci-Tech Innovation Center, 30 Jun 2026)
- Cheetah Mobile CEO Fu Sheng was reported on 2 June 2026 as saying that, by industry standards, harmonic reducers in humanoid joints last about 7,000–10,000 hours at rated load, so core joints would need replacement after 3.5–5 years on single-shift duty (~2,000 h/year) but after only about 1.5–2 years in three-shift operation; the report noted core-component wear accelerates sharply in 24-hour three-shift logistics/manufacturing use and that service life is a key factor customers weigh before deploying humanoids.
- The Economic Observer reported on 2 June 2026 that Robotera's humanoids (column-based lower body) at a China Post logistics centre in Longgang, Shenzhen fell from about 600 parcels/hour in the lab to about 100/hour on arrival (November 2025), reached about 900/hour by March 2026 and a stable 1,200/hour by end-May 2026 versus a human average above 1,000/hour (skilled maximum 1,200), after six months of on-site engineering support including night-shift supervision, with some of its 10+ logistics centres running robots on two to three shifts for 24-hour coverage and new sites now deployed in about 3 days with 2 hours of data.The Dark LineThe "Blue Collar" Shift [1] 机器人同事，开始撑起物流中心的夜班 | 百亿具身系列报道 (Economic Observer) [2] 星动纪元携手中国邮政，助力物流智能化升级 (Xinhua, 20 May 2026)
- Figure AI ended on 22 May 2026 an unedited livestream that began as a response to a challenge to prove humanoids could work an 8-hour shift at human speed without intervention, in which a rotating fleet of Figure 03 humanoids running Helix-02 sorted 249,560 packages over 200 continuous hours at its Sunnyvale headquarters, autonomously rotating units to charge on roughly 4-hour batteries; CEO Brett Adcock said it 'ran for 200 hours without a failure', which Humanoids Daily and Interesting Engineering describe as no catastrophic mechanical breakdown or system-halting crash, though viewers saw occasional dropped or mis-oriented packages (company facility, not a customer workplace; intervention counts not published).The "Blue Collar" Shift [1] Brett Adcock on X: We just wrapped what began as an 8-hour challenge — and it ran for 200 hours without a failure [2] Humanoids Daily: Figure AI pops champagne as autonomous marathon crosses 200 hours without hardware failure [3] Interesting Engineering: Figure's humanoid robots work for 200 hours, process 250k packages without failure
- In a 10-hour 'Man vs. Machine' parcel-sorting contest at Figure AI's own facility, held during its May 2026 livestream marathon and reported on 18 May 2026, a rotating fleet of autonomous Figure 03 humanoids running Helix-02 sorted 12,732 packages (2.83 s/package) versus 12,924 (2.79 s/package) for a human intern who took California-mandated meal and rest breaks, i.e. about 98.5% of the human's shift output on this single task (company-run contest, not a customer site).
- AgiBot's Yao Maoqing said around the 14 April 2026 Longcheer livestream that merging a humanoid into a production line requires 100% success, 7×24-hour stable operation and never becoming the line bottleneck — 24 hours a day for 7 days without error being 'very hard' — and that while traditional industrial robots reach MTBFs of thousands to tens of thousands of hours, humanoids are generally still at hour-level MTBF; Guancha reported that the AgiBot G2's MTBF had been raised from hour-level to a level supporting two-shift continuous operation.
- Four AgiBot G2 humanoids completed an eight-hour live-streamed shift on 14 April 2026 loading and unloading tablets at MMIT test stations on Longcheer's mass-production line in Nanchang, China, where Xinhua reported 18–20 s per operation and 310 units per hour, and AgiBot/Longcheer reported up to 310 units per hour (about 3,000 per shift), over 99.9% success in continuous operation, over 140 hours of cumulative continuous operation with downtime loss below 4% and 'minimal human intervention' (intervention count not disclosed).
- Humanoids Daily reported on 25 March 2026, citing a recent Wall Street Journal report, that at Schaeffler's 750-employee auto-parts plant in Cheraw, South Carolina, an Agility Digit humanoid moves 25-pound baskets of bearing components from a stamping press to an industrial washing machine for eight hours a day, split into two four-hour shifts with a recharge break in between, working alone inside a Plexiglas safety cage, at an estimated operating cost of $10–25 per hour (depending on service model) versus a $20/hour entry-level wage (no throughput or intervention data reported).
- Securities Times and Wenhui Bao reported on 24 March 2026 that AgiBot partner and senior vice-president Yao Maoqing, in a keynote at the recent 'Zhihui Jinling AI Open-Source Talent Summit' in Nanjing, described the company's embodied-AI deployments, with the reports stating that its solutions had achieved 24-hour continuous uninterrupted operation at real industrial sites, with line throughput (UPH) at human level and a mean time between failures (MTBF) 'at the hundred-hour level' (company claim; sites, robot counts and MTBF definition not disclosed).
- Xiaomi said on 2 March 2026 that a humanoid robot completed its first three-hour continuous autonomous run at a self-tapping-nut loading station on its real automotive production line, meeting the station's 76-second cycle time with a 90.2% dual-side installation success rate (i.e. roughly one failed attempt in ten), working alongside an automated feeder, conveyor and tightening station.
- On Tesla's Q4 2025 earnings call on 28 January 2026, Elon Musk said Optimus was 'still in the R&D phase' and 'not in usage in our factories in a material way' (having done 'some basic tasks in the factory'), which Electrek characterized as an admission that no Optimus robots were doing useful work in Tesla's factories, after Tesla had claimed in June 2024 that two Optimus bots performed factory tasks autonomously and Musk had said in January 2025 he was confident several thousand Optimus robots would be doing useful things by end-2025.
- TechSpot reported on 26 January 2026, drawing on the Financial Times, that UBTech acknowledges its Walker S2 humanoids achieve only 30–50% of a human worker's productivity, that the robot is optimized for controlled, repetitive tasks such as moving parts between stations, stacking boxes and visual defect inspection, that it still needs a human operator to swap its hands or tools for different tasks, and that UBTech aims to reach 80% of human efficiency by 2027.The "Blue Collar" Shift [1] TechSpot: These humanoid robots can only work half as fast as humans
- Tesla’s Optimus robot fell over and went viral in 2025.Beyond the frameworkThe "Blue Collar" Shift [1] Forbes: 2025: Another Year Of Humanoid Hype [2] Wayback capture 20260104222424↳ corrects (or adds sources to) an earlier record:
2025: A Tesla Optimus humanoid fell over during a Miami demo in 2025, a viral moment… - CATL said on 17 December 2025 that it had begun deploying Spirit AI's VLA-driven Xiaomo humanoid robots at scale at its Zhongzhou (Luoyang) battery base to plug in high-voltage battery connectors, claiming a 99% task success rate and a daily workload triple a human worker's because the robots work without breaks (company claims; robot count, throughput per hour and interventions not disclosed).
- Figure AI reported on 19 November 2025, retiring Figure 02, that its 11-month deployment at BMW Plant Spartanburg (about six months of daily runtime on an active line, 10-hour shifts Monday–Friday) loaded 90,000+ sheet-metal parts over 1,250+ runtime hours and contributed to 30,000+ BMW X3s, against targets of an 84 s cycle (37 s load), >99% placement accuracy per shift and zero interventions per shift, but it did not disclose achieved accuracy or intervention counts; Assembly Magazine described the pilot as two robots.KPIThe "Blue Collar" Shift [1] Figure AI: F.02 Contributed to the Production of 30,000 Cars at BMW [2] Assembly Magazine: Humanoid robots complete trial project at BMW assembly plant↳ corrects (or adds sources to) an earlier record:
2025: Figure AI completed an 11-month production deployment of Figure 02 humanoids at BMW's… - Figure AI reported on 7 June 2025 that its Helix policy on Figure 02, rotating mixed packages (including poly bags and flat envelopes) on a conveyor so the barcode faces the scanner, cut average handling time to 4.05 s per package (from ~5.0 s at its initial logistics deployment about three months earlier) and raised correct barcode orientation to ~95% (from ~70%); scaling demonstration data from 10 to 60 hours took handling time from ~6.84 s to 4.31 s and barcode success from 88.2% to 94.4%; it also released a 60-minute uninterrupted video of Figure 02 doing the task (company evaluation, not a customer shift).
- Bank of America projected in an April 2025 report that mass adoption of humanoid robots will start in 2028.↳ corrects (or adds sources to) an earlier record:
2025-04-30: Bank of America projected in an April 2025 report that mass adoption of humanoid robots… - Fortune reported on 6 April 2025 that, per BMW spokesperson Steve Wilson, only a single Figure robot worked at BMW's South Carolina plant at any given time and, until sometime in March 2025, only during non-production hours, after which the same robot performed one limited chore in live production hours (picking parts from a logistics container and placing them onto a welding fixture in the body shop), contrasting with CEO Brett Adcock's February claim of a 'fleet of robots performing end-to-end operations'.
- In a 28 February 2024 'Deployment Update: Putwall Automation' post, Agility Robotics said Digit continues to complete real tasks in real facilities, with workflows run on a continuous basis and 100% autonomously, at 75% of a human's speed and a 97% success rate, without gantry or tether (company claim; site, run length and intervention counts not disclosed).
Other events (10)
- Boston Dynamics opened its Robotics Metaplant Application Center (RMAC) at Hyundai Motor Group Metaplant America near Savannah, Georgia on 21 September 2026 as a testbed and training center where Atlas humanoids are 'now in training' on preparing the logistics and sequencing of automotive parts, with component assembly targeted by 2030; Hyundai Motor Group plans to begin Atlas deployment with 25,000 units at Hyundai and Kia plants over the next few years.
- AgiBot announced on 14 April 2026 that multiple AgiBot G2 robots had been officially integrated into Longcheer Technology's tablet mass-production lines in Nanchang, where at MMIT test stations they autonomously pick up tablets, navigate the factory, place devices into test fixtures and sort finished or defective units, claiming up to 310 units per hour, over 99% success in continuous operation, about 3,000 units per shift, and 24/7 autonomous operation with minimal human intervention over 140+ hours of cumulative continuous operation (company claims; robot count and intervention count not disclosed).KPIThe "Blue Collar" Shift [1] AGIBOT: AGIBOT and Longcheer Technology Achieve World's First Embodied AI Deployment in Consumer Electronics Precision Manufacturing Mass-Production Line (14 Apr 2026) [2] AGIBOT G2 User Manual ('AGIBOT G2 wheeled humanoid robot'; general-purpose wheeled-chassis humanoid with four-wheel-steering omnidirectional chassis)↳ corrects (or adds sources to) an earlier record:
2026-04-14: AgiBot announced on 14 April 2026 that multiple AgiBot G2 robots had been officially… - The Wall Street Journal reported that Tesla Optimus demos are often remotely controlled by human engineers.The "Blue Collar" Shift [1] WSJ: Elon Musk Optimus robots
- Kepler's K2 humanoid (about $34,600) ran an 8-hour livestream on a single charge in a trade-show demo, not in production work.↳ corrects (or adds sources to) an earlier record:
2025: Kepler's K2 humanoid (about $34,600) ran an 8-hour livestream on a single charge in a… - Agility Robotics reported on 20 November 2025 that its Digit humanoids had moved over 100,000 totes in commercial deployment at GXO's Flowery Branch facility, picking items on and off an AMR to a conveyor and also stacking totes onto a different floor location (no runtime, fleet-size or intervention figures given).KPIThe "Blue Collar" Shift [1] Agility Robotics: Digit Moves Over 100,000 Totes in Commercial Deployment↳ corrects (or adds sources to) an earlier record:
2025-11: Agility Robotics reported that its Digit humanoids at GXO Logistics had moved more than… - +5 older event(s): the section's event log
Robotaxi Saturation: Autonomous robotaxis outnumber human-driven taxis in a major metropolitan area (driverless transport solved).Progressing: Waymo ~15% of SF ride-hail bookings, flat in 2026; fleet ~4,000 cars, ETA ~2031 · assessed 2026-10-19 · 45 event(s)
- Taxis include ride-hail; count the vehicles in service at peak.
- Driverless (SAE L4) service across the metro's service area qualifies; SAE L5 is not required.
In San Francisco, Waymo took about 15% of the money spent on Waymo, Uber and Lyft trips inside its service zones in June 2026, by YipitData's count of email receipts, slightly down from 16% in January. Meanwhile Waymo's fleet grew from about 2,500 to 4,000 cars between December 2025 and September 2026 as it expanded from 6 to 15 cities. If San Francisco's share stays flat until Waymo's next-generation Ojai vehicles arrive, around early 2027, and then grows as fast as the fleet has, robotaxis would outnumber human-driven taxis and ride-hail cars there around 2031 (80% range 2029-2040, P(by 2100) 92%), later than the previous ETA of 2028, which assumed 2025's ride growth continued. Yellow, not green: the one independent measure of the share shows a stall, so the renewed growth the forecast relies on is not yet corroborated.- ETA
- ~2031 (80% range 2029–2040; P(by 2100) 92%)
- ETA cruxes
- First, the growth mapping: whether SF's robotaxi share resumes growing like Waymo's fleet (0.58/yr) once Ojai supply arrives, or stays near the 2026 ride/share rate (half the fleet rate: p50 2034; the 2026 ride rate 0.14/yr: 2043). Settles it: YipitData's next SF share readings, a CPUC/SFMTA vehicle count, Waymo's SF fleet size. Then the stall length (1.5 yr: 2032). Earlier: the SF share rising above ~20% in 2027, Waymo reaching ~1M rides/week, Tesla or Zoox adding SF driverless fleets. Later: the share still flat through 2027, Ojai imports delayed (tariffs), a suspension after a serious incident.
- Path
- Reading used (owner-decided, framework 66c0e8a): robotaxis in service outnumber human-driven taxis plus ride-hail cars in service (vehicles on road at peak) in one major metro; driverless SAE L4 across the metro's service area qualifies, L5 not required. Metro = San Francisco. Proxy for the level: YipitData's Waymo share of Waymo+Uber+Lyft bookings inside Waymo's zones (dollars, not vehicles); it tracks the vehicle ratio because trips served and vehicles in service at peak scale together, and the bookings-to-vehicles factor is an assumed input (m, 0.7-1.4). Proxy for the rate: Waymo's fleet (vehicles, the milestone's unit), system-wide; its mapping onto SF is assumed. Path: Waymo (~4,000 cars, 15 US cities, Ojai/6th-gen driver ramping, $16B raised); second routes: China (Apollo Go, Pony.ai 1,975 cars and Guangzhou city-centre service, WeRide), Tesla (unsupervised in Texas/Florida, SF still with safety drivers), Zoox (NHTSA exemption up to 2,500/yr).
- Blockers
- Data: no SF vehicle counts; one share series (YipitData, two points) (gather request: Yipit updates, CPUC/SFMTA counts, Waymo SF fleet)
- Supply: Waymo's SF share flat in H1 2026 while its fleet went to new cities; growth depends on the Ojai vehicle ramp
- Safety/regulatory: Santa Monica child struck (NHTSA PE26001), NTSB on remote assistance passing a school bus; China paused new AV licences after the Wuhan mass stall (no reported end as of 2026-10-19)
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- trend (exponential growth of the logistic SF robotaxi/human vehicle odds) from the measured SF booking-share level (YipitData, a proxy), held flat for an assumed supply stall, then growing at the measured Waymo fleet growth (mapping assumed), plus an assumed publication lag; named failure scenarios (each < 10%); Monte Carlo N=200k (no shared driver); prev, proponent and reference-class cross-checks
- ETA calculation
- Level: YipitData SF Waymo share of Waymo+Uber+Lyft gross bookings inside Waymo's zones 16% (Jan 2026) -> 15% (Jun 2026) (2026-06-01-yipit-waymo-15pct-sf-bookings, MEASURED proxy; LA 17->15, Phoenix 19->16): odds 0.190 -> 0.176, g = ln(0.176/0.190)/(5/12) = -0.18/yr (whole-percent rounding range -0.37..0). Bookings-to-vehicles-in-service mapping m = 1 (0.7-1.4), ASSUMED (prev's vehicle composite r0 0.16 at Dec 2025 = 0.84 x Yipit's Jan odds). Growth: Waymo fleet ~2,500 (Dec 2025, 2025-12-01-waymo-450k-weekly-rides) -> ~3,000 (Mar 2026, 2026-03-24-waymo-3000-cars-500k-weekly-rides) -> ~4,000 in 15 cities (24 Sep 2026, 2026-09-24-waymo-fleet-4000-texas-1102; Texas 1,102 registered, +49% in 3 weeks as Ojai arrives): g_fleet = ln(1.6)/0.81 = 0.58/yr MEASURED; rides 450k -> 500k/wk = 0.13/yr (2025: 250k -> 450k = 0.88/yr). Model: SF odds flat for a stall D = 0.55 yr (0.2-1.6) after Jun 2026 (ASSUMED: the named constraint is vehicle supply, easing as Ojai vehicles arrive; SF Ojai riders since 2026-05-28-waymo-ojai-mesa-tens-of-thousands), then growth g ~ N(0.58, 0.30), floor 0.15 (mapping system fleet -> SF odds ASSUMED; spread and floor prev's, ASSUMED, kept), up to odds 1 (logistic odds grow exponentially, no early S-curve slowdown), + publication lag 0.4 yr (0.15-1.2, assumed). Point: 2026.45 + 0.55 + ln(1/0.176)/0.58 (= 2.99) + 0.4 = 2030.4. Failure (assumed, each < 10%): permanent regulatory reversal in every lead metro 0.02 (Cruise ended one operator; Wuhan's city suspension of Apollo Go was temporary, existing service partly resuming in three districts after ~4 months, 2026-08-03-apollo-go-wuhan-service-resumes; China's national pause on new AV licences, fleet additions, test programmes and new cities, 2026-04-29-china-pauses-new-av-licences, 'for an unknown period', has no reported end: unresolved, duration unmeasured. Kept at 0.02 because the scenario needs a permanent reversal in every lead metro including SF, where no halt is reported; the open China pause freezes the second route's expansion. F1 0.04 (failure 0.10): p_a 0.90, p50 2031.11, p90 2067.50); economics/competition plateau below parity in every major metro through 2100 0.05 (raised from 0.02: the share fell in all three Yipit metros in H1 2026); parity never counted 0.01 -> p_achieved 0.92. MC (N=200k) conditional 2029.14/2030.82/2036.55; unconditional p10 2029.18 / p50 2031.05 / p90 2040.57 (p90 sits at the 98th conditional percentile, set by the low-g tail); P(by 2031.80) 0.608; P(by end-2030) 0.493. Attribution vs prev 2027.60/2028.62/2031.58, p_a 0.95: new level and fleet rate, no stall 2028.60/2030.15/2038.08; + stall 2029.16/2030.95/2038.85; + plateau 0.05 2029.18/2031.05/2040.57, p_a 0.92. Prev's own mapping (SF odds grow like system rides) on the 2026 ride rate (0.13/yr, run at 0.14) gives p50 2041.47. Sensitivities (p50): no stall 2030.23; stall 1.5 yr 2032.30, 0.25 yr 2030.56; g 0.88 (2025 rides) 2029.84; g 0.29 (half) 2034.61 (CRUX); g 0.14 without a stall 2043.14 (CRUX); g SD 0.15/0.40 2030.86/2031.15 (p90 2035.04/2041.43); floor 0.05/0.25 p90 2061.82/2036.26; m 0.84/0.6/1.4 2031.36/2031.97/2030.43; lag 0/1.2 2030.45/2032.20; plateau 0.02 (prev failure 0.05) 2030.95 (p90 2038.85), failure 0.15 p_a 0.85, p90 null. Cross-checks: Waymo's 1M rides/wk by end-2026 target (2026-02-24-waymo-four-metros-1m-weekly-target, proponent: 2x in ~10 months) and Goldman's 7k -> 1M global fleet 2025-2030 (2026-04-30-goldman-robotaxi-fleet-1m-2030, 2.7x/yr) discounted 1.5-3x both imply growth at or above g 0.58 once supply flows; the Yipit stall says it has not yet. Reference class: operator stalls so far were incident- or supply-driven and lasted months (Wuhan suspension Apr-Aug 2026, 2026-08-03-apollo-go-wuhan-service-resumes; Cruise ended; China's national licence pause still open). No anchors. Prev (assess:20261015T100057Z) 2027/2028/2031, p_a 0.95, reproduced exactly (2027.60/2028.62/2031.58). p50 lies at 2031.05: a slightly earlier input reads as 2030.
- Data as of
- 2026-09 (Waymo fleet 2026-09-24; SF share 2026-06)
- Change
- yellow -> yellow. New evidence (gather 20261018, 25 robotaxi events published Jan-Sep 2026): Waymo ~3,000 cars and 500k rides/wk (Mar 2026), ~4,000 cars in 15 cities at 500k rides/wk (Sep 2026); YipitData SF booking share 16% -> 15% (Jan -> Jun 2026); Apollo Go Wuhan stall, suspension and partial resumption; China's national licence pause (no reported end); Pony.ai 1,975 cars; Tesla unsupervised in 6 metros (42 registered in Texas); Zoox exemption. Prev's model (odds growing at Waymo's 2025 system ride rate 0.88/yr from a Dec 2025 composite level) is contradicted (rule 7 (a)): system rides grew 0.13/yr and the SF share fell. Replaced inputs: level = Yipit SF share (Jun 2026) x assumed bookings-to-vehicles factor, instead of the assumed SF fleet / on-road / denominator composite; rate = measured Waymo fleet growth 0.58/yr after an assumed supply stall (0.55 yr), instead of the 2025 ride rate; plateau scenario 0.02 -> 0.05 (Yipit: share fell in all three metros). Kept: growth spread SD 0.30 and floor 0.15, lag, F1 0.02 (the China licence pause is open but SF has no halt), F3. Moves: p10 2027 -> 2029 (2027.60 -> 2028.60 new level and rate, -> 2029.16 stall, -> 2029.18 plateau), p50 2028 -> 2031 (2028.62 -> 2030.15 new level and rate, -> 2030.95 stall, -> 2031.05 plateau), p90 2031 -> 2040 (2031.58 -> 2038.08 new level and the lower rate centre, whose spread reaches a slower tail -> 2038.85 stall -> 2040.57 plateau: p_achieved 0.92 puts p90 in the low-growth tail), p_achieved 0.95 -> 0.92 (plateau scenario). Status yellow unchanged.
- Rubric rule
- Milestone yellow (RUBRIC v6 §4): measurable progress on the milestone's quantity (SF robotaxi share, ~15% of bookings; Waymo fleet 2,500 -> 4,000), median ETA 2031 (2031.05) <= made_at + 5 yr (2031.80), but green's corroboration condition fails: the only non-proponent measure of the SF share (YipitData) shows it flat to falling in H1 2026, and the resumed growth the median relies on rests on Waymo's fleet and supply plans. Not achieved: no event shows robotaxis outnumbering taxis + ride-hail in a metro.
by assess:20261019T010556Z · 42 evidence event(s)Evidence and news: 42 cited by the assessment, 3 other event(s)
Cited evidence (42)
- Waymo reported on 24 September 2026 that across over 270 million fully autonomous miles through June 2026 in five metros (Atlanta, Austin, Los Angeles, Phoenix, San Francisco), its vehicles were involved in 82% fewer injury-causing crashes (841 fewer) and 95% fewer serious-injury-or-worse crashes than human-driver benchmarks, regardless of fault (company analysis).Robotaxi Saturation [1] Waymo: From the road — safety data update (24 September 2026)
- TechCrunch reported on 24 September 2026 that Waymo had roughly 4,000 robotaxis serving 15 U.S. cities at an average of 500,000 paid rides a week, about 80% of them in California and Texas (1,102 autonomous vehicles registered in Texas as of 24 September, up 49% in three weeks as Ojai minivans arrived) and about 800 in other states, and that MoffettNathanson expected Waymo to import 5,100 Ojai vehicles by the end of 2026.
- Pony.ai reported on 18 August 2026 that its robotaxi fleet had reached 1,975 vehicles, against a year-end 2026 target of more than 3,500 (raised from 3,000 in May), and that in Guangzhou its robotaxi service now reached the city centre across Haizhu, Tianhe, Huangpu and Panyu districts, its operating area having grown by over 300 km² since January to cover a population of over 7 million.
- Baidu said on its 18 August 2026 Q2 earnings call that Apollo Go delivered around 1 million fully driverless operational rides in Q2 2026, as ride volume was 'temporarily affected by operational adjustments in certain domestic cities due to regulatory considerations' (operations in the affected cities had begun to resume as of August), with cumulative public rides exceeding 23 million as of June 2026; its Q2 results put Apollo Go's global footprint at 28 cities.Robotaxi Saturation [1] Yahoo Finance: Baidu, Inc. (BIDU) Q2 FY2026 earnings call transcript (18 August 2026) [2] Baidu: Baidu Announces Second Quarter 2026 Results (18 August 2026)
- Zoox, which said on 30 July 2026 that NHTSA had granted it the first commercial exemption for a purpose-built robotaxi, announced on 5 August that it would start charging fares in Las Vegas from 10 August 2026, its first commercial launch, after carrying nearly 1 million riders free since 2025 across Las Vegas, San Francisco, Austin and Miami; paid trips elsewhere still need state and local approvals.Robotaxi Saturation [1] CNBC: Amazon's Zoox to launch paid robotaxi service in Las Vegas on Aug. 10 (5 August 2026) [2] Zoox: Zoox granted first-ever NHTSA commercial exemption for purpose-built robotaxi (30 July 2026)
- Bianews (读佳) reported on 3 August 2026 that Baidu's Apollo Go had resumed operating in Wuhan about four months after the 1 April city-wide suspension, with service available in its tests only in the Wuhan Economic-Technological Development Zone and the Caidian and Hanyang districts.Robotaxi Saturation [1] Bianews (读佳): 时隔4个月，百度的萝卜快跑在武汉市恢复运营 (3 August 2026)
- NHTSA announced on 30 July 2026 that it would allow Zoox to commercially deploy its robotaxis through a temporary exemption permitting up to 2,500 vehicles a year for two years under an enhanced, adaptable oversight structure, alongside a three-year, $5 million 'A2SCEND' consortium with SAE Industry Technologies Consortia to develop the first AV performance standards, an interim final rule letting vehicles built before an exemption qualify for commercial-deployment exemptions, and updated guidance to streamline Part 555 exemptions.
- Tesla said in its Q2 2026 update of 22 July 2026 that Cybercab production had begun at Gigafactory Texas and that its Robotaxi service was live in seven metros, ramping unsupervised rides in Austin, Dallas and Houston and, from July, Miami, Orlando and Tampa, while the SF Bay Area still used safety drivers and Phoenix and Las Vegas, both listed for 1H 2026 in January, were still at 'preparations underway'; it disclosed no Robotaxi fleet count.
- Business Insider reported on 16 August 2026 that YipitData, using email receipts from about 1.5 million US consumer accounts and counting only trips that start and end inside Waymo's operating zones, estimated Waymo's share of Waymo, Uber and Lyft gross bookings (dollars spent, not trips or vehicles) at 15% in San Francisco and Los Angeles and 16% in Phoenix in June 2026, down from 16%, 17% and 19% in January.
- Electrive reported on 2 June 2026 that Tesla had registered 42 autonomous robotaxis with the Texas DMV under new reporting rules, mostly in Austin with the rest believed to be in Dallas and Houston, far below Elon Musk's 2025 talk of about 1,000 vehicles within months, and, citing third-party Robotaxi Tracker data via InsideEVs, that unsupervised operation in Austin began with a single vehicle on 22 January 2026 and had reached about 30 vehicles.
- Waymo said on 28 May 2026, as select riders in San Francisco, Phoenix and Los Angeles began riding in its Zeekr-based Ojai, the first vehicle with its 6th-generation Driver, that it was scaling its Mesa, Arizona factory toward a capacity of tens of thousands of Waymo-enabled vehicles per year (a capacity target, not output).
- Baidu reported on 18 May 2026 that Apollo Go delivered 3.2 million fully driverless operational rides in Q1 2026, with weekly rides peaking at over 350,000 in March and total rides up over 120% year over year, cumulative public rides exceeding 22 million as of April 2026 and a footprint of 27 cities as of May.Robotaxi Saturation [1] Baidu: Baidu Announces First Quarter 2026 Results (18 May 2026)
- WeRide reported on 13 May 2026, with its Q1 results, that its global robotaxi fleet had reached approximately 1,300 vehicles as of 30 April 2026, about 1,000 of them in China, and that it planned to deploy 1,200 robotaxis across Abu Dhabi, Dubai and Riyadh by 2027.
- Goldman Sachs Research forecast on 30 April 2026 that the worldwide commercial robotaxi fleet will grow from roughly 7,000 vehicles in 2025 to about 1 million in 2030 and approximately 6 million in 2035.
- Tesla said in its Q1 2026 update of 22 April 2026 that it had further expanded its unsupervised (no safety monitor) Robotaxi operating area in Austin and launched unsupervised Robotaxi rides in Dallas and Houston in April 2026; it gave no fleet count.Robotaxi Saturation [1] Tesla: Q1 2026 Update (22 April 2026)
- Bloomberg reported on 29 April 2026, citing people familiar with the matter (as relayed by Lianhe Zaobao), that China had suspended issuing new autonomous-vehicle licences after the 31 March Apollo Go stall in Wuhan — stopping operators from adding driverless robotaxis to their fleets, starting new test programmes or entering new cities, for an unknown period — and that Baidu's Wuhan robotaxi business remained suspended while the local government investigated.Robotaxi Saturation [1] 联合早报: 萝卜快跑故障后 中国暂停发放自动驾驶新牌照 (29 April 2026, citing Bloomberg)
- On the evening of 31 March 2026 what Wuhan traffic police preliminarily called a 'system failure' left many Baidu Apollo Go robotaxis stalled in the middle of roads in Wuhan, Apollo Go's largest operating city, trapping passengers (one for nearly two hours on an elevated road) and leading to at least one rear-end collision but no injuries, and on 1 April Apollo Go's customer service said service was suspended across all of Wuhan.
- WeRide and Uber began fully driverless, fare-charging robotaxi operations with no vehicle operator on board along routes in Dubai's Jumeirah and Umm Suqeim districts on 31 March 2026, when WeRide had over 200 robotaxis in the Middle East and the partners had committed (February 2026) to deploy at least 1,200 across Dubai, Abu Dhabi and Riyadh.
- Pony.ai reported on 26 March 2026, with its Q4/FY2025 results, that its robotaxi fleet had surpassed 1,400 vehicles, that its Guangzhou and Shenzhen operations had reached unit-economics breakeven within four months of the Gen-7 robotaxi launch, and that it targeted a fleet of over 3,000 vehicles in more than 20 cities by end-2026.
- Waymo co-CEO Dmitri Dolgov said in a podcast released on 24 March 2026 that Waymo had 'about 3,000 cars on the roads' completing roughly half a million rides and over four million fully autonomous miles a week, with riders in 10 U.S. cities, and Waymo said on X that week that it was providing 500,000 paid robotaxi rides per week across 10 U.S. cities, TechCrunch reported.
- Baidu reported on 26 February 2026 that Apollo Go delivered 3.4 million fully driverless operational rides in Q4 2025, with weekly rides peaking at over 300,000 during the quarter and total rides up over 200% year over year, and that by February 2026 its cumulative public rides exceeded 20 million across a footprint of 26 cities.
- Waymo opened its driverless ride-hailing service to its first public riders in Dallas, Houston, San Antonio and Orlando on 24 February 2026, bringing it to 10 commercial metro areas, and co-CEO Tekedra Mawakana said Waymo was on track to serve over one million rides per week by the end of 2026 (a projection).
- Waymo said on 2 February 2026 that it had raised a $16 billion investment round at a $126 billion post-money valuation, that it was providing more than 400,000 rides a week across six U.S. metropolitan areas after more than tripling its annual volume to 15 million rides in 2025 (over 20 million lifetime), and that it was preparing ride-hailing in over 20 additional cities in 2026, including Tokyo and London.
- Tesla said in its Q4 2025 update of 28 January 2026 that, after testing driverless Robotaxis in Austin from December 2025, it began removing the safety monitor from Austin Robotaxi customer rides in January 2026 'on a limited basis', while its SF Bay Area service still ran with safety drivers and Dallas, Houston, Phoenix, Miami, Orlando, Tampa and Las Vegas were listed for 1H 2026; it gave no count of vehicles running without a monitor.Robotaxi Saturation [1] Tesla: Q4 and FY 2025 Update (28 January 2026)
- Waymo began fully autonomous rides to and from San Francisco International Airport on 29 January 2026 for a select number of riders, with pickups and drop-offs initially only at the SFO Rental Car Center (an AirTrain ride from the terminals), all riders to be admitted over the following months and terminal stops planned later.Robotaxi Saturation [1] Waymo: Clear skies and autonomous Waymo rides at SFO (29 January 2026)
- On 23 January 2026 an unoccupied Waymo robotaxi (5th-generation driver, travelling at 17 mph after a drop-off) braked but struck a 9-year-old pupil who stepped out midblock between queued cars in a Santa Monica school zone, causing minor injuries with no medical transport needed; NHTSA opened a preliminary evaluation (PE26001) on 28 January and the NTSB, which published preliminary details on 3 March 2026, is also investigating.
- The NTSB reported preliminarily on 3 March 2026 that on 12 January 2026 an unoccupied Waymo robotaxi in Austin stopped for a school bus loading pupils, asked remote assistance whether it was 'a school bus with active signals', was told 'No' by an agent in Novi, Michigan, and then passed the bus with its stop arms extended (no crash) — one of multiple such incidents reported by Austin ISD in the 2025/26 school year that the NTSB is investigating, after Waymo's December 2025 software recall of 3,067 vehicles.
- By the end of 2025, Tesla's Austin Robotaxi service still had safety riders.Robotaxi Saturation [1] InsideEVs: Tesla's Big Robotaxi Promises Fall Flat As 2025 Comes To A Close [2] Wayback capture 20260101010700↳ corrects (or adds sources to) an earlier record:
2025-12: Tesla missed all its 2025 robotaxi targets: Tesla had promised driverless service in… - By December 2025 Waymo was providing more than 450,000 paid robotaxi rides per week with a fleet of about 2,500 vehicles across 6 US cities.Robotaxi Saturation [1] Forbes: Waymo targets 1M robotaxi rides a week [2] Waymo 2025 Year in Review
- A Baidu Apollo Go robotaxi caused an accident in Zhengzhou, China in December 2025 that left two people in intensive care.Robotaxi Saturation [1] Car News China: Baidu robotaxi accident↳ corrects (or adds sources to) an earlier record:
2025-12: A Baidu Apollo Go robotaxi caused an accident in Zhengzhou, China in December 2025 that… - In December 2025, Waymo recalled its 5th Generation ADS software after 12 citations were identified as related to vehicles proceeding before school buses deactivated their warning signals.Robotaxi Saturation [1] NHTSA: Part 573 Safety Recall Report 25E084↳ corrects (or adds sources to) an earlier record:
2025-12: Waymo issued a recall of its self-driving vehicles in December 2025 after 19+ incidents… - Tesla's robotaxi service remained a pilot in Austin, Texas with roughly 30–50 vehicles in 2025.Robotaxi Saturation [1] CNBC: Tesla obtains Arizona ride-hail permit↳ corrects (or adds sources to) an earlier record:
2025: Tesla's robotaxi service remained a pilot in Austin, Texas with roughly 30–50 vehicles… - Pony.ai reported a combined fleet exceeding 720 fully driverless commercial robotaxis across China's four tier-1 cities; it launched Gen-7 robotaxis in Guangzhou, Shenzhen and Beijing.Robotaxi Saturation [1] Pony.ai: Fully driverless commercial services with Gen-7 [2] Wayback capture 20260130095434↳ corrects (or adds sources to) an earlier record:
2025: Pony.ai operates a fleet of 720+ robotaxis, running fully driverless commercial… - Waymo launched robotaxi freeway operations for riders in San Francisco, Los Angeles and Phoenix in November 2025.Robotaxi Saturation [1] Waymo: Taking riders further with freeways↳ corrects (or adds sources to) an earlier record:
2025-11: Waymo launched robotaxi freeway operations for riders in San Francisco, Los Angeles and… - Baidu's Apollo Go robotaxi service was providing more than 250,000 rides per week as of November 2025, operating in 22 cities globally with 1,000+ vehicles in Wuhan.Robotaxi Saturation [1] Car News China: Baidu Apollo Go↳ corrects (or adds sources to) an earlier record:
2025-11: Baidu's Apollo Go robotaxi service was providing more than 250,000 rides per week as of… - Pony.ai received its first city-wide robotaxi permit, covering all of Shenzhen, in October 2025.Robotaxi Saturation [1] CNBC: Global robotaxi race heats up↳ corrects (or adds sources to) an earlier record:
2025-10: Pony.ai received its first city-wide robotaxi permit, covering all of Shenzhen, in… - In September 2025, Zoox had deployed a fleet of 50 robotaxis between San Francisco and Las Vegas.Robotaxi Saturation [1] CNBC: Zoox kicks off robotaxi rides in San Francisco, facing off with Waymo↳ corrects (or adds sources to) an earlier record:
2025-12: Zoox operated a robotaxi fleet of about 50 vehicles in San Francisco and Las Vegas as… - Waymo's robotaxi fleet reached 100 million fully autonomous miles in July 2025.Robotaxi Saturation [1] Waymo 2025 Year in Review↳ corrects (or adds sources to) an earlier record:
2025-07: Waymo's robotaxi fleet reached 100 million fully autonomous miles in July 2025. - Waymo reached 250,000 paid robotaxi rides per week (about 1 million per month) in April 2025.Robotaxi Saturation [1] Forbes: Waymo targets 1M robotaxi rides a week
- Waymo and Uber launched robotaxi rides in Austin, Texas in March 2025.Robotaxi Saturation [1] CNBC: Waymo, Uber begin robotaxi rides in Austin↳ corrects (or adds sources to) an earlier record:
2025-03: Waymo and Uber launched robotaxi rides in Austin, Texas in March 2025. - In December 2024, General Motors stopped funding Cruise, shifting autonomous-vehicle work toward driver-assistance systems and ending funding for autonomous taxis.Robotaxi Saturation [1] Guardian: GM ends Cruise robotaxi [2] Wikipedia: Cruise [3] Wayback capture 20241212172225↳ follows up on 2023-10: In October 2023 a GM Cruise robotaxi dragged a pedestrian, leading to a recall of… (corrected version)↳ corrects (or adds sources to) an earlier record:
2024-12: General Motors shut down its Cruise robotaxi business in December 2024, ending robotaxi… - In October 2023 a GM Cruise robotaxi dragged a pedestrian, leading to a recall of Cruise's self-driving cars.Robotaxi Saturation [1] Guardian: Cruise recall↳ corrects (or adds sources to) an earlier record:
2023-10: In October 2023 a GM Cruise robotaxi dragged a pedestrian, leading to a recall of…
Other events (3)
- A Waymo robotaxi killed a dog in November 2025, the company's second animal fatality in two months.↳ corrects (or adds sources to) an earlier record:
2025-11: A Waymo robotaxi killed a dog in November 2025, the company's second animal fatality in… - Baidu said Apollo Go's Wuhan operations were nearing break-even and projected to achieve profitability by 2025.Robotaxi Saturation [1] Car News China: Baidu's Apollo Go robotaxi leads global autonomous driving with 17M+ orders, targets profit this year↳ corrects (or adds sources to) an earlier record:
2025-11: Baidu said Apollo Go's Wuhan operations were nearing break-even and projected to… - A Waymo robotaxi killed a cat in October 2025, raising questions about its sensor detection.↳ corrects (or adds sources to) an earlier record:
2025-10: A Waymo robotaxi killed a cat in October 2025, raising questions about its sensor…
The Dark Line: A production line staffed only by general-purpose robots runs a full week, 24/7, with no humans on the floor.Progressing: 15 humanoids load 15 stations beside workers; no 24/7 robot-only line, ETA ~2031 · assessed 2026-10-19 · 23 event(s)
- A real line of 5+ stations; robots also do material moves and routine maintenance, including their own (charging, battery swaps).
- ≥90% of the line's required throughput, with one product changeover by software alone during the run.
- Purpose-built lights-out automation ("dark factories") does not count; it is reported under Beyond. General-purpose = robots that software alone could move to other jobs (humanoids, mobile manipulators), not machines built for this line.
- No teleoperation and no remote fault-fixing; a remote human may only give job-level orders (e.g. start the changeover).
No report shows a 5+ station line run only by robots round the clock with no humans on the floor; the nearest are 15 Leju humanoids loading parts at 15 stations of Zhaofeng's line alongside workers, at about 40% of the line's cycle rate, and 8 AgiBot G2s running Longcheer's test section 10 hours a day with a maintenance team. With the same breakdown, task-failure and assist rates as the Blue Collar Shift estimate, a 7-robot line with 3 spares would need a human about 105 times a week; with those rates falling at the pace Dyna Robotics' napkin-folding robots have cut their reject rate (an assumed carry-over), and allowing for the throughput and changeover requirements and for how often anyone stages an attempt, the first qualifying week comes ~2031 (P(by 2100) 88%). Yellow: the ETA is just inside 5 years, but there is no robot-only multi-station precursor or corroborated human-free 24/7 run.- ETA
- ~2031 (80% range 2028–never/after 2100; P(by 2100) 88%)
- ETA cruxes
- First, the mapping: whether breakdown and intervention rates halve like Dyna's reject fraction (halving twice as slow moves p50 to 2035.15); dated MTBF and intervention-hour logs settle it. Then the attempt rate (1/yr: 2032.71), the economics scenario (own failure 0.25: p_achieved 0.735) and the size of other human assists (1.5/shift: 2032.09). Earlier: a disclosed MTBF >= 100 h at a real site (Zhaofeng's claim, if documented); a staged multi-robot 24/7 week; robots doing line maintenance (Zhaofeng's changeover recalibration) or a software-only changeover. Later: MTBF stuck at hour level, multi-robot collisions, humanoid lines far below required throughput, programmes scaled back.
- Path
- Reading used (owner-decided wording, framework 17ea1bf): a real line of 5+ stations staffed only by robots not purpose-built for it (humanoids/mobile manipulators running general models), no human entering the floor for 168 h, no teleoperation and no remote fault-fixing (a remote human only gives job-level orders). The line's own machines may be automated; the jobs people do (loading, moves, inspection, upkeep) must be done by general-purpose robots. Nearest precursors: Leju Kuavo at Zhaofeng (15 robots at 15 stations of an auto-parts line, workers present, ~40% of the cycle, exploring robot recalibration at changeovers and a lights-out closed loop); AgiBot G2 at Longcheer (8 robots, whole MMIT section, day shifts, maintenance team); Galbot S1 at CATL (7x24 bin moves, claim); Figure 03 HQ fleet with self-charging rotation (claim); autonomous battery swap (UBTech, Atlas). Path: the Blue Collar reliability components driven ~7 halvings lower, plus software changeover, then a staged multi-station week; throughput modelled as a gate on the bottleneck station; scheduled joint service falls outside a one-week run.
- Blockers
- Technical: ~105 human-needed events per robot-line-week today (hour- to two-shift-level MTBF, task failures, assists) vs ~0.7 needed; multi-robot collisions when scaling (Longcheer)
- Technical: robot-performed routine line maintenance and a software-only changeover not demonstrated (UBTech needs human tool swaps; Zhaofeng only exploring robot recalibration)
- Economic: purpose-built automation may be cheaper (Zhaofeng: robots fill gaps between automated processes), so the line may only be staged as a showcase
- Disclosure: no deployer publishes human-presence or intervention data
- Throughput: single stations shown at 90-106% of need (company claims), but Zhaofeng's 15-station line runs at ~40% of its cycle and UBTech at 30-50% of human productivity
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (exponential decline of breakdown, task-failure and assist rates at the shared driver dyna-failure-halving, mapping assumed) in an explicit one-week model (finite spares net of robots away charging, fault absorption, software changeover, >= 90% throughput gate) with dated staged attempts, setup and report lag; named failure scenarios (each < 10%); Monte Carlo N=10,000 on the driver's canonical draws, components shared with Blue Collar; discounted-proponent and reference-class checks
- ETA calculation
- Week: 7 active robots (5 stations + 2 movers, assumed) x 168 h, K = 3 spares (assumed) minus robots away charging round(7 x cfrac), cfrac 0.10 (0.02-0.25, assumed; self-swap/charging per 2025-07-17-ubtech-walker-s2-autonomous-battery-swap, 2026-01-05-atlas-product-self-battery-swap, 2026-02-22-figure-03-hq-24-7-no-babysitter). Per-robot rates (Blue Collar components, same draws): hardware 8/MTBF (MTBF 16 h sourced company claim, spread 5-60 assumed), task failures 0.133/8h measured (2026-06-28-agibot-longcheer-whole-section-8-robots; spread 0.08-0.5 assumed) x 0.5 needing a human (assumed), other assists 0.3/shift (assumed). Absorption 0.6 (assumed). Mid-2026: unabsorbable 105/week (38-328), absorbable 43/week. P(week) = exp(-lam_na) x P(Poisson(lam_m) <= K_eff) x p_changeover (0.3, assumed) x [throughput gate]. All event rates halve every draw i of shared driver dyna-failure-halving halving_months (p50 7.623 mo measured, 2026-08-27-dyna-din-tai-fung-shift-throughput; spread assumed; mapping ASSUMED); changeover odds at its odds_halving_months; bottleneck throughput b0 0.7 (0.4-1.0, assumed; Zhaofeng's 15 s needed / 35-40 s per load = 0.40, 2026-10-18-zhaofeng-kuavo-cycle-gap-lights-out-goal, sits at the low end) x availability 0.95 (assumed) doubling at its throughput_doubling_months (p50 8.10) must reach 0.90: gate 10/50/90/99% 2026.50/2026.76/2027.41/2028.30, nonbinding at base. Point: lam_na 105 -> 0.7 needs 7.2 halvings x 0.635 yr = 4.6 yr -> ~2031.1 for a 50% week. Attempts Poisson 6/yr (assumed) from made_at + setup 0.4 yr (assumed); report lag 0.15 yr (assumed). Failure (assumed, each < 10%): shared stall 0.02; purpose-built automation cheaper so no qualifying staged line 0.08 (Zhaofeng's GM: robots fill the manual gaps between automated processes; its chairman's lights-out goal points the other way; neither moves it); never verifiably disclosed 0.02 -> p_achieved 0.882. MC (N=10,000 driver draws, made_at 2026.797) conditional 2028.88/2030.89/2034.47; unconditional p10 2028.97 / p50 2031.28 / p90 null; P(<=2031.80) 0.578, P(<=2036.80) 0.857. Sensitivities (p50): mapping slow x2 2035.15 (CRUX), x4 2042.55, fast (task-horizon doubling_software_months 3.75 mo / 7.623) 2029.23; b0 0.4 (Zhaofeng) 2031.28 (nonbinding), b0 0.4 with doubling x4 2031.47; MTBF 60 h 2030.80, MTBF per the Zhaofeng claim 300 h (100-1,000) 2030.67; other assists 1.5/shift 2032.09; attempts 1/yr 2032.71; 15 active robots (Zhaofeng's scale) 2032.12; own failure 0.05/0.25 -> p_achieved 0.931/0.735 (p50 2031.11/2031.99); shared stall 0.08 -> 0.827, 2031.49; prev's full table (cfrac, absorption, K, changeover, setup, lag, dark-floor x2) unchanged in structure, each within +-0.7 yr. Discounted proponents: Aitu 2028-2030 multi-process scale-up (2026-07-23-xinhua-single-process-not-full-line) x1.5-3 -> 2030.1-2033.4; Zhaofeng's chairman expects the 15 s cycle in 6-12 months -> 2027.5-2029.8 for line-rate throughput, consistent with the nonbinding gate. Prev (assess:20261015T100057Z) 2028/2031/null, p_achieved 0.882, reproduced exactly (2028.97/2031.28).
- Data as of
- baseline 2026-06 (Longcheer); driver 2026-08 (Dyna); context to 2026-10 (Zhaofeng)
- Change
- yellow -> yellow. New evidence: 2026-10-16-leju-kuavo-zhaofeng-parts-loading-r2 and 2026-10-18-zhaofeng-kuavo-cycle-gap-lights-out-goal (15 humanoids at 15 stations with workers present, 35-40 s vs 15 s cycle, robot recalibration and lights-out explored, robots fill gaps between automated processes). No input contradicted (rule 7): the cycle ratio 0.40 is at the low end of the assumed b0 range and the gate stays nonbinding; the MTBF claim is ambiguous company data (sensitivity 2030.67); the lights-out goal and the gap-filling quote bear on scenario D1 in opposite directions and leave it at 0.08. Prev reproduced exactly; rerun at made_at 2026-10-19 on the same driver draws: 2028.97/2031.28/null, p_achieved 0.882 (conditional p10 2028.89 -> 2028.88). No move. The fast-mapping sensitivity now uses the task-horizon driver (metr-horizon retired): 2029.41 -> 2029.23. Status yellow.
- Rubric rule
- Milestone yellow (RUBRIC v6 §4): measurable progress on the drivers the milestone is tied to (multi-robot real-line operation, now 15 robots at 15 stations; MTBF, task success, autonomous charging/battery swap), median ETA 2031 <= made_at + 5 yr (2031.80) but the path is not fully evidenced (no robot-only multi-station precursor, no corroborated human-free 24/7 run), so yellow, not green.
by assess:20261019T010556Z · 23 evidence event(s)Evidence and news: 23 cited by the assessment
Cited evidence (23)
- Cailian Press reported on 18 October 2026, after a visit to Zhaofeng's Hangzhou auto-parts plant, where 15 wheeled Leju Kuavo humanoids load parts at 15 stations on workers' instructions, that each robot load took 35–40 s versus about 15 s for a skilled worker; Leju's CEO said accuracy and MTBF were already up to standard and the next step was to cut load time to the 15 s that matches the production cycle, which Zhaofeng's chairman expected in six months to a year; Zhaofeng is also exploring having robots do the pre-shift recalibration of a 26–28-station line at model changeovers (now 2–3 hours of work by three engineers) and is testing robot camera inspection toward 'lights-out factory' closed-loop operation (company statements; humans still work on the line, and Jiemian quoted Zhaofeng's general manager on 16 October as saying the robots fill the manual gaps between automated processes rather than replacing the whole automated line).The Dark LineThe "Blue Collar" Shift [1] 财联社: 人形机器人“上工”记 (18 Oct 2026) [2] 界面新闻 (Tencent News): 人形机器人开始进入汽车产线，但工厂要的不只是会干活 (16 Oct 2026)
- 21st Century Business Herald reported on 16 October 2026, after a site visit on 15 October, that 15 Leju humanoid robots were in routine operation on Zhaofeng Mechanical & Electronic's automotive-parts production line in Hangzhou, autonomously completing 10+ consecutive actions per cycle to fetch, transfer and load small parts onto loading stations in under 40 seconds with no human teleoperation, with the solution adaptable to 15 stations and covering 10+ part types, and a company-stated loading success rate above 98%, about 1,000 cumulative running hours over more than two months without failure and 60,000+ parts produced (company-provided data; the project is billed as an embodied-AI 'real-scene practical-training demonstration project').KPIThe "Blue Collar" Shift [1] 21世纪经济报道: 实拍乐聚人形机器人进厂上料，40秒内完成，全程无需遥操作 (16 Oct 2026) [2] 网通社 (ifeng): 乐聚机器人与兆丰机电联合发布“具身智能+汽车零部件上料解决方案” (16 Oct 2026; 15 Kuavo robots)↳ corrects (or adds sources to) an earlier record:
2026-10-16: 21st Century Business Herald reported on 16 October 2026, after a site visit on 15… - Boston Dynamics opened its Robotics Metaplant Application Center (RMAC) at Hyundai Motor Group Metaplant America near Savannah, Georgia on 21 September 2026 as a testbed and training center where Atlas humanoids are 'now in training' on preparing the logistics and sequencing of automotive parts, with component assembly targeted by 2030; Hyundai Motor Group plans to begin Atlas deployment with 25,000 units at Hyundai and Kia plants over the next few years.
- Unveiling Digit 5 on 15 September 2026, Agility Robotics said Digit 4 had logged more than 65,000 operating hours across customer sites including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada over nearly three years and kept approximately 98% accuracy while on-task through its 100,000-tote GXO milestone, and claimed Digit 5's 90-minute battery with 9-minute charging gives a 10:1 run-to-charge ratio (vs 2:1 for Digit 4) enabling 20+ productive hours per day, with early access in H1 2027.KPIThe "Blue Collar" ShiftPower/Efficiency/Cost [1] Agility Robotics: Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale↳ updates 2025-11-20: Agility Robotics reported on 20 November 2025 that its Digit humanoids had moved over… (corrected version)
- Dyna Robotics reported on 27 August 2026 that at Din Tai Fung restaurants, where each robot must produce 1,500 table-ready napkins per 18-hour shift, its Dyna-1 model at launch (June 2025) folded about 35 napkins/hour with 75% meeting the quality bar (about 480 per shift, roughly 32% of the requirement), while Dyna-2 (August 2026) folds 95/hour with 93% passing, about 1,590 per day (106% of the requirement), and that Din Tai Fung is rolling the robots out across its network, with Dyna projecting a deployed fleet of hundreds by H1 2027.The "Blue Collar" ShiftGeneralization [1] Dyna Robotics: Not Just a Model, But a Product [2] Dyna Robotics on X: Din Tai Fung rollout thread
- In a Xinhua survey of Zhejiang factory deployments published 23 July 2026, Aitu Embodied Intelligence general manager Lou Lingtong said humanoids can now do single processes well but cannot yet run whole processes unmanned, projecting 2026–2028 as the validation period for single-process capability and 2028–2030 as the scale-up period for multi-process collaboration, while other interviewees said there is no certification system for the 'industrial-grade lifetime' reliability of robots and their components (firms validate by repeated on-site testing) and that fine factory manipulation places very high demands on fatigue-resistant design (structural durability, thermal management, power optimisation) and high-load continuous operation.The Dark Line [1] 从“上场”到“进厂”——“人形机器人量产元年”新观察 (Xinhua, 23 Jul 2026)
- AgiBot said on 29 June 2026 that during a six-consecutive-day livestream from a tablet mass-production quality-inspection section at Longcheer's Nanchang factory, multiple AgiBot humanoids logged more than 64 hours of operation, performed 64,828 production-line tasks across more than 4 production workflows alongside human operators with a 99.99% task success rate, and contributed to cumulative line output of 17,625 (company figures; interventions and teleoperation not quantified).The "Blue Collar" Shift [1] AGIBOT: AGIBOT Concludes Six-Day Real Factory Livestream
- Xinhua reported on 28 June 2026 that eight AgiBot G2 humanoids had taken over Longcheer's entire MMIT tablet-testing section in Nanchang (up from four robots at single stations in April), running 10 hours a day for six days (23–28 June) on a public livestream and logging, by day five, over 54 cumulative robot-hours, 14,925 units, 55,107 operations and a 99.987% success rate, while Longcheer said scaling from four to eight robots surfaced robot-to-robot collisions, arm clashes and lidar frame drops and that the line relies on a supporting maintenance team (day shift only, not 24/7; humans on the floor).
- Galbot and CATL reported in late June 2026 that a Galbot S1 heavy-duty humanoid (50 kg dual-arm payload, up to 8 hours of continuous operation on CATL batteries, vision-only positioning with 360° obstacle avoidance) had worked 7×24 hours on CATL's mass-production lines at its Ningde base for over three months since passing acceptance in March 2026, autonomously carrying heavy material bins between module and battery-pack stations amid human workers without remote control, described as the world's first 'routine autonomous operation' of a humanoid in a new-energy smart-manufacturing production setting (company/partner claim; robot count, uptime, interventions and battery-maintenance arrangements not disclosed).The Dark Line [1] First Heavy-Duty Humanoid Robot Powered by CATL Batteries Goes Live in CATL Factory (CATL, 24 Jun 2026) [2] Galbot's Galbot S1 begins operations at CATL's factory (Gasgoo, 26 Jun 2026) [3] 全球首次！北京人形机器人实现在新能源智能制造生产场景“常态化自主作业” (Beijing International Sci-Tech Innovation Center, 30 Jun 2026)
- Cheetah Mobile CEO Fu Sheng was reported on 2 June 2026 as saying that, by industry standards, harmonic reducers in humanoid joints last about 7,000–10,000 hours at rated load, so core joints would need replacement after 3.5–5 years on single-shift duty (~2,000 h/year) but after only about 1.5–2 years in three-shift operation; the report noted core-component wear accelerates sharply in 24-hour three-shift logistics/manufacturing use and that service life is a key factor customers weigh before deploying humanoids.
- The Economic Observer reported on 2 June 2026 that Robotera's humanoids (column-based lower body) at a China Post logistics centre in Longgang, Shenzhen fell from about 600 parcels/hour in the lab to about 100/hour on arrival (November 2025), reached about 900/hour by March 2026 and a stable 1,200/hour by end-May 2026 versus a human average above 1,000/hour (skilled maximum 1,200), after six months of on-site engineering support including night-shift supervision, with some of its 10+ logistics centres running robots on two to three shifts for 24-hour coverage and new sites now deployed in about 3 days with 2 hours of data.The Dark LineThe "Blue Collar" Shift [1] 机器人同事，开始撑起物流中心的夜班 | 百亿具身系列报道 (Economic Observer) [2] 星动纪元携手中国邮政，助力物流智能化升级 (Xinhua, 20 May 2026)
- Figure AI ended on 22 May 2026 an unedited livestream that began as a response to a challenge to prove humanoids could work an 8-hour shift at human speed without intervention, in which a rotating fleet of Figure 03 humanoids running Helix-02 sorted 249,560 packages over 200 continuous hours at its Sunnyvale headquarters, autonomously rotating units to charge on roughly 4-hour batteries; CEO Brett Adcock said it 'ran for 200 hours without a failure', which Humanoids Daily and Interesting Engineering describe as no catastrophic mechanical breakdown or system-halting crash, though viewers saw occasional dropped or mis-oriented packages (company facility, not a customer workplace; intervention counts not published).The "Blue Collar" Shift [1] Brett Adcock on X: We just wrapped what began as an 8-hour challenge — and it ran for 200 hours without a failure [2] Humanoids Daily: Figure AI pops champagne as autonomous marathon crosses 200 hours without hardware failure [3] Interesting Engineering: Figure's humanoid robots work for 200 hours, process 250k packages without failure
- In a 10-hour 'Man vs. Machine' parcel-sorting contest at Figure AI's own facility, held during its May 2026 livestream marathon and reported on 18 May 2026, a rotating fleet of autonomous Figure 03 humanoids running Helix-02 sorted 12,732 packages (2.83 s/package) versus 12,924 (2.79 s/package) for a human intern who took California-mandated meal and rest breaks, i.e. about 98.5% of the human's shift output on this single task (company-run contest, not a customer site).
- AgiBot's Yao Maoqing said around the 14 April 2026 Longcheer livestream that merging a humanoid into a production line requires 100% success, 7×24-hour stable operation and never becoming the line bottleneck — 24 hours a day for 7 days without error being 'very hard' — and that while traditional industrial robots reach MTBFs of thousands to tens of thousands of hours, humanoids are generally still at hour-level MTBF; Guancha reported that the AgiBot G2's MTBF had been raised from hour-level to a level supporting two-shift continuous operation.
- Four AgiBot G2 humanoids completed an eight-hour live-streamed shift on 14 April 2026 loading and unloading tablets at MMIT test stations on Longcheer's mass-production line in Nanchang, China, where Xinhua reported 18–20 s per operation and 310 units per hour, and AgiBot/Longcheer reported up to 310 units per hour (about 3,000 per shift), over 99.9% success in continuous operation, over 140 hours of cumulative continuous operation with downtime loss below 4% and 'minimal human intervention' (intervention count not disclosed).
- Securities Times and Wenhui Bao reported on 24 March 2026 that AgiBot partner and senior vice-president Yao Maoqing, in a keynote at the recent 'Zhihui Jinling AI Open-Source Talent Summit' in Nanjing, described the company's embodied-AI deployments, with the reports stating that its solutions had achieved 24-hour continuous uninterrupted operation at real industrial sites, with line throughput (UPH) at human level and a mean time between failures (MTBF) 'at the hundred-hour level' (company claim; sites, robot counts and MTBF definition not disclosed).
- Xiaomi said on 2 March 2026 that a humanoid robot completed its first three-hour continuous autonomous run at a self-tapping-nut loading station on its real automotive production line, meeting the station's 76-second cycle time with a 90.2% dual-side installation success rate (i.e. roughly one failed attempt in ten), working alongside an automated feeder, conveyor and tightening station.
- Figure AI CEO Brett Adcock said on 21–22 February 2026 that a Figure 03 fleet (seven units shown) now runs autonomously 24/7 at its Sunnyvale headquarters 'without any human babysitters': robots work until their battery is low, a docked robot is signalled to vacate the 2 kW foot-inductive charger (about 1 hour to full charge) and swap in, and units with hardware or software faults walk to a triage area while a replacement takes over — company claim with the tasks performed, uptime and intervention counts undisclosed.The Dark LinePower/Efficiency/Cost [1] Brett Adcock on X: 'Running 24/7 without any human babysitters has been really hard...' [2] Brett Adcock on X: 'Rain or shine, the machines don't sleep. Figure robots operate autonomously, 24/7' [3] Lights-Out in Sunnyvale: Figure 03 Fleet Shifts to 24/7 Autonomous Operation (Humanoids Daily)
- TechSpot reported on 26 January 2026, drawing on the Financial Times, that UBTech acknowledges its Walker S2 humanoids achieve only 30–50% of a human worker's productivity, that the robot is optimized for controlled, repetitive tasks such as moving parts between stations, stacking boxes and visual defect inspection, that it still needs a human operator to swap its hands or tools for different tasks, and that UBTech aims to reach 80% of human efficiency by 2027.The "Blue Collar" Shift [1] TechSpot: These humanoid robots can only work half as fast as humans
- Boston Dynamics unveiled the production version of its electric Atlas humanoid at CES on 5 January 2026, stating that when its battery runs low it autonomously navigates to a charging station, swaps its own batteries and returns to work (stated 4-hour battery, swapped in about 3 minutes), with 2026 deployments fully committed to Hyundai's RMAC and Google DeepMind — a self-maintenance capability claim, not a demonstrated continuous run.The Dark LinePower/Efficiency/Cost [1] Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry [2] CES 2026: Boston Dynamics Set to Ship First Atlas Humanoids (A3)
- UBTech announced on 17 July 2025 that its Walker S2 industrial humanoid can autonomously swap its own hot-swappable battery in about 3 minutes without human intervention or shutdown (dual-battery design, choosing swap or charge by task priority), which it says enables 24/7 operation; a Shenzhen Special Zone Daily report of 5 March 2026, republished by Shenzhen's Development and Reform Commission, described a Walker S2 at Zeekr's 5G factory walking to a charging station after moving a box of parts, hot-swapping its battery within three minutes and returning to its station, and said the robot heads to the nearest swap station when its charge falls below 20% (capability claim; no continuous multi-day run data published).The Dark LinePower/Efficiency/Cost [1] UBTech shows how its humanoid robot can work 24/7 with autonomous battery swap (CnEVPost) [2] UBTECH Walker S2 product page: Autonomous Battery Swapping, 24/7 Continuous Operation [3] 从“三分钟换电”到“千台量产下线” 优必选人形机器人加速落地千行百业 (Shenzhen Development and Reform Commission, Mar 2026)
- UBTech said on 3 March 2025 that dozens of its Walker S1 humanoids had been deployed in 'practical training' across several production areas of Zeekr's 5G Intelligent Factory (final assembly workshop, SPS instrumentation zone, quality inspection area, vehicle assembly station), coordinated by its BrainNet multi-robot software framework to perform collaborative sorting, collaborative handling of large, heavy workpieces and precision assembly — which UBTech framed as a move from single-agent autonomy to multi-robot "Swarm Intelligence" — with no run length, throughput, uptime or intervention figures disclosed.
- In a 28 February 2024 'Deployment Update: Putwall Automation' post, Agility Robotics said Digit continues to complete real tasks in real facilities, with workflows run on a continuous basis and 100% autonomously, at 75% of a human's speed and a 97% success rate, without gantry or tether (company claim; site, run length and intervention counts not disclosed).
Major Open Challenges
Dexterous Manipulation: Human-level hands requiring dense tactile sensing, compliant actuators, and real-time control across 10,000+ object types including deformables.22 event(s)
Events (22)
- Boston Dynamics unveiled on 1 October 2026 a new-generation hand for Atlas with four fingers and 13 directly actuated, backdrivable degrees of freedom (up from 7 in the previous hand, including a 4-DoF opposable thumb), dense pressure tactile sensors over the fingertips and palm and the previous hand's strength (carrying a 100 lb+ loaded minifridge), deliberately omitting a pinky to save cost and failure points; it said the hand is designed for high-fidelity simulation and sim-to-real reinforcement learning, with only initial sim-to-real results so far.
- Sharpa launched at IROS 2026 on 28 September 2026 its W02 dexterous hand — 21 active degrees of freedom in a body about 30% smaller than its W01, with a Dynamic Tactile Array at the fingertips (5 mN–30 N force range, 1 mm spatial resolution) plus electronic skin over most of the hand (0.1–20 N, 5 mm resolution) and IP54 protection — alongside a D01 manipulation robot with whole-body e-skin and an AE01 haptic exoskeleton data glove (product specifications, no task results).
- The Information reported on 25 September 2026 that hands, along with factory equipment and suppliers, are holding back Tesla's Optimus V3 ramp (from a few dozen robots a week in Q2 to several hundred, against a goal of 1,000+ a week by end-2026): each hand and forearm has more than 100 screws and small parts that workers assemble by hand, assembly equipment does not always align parts precisely so robots need rework, and some touch sensors have been unreliable, so Tesla developed a replaceable 'sensing glove' to add in next year's model (per The Next Web's summary of the paywalled report).
- GGII's humanoid-robot media arm counted about 35 financing rounds in China's dexterous-hand sector in 2026 through 19 September totalling over ¥6.73 billion, up 821% from ¥730 million in the same period of 2025, and 61 new dexterous-hand products launched in the year, adding that hybrid tendon-plus-direct-drive designs had become a hotspot and that tactile sensors are becoming standard equipment, with piezoresistive, capacitive and Hall-effect sensors dominating mass production.Dexterous Manipulation [1] 高工人形机器人 via Tencent News: 灵巧手的2026：融资涨8倍，收入、销量增2倍 (22 Sep 2026)
- Unitree launched the Dex5-S dexterous hand on 21 September 2026 at a starting price of ¥39,900: human-hand 1:1 size, 22 degrees of freedom with all 22 motors direct-drive and backdrivable, 620 g, 1 kg rated (2 kg maximum) payload and up to 1,000 Hz control — which trade press framed as a price cut for high-DoF hands, at a time when, per OFweek, 6-DoF humanoid hands sell for ¥6,000–8,000 and Inspire's hand fell from ¥50,000 in 2023 to ¥18,000 in 2026.
- +17 older event(s): the section's event log
Generalization: Few-shot skill acquisition, robust task transfer across environments, recovery from unexpected failures, and long-horizon replanning without human intervention.41 event(s)
Events (41)
- Vals AI's Robo-bench, introduced on 13 October 2026, has frontier models' coding agents control a simulated 7-DoF Franka arm and 29-DoF Unitree G1 humanoid in real time from only datasheets, cameras and sensors (no motion planner, grasping or walking controller; 24 mostly held-out tasks of 3 minutes each) and found the best agent, OpenAI's GPT-6.1 Sol, averaging 28.33% of milestone points overall, with the best arm-suite average about 50.8 points and every model below 2 points on the humanoid tasks.
- Stanford and Caltech researchers' HomeBody system (reported 27 September 2026) let a Unitree G1 humanoid driven by GPT-6 Astra explore an unfamiliar kitchen, build a digital twin and spatial memory, then tidy it and fetch items from drawers by calling a skill library, with limitations including model latency, overheating finger servos and high compute costs; no success rates were reported.
- Electrek reported on 25 September 2026, citing The Information, that three people familiar with Tesla's Optimus said its AI cannot yet reliably handle a wide range of tasks and behaves unpredictably in situations it was not trained on, that Optimus reportedly still takes several days to learn even basic tasks despite over 500,000 hours of training data, and that the units working in Tesla's factories are confined to controlled, supervised areas and programmed for specific tasks, even as Tesla builds several hundred V3 robots a week at Fremont, mostly for internal testing, training and data collection.
- Figure AI reported on 17 September 2026 that its Helix 2.5 model, pretrained on Figure's Index dataset of human behavior, ran three long-horizon whole-body tasks (living-room tidying, towel folding, bed making) zero-shot in 30 unseen Bay Area homes with unseen objects, reaching 56% full-task success (no partial credit) versus 9% for the same model without Index pretraining, and matched a Helix 02 policy trained in its evaluation environment using half as much task data.
- AgiBot reported on 9–10 September 2026 that its GE-Act 2.0 world-action model, evaluated zero-shot (no task fine-tuning or demonstrations) on 100 atomic tasks in 20 skill groups with scenes, backgrounds, lighting and object instances unseen in training, raised average success from 17.1% to 44.1% on its G1-OP robot (tasks with non-zero success from 39 to 76) and from 13.4% to 31.1% on its G2-90D robot as co-training data grew from 300 to 30,000 hours, without clear saturation, the G2-90D gain despite that robot making up under 2% of the data indicating cross-embodiment transfer (company evaluation).Generalization [1] AGIBOT Releases GE-Act 2.0, Providing the First Systematic Validation of a Scaling Path for Native World-Action Models (10 Sep 2026) [2] GE-Act 2.0: Pretraining and Scaling a World-Action Model (arXiv 2609.05588) [3] 智元发布GE-Act 2.0 能力随数据规模扩大可'顿悟' (Shanghai Securities News via Sina, 9 Sep 2026)
- +36 older event(s): the section's event log
Power/Efficiency/Cost: Extending 1-2 hour runtimes to full 8-hour shifts via battery breakthroughs or radical actuator efficiency gains, at commercially viable price points.31 event(s)
Events (31)
- 36Kr reported on 14 October 2026 that low-priced consumer humanoids reach their price points by cutting hardware — Unitree's R1 AIR at ¥29,900 (versus ¥39,900 for the standard R1) drops from 26 to 20 joint degrees of freedom and loses force control, and Noetix's ¥9,998 Bumi lacks a five-finger hand and depth sensing — and, citing manufacturers' specifications and third-party reports, that entry-level models under ¥100,000 typically run 2–3 hours per charge (some about 1 hour) with none offering autonomous recharging, versus 6–8 hours for commercial or flagship models above ¥100,000.Power/Efficiency/Cost [1] Is the 10,000-Yuan Level Humanoid Robot Merely a "Cyber Toy"? (36Kr)
- At its 6 October 2026 analyst and investor day (transcript filed with the SEC on 9 October), Agility Robotics said Digit 5's bill of materials will be about $150,000 per unit at commercial launch with a target of about $30,000 per unit at volumes of 10,000 robots a year, that Digit 4's bill of materials is roughly $125,000, and that actuators make up roughly 25% of Digit's current cost (company targets).
- The Information reported on 25 September 2026 that hands, along with factory equipment and suppliers, are holding back Tesla's Optimus V3 ramp (from a few dozen robots a week in Q2 to several hundred, against a goal of 1,000+ a week by end-2026): each hand and forearm has more than 100 screws and small parts that workers assemble by hand, assembly equipment does not always align parts precisely so robots need rework, and some touch sensors have been unreliable, so Tesla developed a replaceable 'sensing glove' to add in next year's model (per The Next Web's summary of the paywalled report).
- KAIST researchers reported in Nature on 23 September 2026 that RAIBO2, a quadruped designed from a whole-robot energy-loss model (force-transparent lightweight hardware, a low-resistance motor-drive circuit and a low-dissipation locomotion policy), completed a full 42.195 km marathon (Sangju, November 2024) in 4 h 19 min 52 s on a single battery charge with a total cost of transport of 0.25, below the human benchmark of 0.37, giving more than three times the travel range per charge of existing quadrupeds.
- Unveiling Digit 5 on 15 September 2026, Agility Robotics said Digit 4 had logged more than 65,000 operating hours across customer sites including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada over nearly three years and kept approximately 98% accuracy while on-task through its 100,000-tote GXO milestone, and claimed Digit 5's 90-minute battery with 9-minute charging gives a 10:1 run-to-charge ratio (vs 2:1 for Digit 4) enabling 20+ productive hours per day, with early access in H1 2027.KPIThe "Blue Collar" ShiftPower/Efficiency/Cost [1] Agility Robotics: Agility Unveils Digit 5 Humanoid Robot Built for Cooperatively Safe Work at Scale↳ updates 2025-11-20: Agility Robotics reported on 20 November 2025 that its Digit humanoids had moved over… (corrected version)
- +26 older event(s): the section's event log
Beyond the framework: 58 event(s)
- The IFR's World Robotics 2026 Service Robots report (30 September 2026) said global sales of professional service robots rose 24% to almost 250,000 units in 2025 (transport and logistics 117,500, a 47% share) and consumer service robots 37% to almost 34.2 million, and that about 7,000 full-size humanoids (above 140 cm) intended for commercial and professional uses beyond R&D and entertainment were sold worldwide in 2025, while most current humanoid applications remain specialized and often require human teleoperation.Beyond the framework [1] IFR: Global Sales of Professional Service Robots Surge 24% (30 Sep 2026)
- The IFR's World Robotics 2026 report (24 September 2026) said the global operational stock of industrial robots rose 9% to a record 5 million units in 2025 as annual installations rose 11% to more than 600,000, with China installing 354,000 (up 20%, 59% of the world total; Chinese suppliers 55% of their home market), the United States overtaking Japan as the second-largest market with almost 38,500 (up 12%) versus Japan's 36,219 (down 19%), and forecast installations of 655,000 in 2026 (up 9%) and 806,000 in 2029.
- CNMO reported on 18 September 2026 that Lei Jun, sharing a CCTV News visit to Xiaomi's Changping phone factory, said its in-house iBrain industrial-AI system cut engineers' manual parameter tuning (e.g. battery-cover assembly to a 0.05 mm gap) from 4 hours to 2 minutes and on-site anomaly handling from 2 hours to 3 minutes; CNMO added, citing official information, that the factory's modular lines (key processes 100% automated, 96.8% self-developed hardware) can switch products within 10 hours and turn out one phone every 6 seconds per line (600 per hour per line) — iBrain recommends parameters and anomalies are still handled on site, and no staffing or actual output figure was given (purpose-built lines, no general-purpose robots).Beyond the framework [1] 雷军发文展示小米昌平手机智能工厂 iBrain三脑协同 (CNMO via Sohu, 18 Sep 2026)
- China's National Bureau of Statistics reported on 16 September 2026 that industrial enterprises above designated size produced 729,352 sets of industrial robots in January–August 2026 (up 29.0% year on year; 96,174 in August alone, up 34.6%) and 13,931,863 sets of service robots (up 12.1%).
- Security researcher Olivier Laflamme disclosed on 27 August 2026 two independent root remote-code-execution chains in Unitree's G1 EDU humanoid (CVE-2026-76639, network-adjacent, and CVE-2026-76640, starting over Bluetooth Low Energy without pairing), after Unitree had patched a related cloud robot-ownership check in July 2026, with no fixed firmware release for either flaw verified as of the disclosure, The Hacker News reported.
- +53 older event(s): the section's event log
Latest news: the 8 most recently published of 270 events
- Cailian Press reported on 18 October 2026, after a visit to Zhaofeng's Hangzhou auto-parts plant, where 15 wheeled Leju Kuavo humanoids load parts at 15 stations on workers' instructions, that each robot load took 35–40 s versus about 15 s for a skilled worker; Leju's CEO said accuracy and MTBF were already up to standard and the next step was to cut load time to the 15 s that matches the production cycle, which Zhaofeng's chairman expected in six months to a year; Zhaofeng is also exploring having robots do the pre-shift recalibration of a 26–28-station line at model changeovers (now 2–3 hours of work by three engineers) and is testing robot camera inspection toward 'lights-out factory' closed-loop operation (company statements; humans still work on the line, and Jiemian quoted Zhaofeng's general manager on 16 October as saying the robots fill the manual gaps between automated processes rather than replacing the whole automated line).The Dark LineThe "Blue Collar" Shift [1] 财联社: 人形机器人“上工”记 (18 Oct 2026) [2] 界面新闻 (Tencent News): 人形机器人开始进入汽车产线，但工厂要的不只是会干活 (16 Oct 2026)
- At Beijing E-Town's 2026 Robot Warrior Challenge fall competition on 18 October 2026 — a ~6,500 m² outdoor staged emergency course of 14 tasks, run on surfaces left wet by the previous day's heavy rain, whose humanoid-only tasks were pulling open a door, explosive-hazard removal, emergency valve closure and fire extinguishing — 24 of the 29 teams competed by remote control and 5 by autonomous navigation, and the published results named placings in the remote-controlled humanoid, quadruped and wheeled-quadruped groups but only one autonomous winner (Zhishen 'Milk long B' in the autonomous wheeled-quadruped group), reporting no autonomous humanoid placing, whereas X-Humanoid's Tien Kung 3.0 had won the autonomous humanoid run of the spring edition in April.The "Fukushima" Test [1] Beijing Youth Daily (北青网): 14项实景关卡、29支赛队 2026机器人勇士挑战赛在北京亦庄举办 (18 Oct 2026; 24 remote vs 5 autonomous teams, full placings) [2] Beijing Daily (京报网): 机器人也“过五关斩六将”？这场面太燃了 (18 Oct 2026; humanoid tasks incl. 紧急关阀, group results) [3] Beijing Economic-Technological Development Area (organiser): 2026机器人勇士挑战赛秋季赛即将举办 (9 Oct 2026; 14 tasks, remote/autonomous × humanoid/quadruped/wheeled-quadruped groups)
- Generalist AI reported in August 2026 that its GEN-1.5 robot foundation model learns new short-horizon dexterous tasks in context from a single 3–12-second demonstration with no gradient updates, averaging 59% (±10% s.d.) success across 10 tasks such as unzipping a pouch and opening jars, rising to 83% (±9%) after 10 gradient steps on about 5 minutes (~50 demonstrations) of data per task and reaching 66.5% on a held-out task after one gradient step on one minute of data, while noting the tasks are simple and short-horizon and in-context skills are more brittle than fine-tuned ones (company evaluation).
- 21st Century Business Herald reported on 16 October 2026, after a site visit on 15 October, that 15 Leju humanoid robots were in routine operation on Zhaofeng Mechanical & Electronic's automotive-parts production line in Hangzhou, autonomously completing 10+ consecutive actions per cycle to fetch, transfer and load small parts onto loading stations in under 40 seconds with no human teleoperation, with the solution adaptable to 15 stations and covering 10+ part types, and a company-stated loading success rate above 98%, about 1,000 cumulative running hours over more than two months without failure and 60,000+ parts produced (company-provided data; the project is billed as an embodied-AI 'real-scene practical-training demonstration project').KPIThe "Blue Collar" Shift [1] 21世纪经济报道: 实拍乐聚人形机器人进厂上料，40秒内完成，全程无需遥操作 (16 Oct 2026) [2] 网通社 (ifeng): 乐聚机器人与兆丰机电联合发布“具身智能+汽车零部件上料解决方案” (16 Oct 2026; 15 Kuavo robots)↳ corrects (or adds sources to) an earlier record:
2026-10-16: 21st Century Business Herald reported on 16 October 2026, after a site visit on 15… - MarketWatch reported on 16 October 2026, after reporter Christine Ji hosted two Tau Robotics humanoids (modified Unitree G1s costing about $55,000 each before Tau's modifications) in her San Francisco apartment, that the robots climbed the stairs to her third-floor apartment with teleoperator assistance, vacuumed, mopped, wiped counters, loaded a dishwasher (struggling to grip a smooth glass and to retrieve a small knife) and slowly folded a towel (about four minutes) and denim pants while remote human operators at Tau's headquarters guided them, with Tau operating eight robots and having logged roughly 160 paid robot hours across nearly 100 cleanings ($30 per robot-hour); Tau's website states its robots are not fully autonomous, a supervising human deciding when to let them do tasks autonomously and when to take over and steer them — i.e. Tau's commercial in-home humanoid cleaning still depends on human help.The "Wozniak" Test [1] Can a $55,000 robot load the dishwasher? Inside my afternoon with two humanoid housekeepers (MarketWatch, 16 Oct 2026) [2] humanoid.guide — Tau Robotics humanoids clean homes for $30 an hour, with help (17 Oct 2026; summary of the MarketWatch report) [3] Tau Robotics — Humanoid Cleaning Service in San Francisco (FAQ: 'Are the robots fully autonomous? No')
- A Beijing News (新京报) survey of Chinese humanoid factory deployments published on 14 October 2026 (SAIC-GM/AgiBot's 'Nengzai No. 1', AgiBot G2 at Longcheer, Galbot S1 at CATL) concluded that the applications disclosed so far show some robots can take on specific tasks but that whether this holds over longer periods and more complex conditions still needs sustained-operation testing, warning that a rated runtime such as Galbot S1's official 8 hours with hot-swappable batteries is not the same as continuously completing production tasks, so line suitability must be judged from effective work, fault downtime and human takeovers during operation rather than battery life, and noting that factory training, line trials and routine production still coexist (e.g. Dongfeng's 'Xiaodong' humanoid was planned to enter its plant for practical training at end-October 2026 and join trial production by year-end).
- 36Kr reported on 14 October 2026 that low-priced consumer humanoids reach their price points by cutting hardware — Unitree's R1 AIR at ¥29,900 (versus ¥39,900 for the standard R1) drops from 26 to 20 joint degrees of freedom and loses force control, and Noetix's ¥9,998 Bumi lacks a five-finger hand and depth sensing — and, citing manufacturers' specifications and third-party reports, that entry-level models under ¥100,000 typically run 2–3 hours per charge (some about 1 hour) with none offering autonomous recharging, versus 6–8 hours for commercial or flagship models above ¥100,000.Power/Efficiency/Cost [1] Is the 10,000-Yuan Level Humanoid Robot Merely a "Cyber Toy"? (36Kr)
- Vals AI's Robo-bench, introduced on 13 October 2026, has frontier models' coding agents control a simulated 7-DoF Franka arm and 29-DoF Unitree G1 humanoid in real time from only datasheets, cameras and sensors (no motion planner, grasping or walking controller; 24 mostly held-out tasks of 3 minutes each) and found the best agent, OpenAI's GPT-6.1 Sol, averaging 28.33% of milestone points overall, with the best arm-suite average about 50.8 points and every model below 2 points on the humanoid tasks.
Rockets and Space
189 events · latest news published 2026-10-15 (7 days ago)news per week
What child hasn't looked at the starry sky in wonder? The universe promises many worlds humanity could spread to — removing some eggs from the "one basket", a new age of discovery, making all raw materials cheap, and vast opportunities for community independence. And if nothing else, transport by orbital flights is the next best thing to teleportation, allowing you to reach the opposite side of the globe in 90 minutes.
KPI
Cheapest launch to low Earth orbit, per kg (constant 2025 $)
1673 $/kg
2025 · source
next update due by 15 Jan 2027
Stalled: $1,673/kg (2025, Falcon Heavy); 2021-25 -0.04%/yr vs its own long-run -2.4%/yr · assessed 2026-10-18
The cheapest ride to orbit has cost about $1,673/kg (Falcon Heavy) every year since 2021: -0.04%/yr over 2021-2025, after falling 19%/yr over 2016-2020. The benchmark is the price's own long-run decline of 2.4%/yr since 1962; against it the 2021-2025 trend is flat, so the status is red (Stalled). Prices published in 2026 do not break the stall: Falcon 9's list price rose to $74M (about $3,360/kg) and SpaceX's rideshare price rose too. The one thing that could change this: Starship's first disclosed launch price ($90M for the Starlab station, about $900/kg at its 100-tonne design capacity) would roughly halve the 2026 figure if it counts as a published price for a rocket in service.
The KPI's precise definition
The KPI, as defined: Cost to LEO ($/kg): the cheapest published price per kg to low Earth orbit of a dedicated launch among the rockets in service, inflation-adjusted (constant 2025 $)
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- cost-to-leo-best-real: cheapest published price per kg to LEO of a dedicated launch among the rockets in service each year (JSR GCAT), constant 2025 $ (BLS CPI-U); list prices (Falcon 9, Falcon Heavy) where published, CSIS $/kg otherwise; goal_direction down. A year without a list row holds the last listed year's real value (collector rule, commit 1d31c3c).
- Window
- Current
- Span
- 2021-2025
- Values 2025 $ per kg
- 2021167620221673202316732024167320251673
- OLS log slope per yr
- -0.0004
- OLS log slope SE
- 0.0002
- % per yr
- -0.04
- Endpoint change %
- -0.18
- Half means
- 2021-2022
- 1674
- 2023-2025
- 1673
- Cheapest
- Falcon Heavy every year; 2021 list $90M/63,800 kg, 2022 $97M/63,800 kg; 2023-2025 hold the 2022 list price in real terms
- Previous
- Span
- 2016-2020
- Values 2025 $ per kg
- 2016364820173572201818092019177620201755
- OLS log slope per yr
- -0.2162
- OLS log slope SE
- 0.0666
- % per yr
- -19.44
- Endpoint cagr %
- -16.72
- Cross check falcon9 nominal usd per kg
- 2021
- 2719
- 2022
- 2939
- 2023
- 3045
- 2024
- 3170
- 2025
- 3175
- 2026
- 3364
- Note
- list-price-falcon-9 / list-leo-kg-falcon-9 (2026: $74M / 22,000 kg); real 2021-25 ~3,231 -> 3,175 (prev), flat; 2026 CPI not yet published
- Cross check rideshare nominal usd per kg
- 2020
- 5000
- 2021
- 5000
- 2022
- 5500
- 2023
- 5500
- 2024
- 6000
- 2025
- 6500
- 2026
- 7000
- Note
- rideshare-price-per-kg-spacex (side series): rising
- Noise
- the series moves only by list-price steps and CPI timing; Falcon 9 real $/kg wanders ±1% year to year, so -0.04%/yr is flat within noise
- Source
- uv run core/ledger.py series rockets cost-to-leo-best-real --as-of 2026-10-18 (64 rows, 1962-2025, unchanged since prev); OLS by kpi_calc.py in ledger/assessments/notes/rockets/20261018T191702Z.md
- Benchmark
- tier 3: own long-run decline of the best $/kg: OLS -2.37%/yr over 1962-2025 (17,703 -> 1,673 $/kg); 1996-2025 OLS -5.16%/yr. The current window, -0.04%/yr, is far below either and not clearly negative.
- Benchmark source
- own series (cost-to-leo-best-real: published list prices deflated with CPI, CSIS $/kg where no list price; rockets in service per JSR GCAT)
- Transients discounted
- None. The stall spans the full window and has no named cause with an expected reversal. Starship pricing is not yet an observation: the $90.0M Starlab launch commitment (2026-03-10-starlab-starship-launch-90m; ~900 $/kg nominal at the 100 t V3 capacity, 2026-05-20-spacex-s1-starship-payload-h2-2026) is a contract disclosure for a future launch, Starship reached orbit with payload only on 2026-09-28, and whether the collector's 2026 row counts it is open (crux). The ~$100/kg figure and the S-1 -99% target are projections (rule 5).
- Data as of
- 2025 (row published 2026-01-15); annual cadence, the 2026 row is due after year-end: not stale.
- Change
- No change in status or numbers: no new observation of cost-to-leo-best-real (still 1962-2025, the 2026 row is due after year-end); kpi_calc.py reproduces prev's window statistics exactly. New since prev (all published before prev's made_at, merged by the 2026-10-18 gather): Falcon 9 2026 list price $74M (3,364 $/kg nominal, up), rideshare 7,000 $/kg (up), Starship's first disclosed dedicated-launch price ($90M for Starlab), Kinetica-2 cost claim, the PNAS Nexus experience-curve study (an average-cost series, not the cheapest list price), and SpaceX turning away dedicated Falcon customers after 2028; none is an observation of the assessed metric. Label: removed '(tier 3)', which RUBRIC §6 keeps out of labels (the tier stays in basis.benchmark).
- Rubric rule
- RUBRIC.md v6 §3: red when the window trend is flat within noise (stalled). Current 5-yr window 2021-2025 vs previous non-overlapping 2016-2020; benchmark tier 3 (own long-run trend).
by assess:20261018T191702Z · 10 evidence event(s)
Evidence and news: 10 cited by the assessment, 6 other KPI news
Cited evidence (10)
- Starship Flight 14 on 28 September 2026 became the first Starship to reach Earth orbit and its first to deliver meaningful payload there, deploying 26 Starlink V3 satellites before performing its first deorbit burn and splashing down on target in the northern Pacific.
- Bloomberg and SatNews reported in July 2026 that SpaceX was turning away satellite operators seeking dedicated Falcon 9 launches beyond 2028 and had stopped taking new Transporter and Bandwagon rideshare bookings beyond late 2028; Ars Technica reported that one source said SpaceX told the US government it intends to end commercial launches after 2028, a plan not publicly confirmed by SpaceX.
- A PNAS Nexus study (Terzi and Nicoli, 4,400+ launches 1960–2025) estimates the average cost of sending a kilogram to orbit fell from $87,023 in 1960 to $3,868 in 2025, by 21.2% per doubling of cumulative payload, and projects in its central estimate about $1,600/kg by 2030 and $300/kg by 2040.
- SpaceX's S-1 (filed 20 May 2026) says Starship V3 is designed to deliver 100 metric tons to orbit fully reusably, that SpaceX expects Starship to commence payload delivery to orbit in the second half of 2026, and that deploying 100 GW per year of orbital AI compute would need thousands of launches and about one million metric tons to orbit annually.
- SpaceX's IPO registration statement (S-1, filed 20 May 2026) cites NASA figures of about $2,700/kg to orbit for the first Falcon 9 (2010) and about $1,400/kg for the first Falcon Heavy (2018) against a historical average of $18,500/kg, and says SpaceX aims for Starship to cut the cost to reach orbit by 99% or more relative to that average.
- CAS Space's Kinetica-2 (Lijian-2, 12 t to 200 km LEO) reached orbit on its 30 March 2026 debut, and its chief commander Yang Haoliang said its single-launch cost without booster recovery is 30,000 yuan per kg (about $4,350/kg), which he called roughly on par with Falcon 9's $5,000/kg, and said booster recovery could later bring it to about half of SpaceX's figure.
- Voyager Technologies' 2025 annual report (10-K, filed 10 March 2026) discloses a $90.0 million commitment for one launch for the Starlab space station program, whose launch contract is with SpaceX's Starship, a rare public price for a dedicated Starship launch.
- SpaceX's Capabilities & Services sheet, first archived in its 2026 version on 4 February 2026, lists a dedicated Falcon 9 launch at $74 million (standard payment plan, through 2026) for up to 22,000 kg to LEO on a fully expendable vehicle, about $3,364/kg nominal; Ars Technica reported the increase as from $70 million.
- SpaceXStock’s comparison lists LEO launch costs of $1,500/kg for Falcon Heavy, $2,720/kg for Falcon 9, $4,044/kg for Vulcan, and $19,039/kg for Electron; it gives no New Glenn cost.↳ corrects (or adds sources to) an earlier record:
2025: SpaceXStock's 2025 launch-market comparison lists launch cost to LEO per kg of about… - A NextBigFuture analysis (Aug 2025) projected that SpaceX Starship with reuse could bring launch cost to LEO down to about $100/kg.↳ corrects (or adds sources to) an earlier record:
2025-08: A NextBigFuture analysis (Aug 2025) projected that SpaceX Starship with reuse could…
Other events (6)
- ESA's annual contract disclosure shows that launching Sentinel-1D on an Ariane 6 in November 2025 cost €82,070,773, European Spaceflight reported.
- Ill-Defined Space's 2025 global orbital launch summary puts SpaceX's share of global payload mass to orbit at ~85% in 2025, versus ~86% in 2024 and 60% in 2020.
- China-in-Space estimates the cost to LEO of China's Long March 3B at about $4,412/kg.
- Jonathan's Space Report counts 3,194.0 tons of payload launched worldwide in 2025, its biggest annual total, and 2,625.9 tons in 2024 (payload tonnage).KPIThe Deluge [1] Jonathan's Space Report: annual payload mass statistics [2] Jonathan's Space Report: Satellite statistics: Payloads↳ corrects (or adds sources to) an earlier record:
2025: Jonathan's Space Report's annual table of launched payload mass gives world totals of… - Wikipedia lists Electron’s starting price for delivering payloads to orbit at about $25,000 per kg.↳ corrects (or adds sources to) an earlier record:
2025: Rocket Lab's Electron small launcher costs about $19,039/kg to LEO (Wikipedia figure… - +1 older event(s): the section's event log
Milestone Countdown
A sequence: each milestone is expected after the one before it.
Expected year: dot = median (hollow: after 2100), bar = 80% range; arrows: each milestone is expected after the one before. The time axis is compressed: the farther out, the more years per step.
The Deluge: 100,000 tonnes delivered to orbit in a single year — the threshold to enable Mars colonization.Progressing: 3,194 t to orbit in 2025 (3.2% of goal); +45%/yr 2021-25, 2026 ~flat; ETA ~2036 · assessed 2026-10-18 · 30 event(s)
- Cargo plus propellant delivered by tanker flights counts; the launch vehicles themselves do not.
Annual payload mass to orbit rose from 797 t in 2021 to 3,194 t in 2025, +45%/yr, but quarterly industry tallies show 2026 running only about 6% above 2025 in its first half: Falcon 9's launch rate has plateaued and Starship delivered its first orbital payload only in September 2026. Projecting the 2021-25 growth rate forward from that lower 2026 level gives an 87% chance of a 100,000-tonne year by 2100 and a median of 2036 (10% chance by 2034). Yellow, not green: the median is about 10 years out, and reaching it depends on Starship flying hundreds of times a year, which it has not yet demonstrated.- ETA
- ~2036 (80% range 2034–never/after 2100; P(by 2100) 87%)
- ETA cruxes
- Earlier: BryceTech Q3/Q4 2026 and JSR's 2026 total above ~3,700 t (H2 >= +25% y/y); monthly operational Starship payload flights in 2027; the LC-39A operator licence and an approved Starbase PEIS (authorization beyond today's 25/yr toward ~500 flights/yr), a third site; evidence that JSR's tally omits tanker propellant. Later: a 2026 total near or below 2025's (~3,180 t) or a flat 2027 (-> 2037); Starship grounding or losses (stall branch P 0.25 -> 2039); demand stalling beyond constellations.
- Path
- Own quantity: annual mass delivered to orbit, read as metric tonnes of cargo plus tanker-delivered propellant delivered to Earth orbit or beyond in one calendar year, launch vehicles excluded; measured by JSR annual payload tonnage, nowcast for 2026 from BryceTech's quarterly upmass (the same outcome, reported quarterly; JSR's year-on-year ratio assumed equal to Bryce's). Near term: Falcon 9/Starlink (~70-80% of mass, cadence flat in H1 2026), Amazon Leo, Chinese constellations and new reusable boosters; 100 kt/yr needs ~1,000 flights/yr of a ~100 t fully reusable Starship-class vehicle (or ~500 at 200 t), so growth must pass from Falcon 9 to Starship over 2027-2030. Starship V3 delivered its first orbital payload (26 Starlink V3) on 2026-09-28.
- Blockers
- Technical: Starship not yet operational (first orbital payload flight 2026-09-28; flights 12-13 suborbital; ship-to-ship propellant transfer slipped to Q1 2027)
- Regulatory capacity: Starship currently authorized for 25 launches/yr at Starbase; LC-39A's 44/yr has environmental clearance (EIS/ROD) but no FAA operator licence yet; expansion to 100/456 Starbase launches/yr is only proposed (PEIS begun October 2026)
- Momentum: 2026 runs ~+6% y/y worldwide in H1 (SpaceX's own mass -5.5%), ~0.8 yr behind the 2021-25 trend
- Measurement: ton unit (metric vs short) and whether JSR's payload tonnage counts tanker-delivered propellant; JSR's 2026 total not before January 2027
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend: log-linear on payload-mass-to-orbit via the shared driver launch-mass (draw i of its mass_2025_t and log_slope -> draw i of this simulation, n=10000), with the 2026 step nowcast from the faster-reporting BryceTech quarterly upmass and three named failure scenarios; one Monte Carlo distribution; cross-checked by a bottom-up Starship-cadence and launch-licence count, a discounted proponent date and the Falcon 9 ramp
- ETA calculation
- Gap 100000/3194.0 = 31.31x (4.97 doublings). series rockets payload-mass-to-orbit --as-of 2026-10-18 (unchanged), 2021-25: 797.4, 1046.4, 1462.8, 2625.9, 3194.0 t -> OLS ln slope 0.3695 (se 0.0321) = +44.7%/yr, doubling 1.88 yr; on trend 2026 4,622 t. 2026 measured by the faster series (BryceTech): 2025 2,712.0 t (2026-04-08-brycetech-2025-upmass-2712-t); Q1 2026 647.4 t +18% y/y (2026-06-15-brycetech-q1-2026-upmass), Q2 732.5 t -2% (2026-08-24-brycetech-q2-2026-upmass) -> H1 1,379.9 vs 1,296.1 t = +6.47%; SpaceX own H1 1,041 vs 1,102 t = -5.5% (2026-08-04-spacex-10q-h1-2026-mass-down). Model: ln M(2026) = ln mass_2025_t + ln r_i, r_i = (H1 2026 + q_i x H2 2025 1,415.9 t)/2,712.0, q_i ~ split-LN(1.07; p10 0.93, p90 1.25) assumed (H1 rate, Starship's first orbital payloads offsetting Q2's -2%); r p10/p50/p90 0.995/1.068/1.162 -> 3,178/3,412/3,710 t JSR-equivalent (JSR ratio = Bryce ratio, assumed), ~0.8 yr of trend growth behind; then growth at driver launch-mass log_slope_i from 2026 (ledger.py driver-draws launch-mass --as-of 2026-10-18, seed 20261015, n 10000; split-normal p10 0.20/p50 0.3695/p90 0.47, draws 0.1998/0.3684/0.4707; p50 measured, spread assumed); first year with ln M >= ln 100000 + e_i, e ~ N(0, 0.07) assumed. Failure (MC masses): S1 demand plateau 0.084 (class: 2 earlier ramps in this series, the demand-driven one 1957-68 followed by a 51-yr plateau; failing now needs <4.7%/yr for 75 yr), S2 debris/regulatory cap 0.027 (ESA 2026 runaway projection vs a proposed expansion to 456 Starbase launches/yr (PEIS begun) and NSPM-17: both directions, kept), S3 collapse of access 0.010, trend too slow 0.006; none >= 10%. Result p_achieved 0.872; p10 2034, p50 2036 (P by end-2035 0.411, 2036 0.516, 2037 0.589), p90 null; if achieved 2034/2036/2043; P by 2040 0.728. Sensitivities (same draws): H2 y/y 0.93 -> 2037, 1.25 -> 2036; SpaceX's H1 ratio for the year -> 2037; second stalled year (2027 = 2026's step) -> 2037; reversion to the 2021-25 line (prev's model) -> 2035; fixed +21.6%/yr -> 2044, +30% -> 2039, +60% -> 2034; stall branch P 0.25 of 5-15 yr -> 2039 (unmeasured, not in median); failure 0.05 -> 0.942, p90 2049; 0.25 -> 0.741, 2037; short tons -> 2037; JSR omitting tanker propellant (share -> 0.5 by 2035) -> 2035, x2 bound -> 2034 (earlier only). Bottom-up: 100 kt needs ~1,000 flights/yr at 100 t or ~500 at 200 t; currently authorized 25/yr at Starbase (~2.5 kt at 100 t; 2026-10-14-faa-starbase-456-launches-peis); LC-39A 44/yr environmentally cleared (final EIS/ROD) but its FAA operator licence still outstanding (2026-01-29-faa-lc39a-starship-44-launches); Starbase 100 (Phase 1) / 456 (Phase 2) only proposed, PEIS just begun. Even 456 + 44 = 500 is half of 1,000: 2036 needs the LC-39A licence, the PEIS approved, plus another site or 200-t flights; regulatory capacity is no model input (cap risk in S2), so no number moves. Proponent: Musk 1 Mt/yr in ~3 yr (2026-06-09-musk-million-tons-three-years) x1.5-3 on 2.6 yr remaining -> 2031-2035 for 1 Mt/yr, earlier than our p10; weighted down (roadmap, measured 2026 step ~+7%, Starship schedules slipped >3x). Driver note: the launch-mass fit (2021-25) predates the 2026 quarters, whose ~+7% step is below its p10 slope; used as is, flagged for the owner. Prev (2026-10-18, assess:20261018T191702Z): 0.872, 2034/2036/null, reproduced byte-identically on unchanged data and driver draws; all inputs kept.
- Window
- Current
- Span
- 2021-2025
- Values t
- 2021797.420221046202314632024262620253194
- Loglin growth % per yr
- 44.7
- Loglin slope
- 0.3695
- Loglin slope SE
- 0.0321
- Doubling yr
- 1.88
- Previous
- Span
- 2016-2020
- Values t
- 2016337.42017394.52018388.92019399.22020572
- Loglin growth % per yr
- 11.3
- Loglin slope
- 0.1068
- Loglin slope SE
- 0.0364
- Nowcast 2026
- Bryce t
- 2025
- 2712
- H1 2025
- 1296
- H1 2026
- 1380
- Q1 2026
- 647.4
- Q2 2026
- 732.5
- H1 year on year
- 1.065
- Spacex h1 year on year
- 0.9446
- Jsr equivalent 2026 t p10 p50 p90
- 3178, 3412, 3710
- On trend 2026 t
- 4622
- Long run
- 2016-2025_pct_per_yr
- 30
- 2011-2025_pct_per_yr
- 14.7
- Source
- uv run core/ledger.py series rockets payload-mass-to-orbit --as-of 2026-10-18; BryceTech quarterly events; eta.py in ledger/assessments/notes/rockets/20261018T191702Z.md
- Change
- Prev (2026-10-18-milestone-the-deluge, assess:20261018T191702Z, same made_at) reproduced byte-identically: its eta.py, run unchanged on the same ledger, gives the same output (0.872, 2034/2036/null). No new evidence: the ledger's rockets events and observations are unchanged since prev's run (last data merge f980611, which prev used), the launch-mass driver draws are identical (md5 5a3042a8...). Prev's inputs were re-checked against the event claims and the rubric and no error was found, so every input is kept (rule 7) and nothing moved; status yellow unchanged (rule 1: no newer evidence). Review-1 D1: the inherited statement '69 launches/yr licensed' was corrected to 25 authorized (Starbase) + 44 environmentally cleared awaiting a licence (LC-39A) + 100/456 proposed; no licence count is a model input, so no number moved.
- Rubric rule
- Milestone yellow: measurable progress on the milestone's own quantity and median ETA 2036 <= made_at + 10 yr (RUBRIC.md §4); not green because p50 > made_at + 5 yr and there is no demonstrated precursor at relevant scale (one orbital Starship payload flight, no operational cadence).
by assess:20261018T211044Z · 26 evidence event(s)Evidence and news: 26 cited by the assessment, 4 other event(s)
Cited evidence (26)
- On 14 October 2026 the FAA announced its intent to prepare a Programmatic EIS on SpaceX’s proposal to raise Starship operations at Starbase from the currently authorized 25 launches and 50 landings per year to up to 100 launches and 200 landings a year in Phase 1 and up to 456 launches and 912 landings a year in Phase 2.
- NASA's Human Landing System presentation at IAC 2026 (5-9 October) schedules SpaceX's Starship ship-to-ship propellant transfer demo for Q1 2027, after a first ship catch and first Cape flight in Q4 2026, whereas SpaceX's October 2025 update had targeted the in-space propellant transfer flight test for 2026.
- The U.S. DOT announced five proposed FAA rules to streamline Part 450 commercial launch and reentry licensing (electronic applications, more flexibility to show equivalent safety, combining three safety analyses into one submission), citing 217 of 329 commercial launches worldwide in 2025 by the U.S. and an expectation of 10,000 FAA-licensed launches and reentries a year by 2035.
- Starship Flight 14 on 28 September 2026 became the first Starship to reach Earth orbit and its first to deliver meaningful payload there, deploying 26 Starlink V3 satellites before performing its first deorbit burn and splashing down on target in the northern Pacific.
- ESA's Space Environment Report 2026 (data to end-2025) says more than 300 launches put over 4,000 payloads in orbit in 2025, more than three intact satellites or rocket bodies now reenter daily, and projected LEO object and collision numbers are rising in a runaway (Kessler-type) effect even with the projection horizon cut from 200 to 100 years, making active debris removal necessary.Beyond the framework [1] ESA Space Environment Report 2026 (ESA)
- BryceTech's Q2 2026 Global Space Activity briefing (24 August 2026) estimates 732,519 kg of spacecraft mass launched worldwide in Q2 2026, down 2% from Q2 2025, on 81 orbital launches.
- The new U.S. National Space Transportation Policy (NSPM-17, signed 20 August 2026, replacing the 2013 policy) says U.S. space transportation ranges must grow to support more than 1,000 launches and reentries every year by 2030 and makes reentry infrastructure a national priority.
- SpaceX's 10-Q for the quarter ended 30 June 2026 (filed 4 August 2026) reports mass to orbit of 485 metric tons in Q2 2026 versus 652 t in Q2 2025, and 1,041 t in H1 2026 versus 1,102 t in H1 2025, a year-on-year decline (Falcon launches fell from 81 to 77 in the half).
- Starship Flight 13 on 24 July 2026, the second V3 flight, deployed 20 Starlink V3 satellites on its suborbital trajectory (expected to demise on reentry), relit a Raptor in space and made a soft splashdown intact in the Indian Ocean, while the Super Heavy booster again ended in a hard splashdown in the Gulf.
- BryceTech's Q1 2026 Global Space Activity briefing (15 June 2026) estimates 647,412 kg of spacecraft mass launched worldwide in Q1 2026, up 18% from Q1 2025, with 1,216 spacecraft launched.
- In an IPO-roadshow interview filed with the SEC on 9 June 2026, Elon Musk said SpaceX aims to go with Starship from somewhere around 2,500 tons a year to orbit to millions of tons per year, and that it can probably reach a million tons to orbit per year in about three years.
- Starship Flight 12 on 22 May 2026, the first flight of the V3 Starship and Super Heavy with Raptor 3 engines and from Starbase Pad 2, put the ship on its planned suborbital trajectory, deployed 20 Starlink simulators and two modified Starlink satellites and splashed down in the Indian Ocean, while the booster ended its boostback early and hit the Gulf in a hard splashdown.
- SpaceX's S-1 (filed 20 May 2026) says Starship V3 is designed to deliver 100 metric tons to orbit fully reusably, that SpaceX expects Starship to commence payload delivery to orbit in the second half of 2026, and that deploying 100 GW per year of orbital AI compute would need thousands of launches and about one million metric tons to orbit annually.
- SpaceX's IPO registration statement (S-1, filed 20 May 2026) reports its own mass to orbit (payload deployed by successful launches) of 2,213 metric tons in 2025, 1,699 t in 2024 and 1,210 t in 2023, and 556 t in Q1 2026 versus 450 t in Q1 2025, over 80% of the world's mass to orbit each year since 2023.
- BryceTech's Orbital Launches Year in Review 2025 (8 April 2026) estimates 2,712,023 kg of spacecraft mass launched worldwide in 2025 on 325 orbital launches, 2,225,403 kg of it by SpaceX.
- BryceTech's Q4 2025 Global Space Activity briefing (30 March 2026) estimates that 770,060 kg of spacecraft mass was launched worldwide in Q4 2025, with 1,404 spacecraft launched.
- On 4 February 2026 the FCC Space Bureau accepted for filing SpaceX's application for a non-geostationary system of up to one million satellites designed to operate as orbital data centers.
- Announcing its acquisition of xAI on 2 February 2026, SpaceX said only about 3,000 tons of payload were launched into orbit in 2025 and projected that Starship, launching every hour with 200 tons per flight, will deliver millions of tons to orbit per year, a million tons per year of compute satellites adding 100 GW of AI compute annually.
- The FAA's Final Environmental Impact Statement and Record of Decision for Starship-Super Heavy at Kennedy Space Center's LC-39A (notice dated 29 January 2026) covers up to 44 Starship-Super Heavy launches per year, with up to 44 Super Heavy and 44 Starship landings (SpaceX still needs an FAA license for these operations).
- Ill-Defined Space's 2025 global orbital launch summary puts SpaceX's share of global payload mass to orbit at ~85% in 2025, versus ~86% in 2024 and 60% in 2020.
- Jonathan's Space Report counts 3,194.0 tons of payload launched worldwide in 2025, its biggest annual total, and 2,625.9 tons in 2024 (payload tonnage).KPIThe Deluge [1] Jonathan's Space Report: annual payload mass statistics [2] Jonathan's Space Report: Satellite statistics: Payloads↳ corrects (or adds sources to) an earlier record:
2025: Jonathan's Space Report's annual table of launched payload mass gives world totals of… - China launched the 12th batch of Guowang satellites on Oct. 16, 2025.Beyond the framework [1] SpaceNews: China launches 12th group of Guowang broadband satellites as rocket milestone highlights accelerating launch rate↳ corrects (or adds sources to) an earlier record:
2025: China launched its 12th batch of Guowang satellites in 2025 for the planned… - Blue Origin's New Glenn landed its booster on its second flight on 13 November 2025 (launching NASA's ESCAPADE Mars mission), the first propulsive orbital-booster landing by a company other than SpaceX.Beyond the framework [1] CNN: New Glenn first booster landing↳ corrects (or adds sources to) an earlier record:
2025-11-13: Blue Origin's New Glenn landed its booster on its second flight on 13 November 2025… - ESA’s 2025 Space Environment Report estimates 54,000 objects larger than 10 cm and 1.2 million objects from 1 cm to 10 cm.Beyond the framework [1] ESA Space Environment Report 2025↳ corrects (or adds sources to) an earlier record:
2025-10-21: ESA's Space Environment Report 2025 counts more than 40,000 tracked objects in orbit… - SpaceX Starship Flight 11 on 13 October 2025 achieved its test objectives.↳ corrects (or adds sources to) an earlier record:
2025-10-13: SpaceX Starship Flight 11 on 13 October 2025 met all test objectives, closing out… - SpaceX Starship’s test launch on 26 August 2025 succeeded, releasing eight test satellites and splashing down in the Indian Ocean.The Deluge [1] Scientific American: SpaceX Successfully Launches Starship Spacecraft after String of Mishaps↳ corrects (or adds sources to) an earlier record:
2025-08-26: SpaceX Starship Flight 10 on 26 August 2025 succeeded, performing Starship's first…
Other events (4)
- A NextBigFuture analysis (Aug 2025) projected that SpaceX Starship with reuse could bring launch cost to LEO down to about $100/kg.↳ corrects (or adds sources to) an earlier record:
2025-08: A NextBigFuture analysis (Aug 2025) projected that SpaceX Starship with reuse could… - SpaceX Starship Flight 9 on 27 May 2025 flew the first reused Super Heavy booster, but the booster was lost.The Deluge [1] NBC News: Starship Flight 9 live updates↳ corrects (or adds sources to) an earlier record:
2025-05-27: SpaceX Starship Flight 9 on 27 May 2025 flew the first reused Super Heavy booster, but… - SpaceX Starship Flight 8 on 6 March 2025 failed when Ship 34 was lost during ascent.↳ follows up on 2025-01-16: Starship Flight 7 lifted off on January 16, 2025; Ship 33 was lost before second-stage… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-03-06: SpaceX Starship Flight 8 on 6 March 2025 failed when Ship 34 was lost during ascent. - Starship Flight 7 lifted off on January 16, 2025; Ship 33 was lost before second-stage engine cutoff, while Booster 14 was caught on the tower.↳ corrects (or adds sources to) an earlier record:
2025-01-16: SpaceX Starship Flight 7 on 16 January 2025 lost its upper stage (Ship 33 exploded),…
The Export: Product manufactured in space is sold commercially on Earth (beyond novelty items).Progressing: no recurring space-made product sold yet; drug-making flight due early 2027; ETA ~2036 · assessed 2026-10-18 · 21 event(s)
- Recurring commercial sales of a space-manufactured product line, from repeated production, to customers who use it for its function: sales from ongoing production, not stock from a few experimental flights (e.g. NIST's "space beads", sold from 1985), and not material supplied for clinical trials.
No space-made product is yet sold on Earth from repeated production, and none of today's ventures has reported even a first sale to customers who use its product. But the routes are funded and flying: Varda has returned six capsules and raised $251M to pursue a space-made drug, with its first flight to make a drug formulation for patients (with United Therapeutics) due in early 2027, while semiconductor, fibre and bioprinting firms keep flying demonstrations. A drug must still pass trials and approval, and materials sales are slower to become recurring; together that gives an 88% chance by 2100 and a median of 2036 (10% chance by 2031). Yellow, not green: the median is about 10 years out, every candidate is still at demonstration scale, and the evidence comes mostly from the companies themselves.- ETA
- ~2036 (80% range 2031–never/after 2100; P(by 2100) 88%)
- ETA cruxes
- Earlier: the early-2027 Varda/United Therapeutics flight returning a formulation that reaches an IND or first dosing by 2028; a paid repeat supply contract for ZBLAN fibre, semiconductor crystals or bioprinted implants. Later: the 2027 flight slips or yields no candidate (trial start 2028+ -> 2036-37); Varda's flights remain reentry/defence services (P(A) 1/3 -> 2041); first materials sales stay one-off batches (hazard 1.25%/yr -> 2040, P 0.66). Either way: a report that a candidate line (ZBLAN, crystals, implants) has already sold to a functional customer (P 0.25 -> 2034).
- Path
- Reading (d964534 details): recurring commercial sales of a space-made product line from repeated production, to Earth customers who use it for its function (materials to industry or research buyers, or a marketed drug); one-off batches and clinical-trial supply excluded. Paths: Varda capsules (space formulation with United Therapeutics from 2027 -> trials -> marketed product); Flawless Photonics ZBLAN; Space Forge and Besxar semiconductor crystals/wafers; Auxilium and Redwire biofabrication; return capacity from Varda and SpaceX Starfall; platforms Vast Haven-1 (Q1 2027), Axiom, NASA commercial stations (awards 2027).
- Blockers
- Regulatory/technical: no product has passed qualification for sale (pharma needs trials and approval; ZBLAN quality verification has no reported result)
- Commercial: a first sale to functional customers must become a recurring line from repeated production; no such sale and no repeat customer is reported in the ledger for any current candidate (historic one-off stock such as NIST's 1985 beads is excluded by the wording)
- Evidence: all supporting events are proponent-sourced (gather request: ZBLAN verification result and sales, Space Forge/Besxar crystal sales, Varda IND or trial start)
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- structural floor (drug trials and review, partly sourced) plus a thinned flat hazard for recurring materials lines, earliest of competing paths, named common-mode failure scenarios; one Monte Carlo distribution (prev's model, reproduced; trial start conditioned on the sourced first formulation flight, materials hazard restarted at made_at); cross-checked against a reformulation reference class, a discounted proponent date and the previous ETA
- ETA calculation
- Gap 0 -> 1 recurring product line (none reported in the ledger; historical one-off stock like NIST's 1985 beads is excluded by the wording; Varda in May 2026 'could not name anything manufactured in space, returned and sold', 2026-05-13-varda-united-therapeutics: one company's statement). Path A (Varda pharma, P 0.5): trial start 2025.91 + LN(1.7 yr; 80% 0.8-3.6), conditioned on >= 2027.5 (first patient-directed formulation flight with United Therapeutics 'early 2027', sourced; + ~2-3 months flight and return, assumed; Varda's 2026 flights carried defence/thermal payloads: 2026-01-29-varda-w5-reentry, 2026-10-01-varda-w8-w9-dual-launch) -> start p10/p50/p90 2027.63/2028.33/2030.25 (prev 2026.71/2027.61/2029.51); + programme LN(4.0; 2.5-7.0; FDA review ~1 yr sourced, rest assumed) + recurrence lag LN(0.5; 0.25-1.0). Path B (materials): first sales 5%/yr from made_at 2026.80 (assumed state, not a measured absence: every ledger candidate - ZBLAN quality verification ongoing, ForgeStar-1 no return, Besxar suborbital test 1 of 12, Auxilium/Redwire output for research, LambdaVision pre-clinical - is pre-sale as of 2026-10-18; memoryless) x P(recurring) 0.6 = 3%/yr, + repeat-order lag LN(1.5; 0.5-4.0). MC N=400k: p_achieved 0.877; p10 2031, p50 2036 (P by end-2035 0.498, 2036 0.532: on the 2035/2036 boundary), p90 null; if achieved 2030/2034/2068. Failure 0.123 (MC masses): C1 no product beats terrestrial alternatives 0.050, C2 loss of space access 0.020 (inherited; 0.01 as in Deluge/Homestead -> 2035, 0.886), C3 path A fails and no recurring materials line by 2100 0.053; none >= 10%. Path-A failure (0.5, within the model) cross-checked against a counted class: commercial microgravity production aimed at a marketed drug, ended: McDonnell Douglas/Ortho EOS = 0 of 1 -> 1/3 (sensitivity). Sensitivities (same draws): hazard origin 2026.0 -> 2035, 0.878; 2024.8 -> 2035, 0.880; P 0.10 / 0.25 that a candidate already had a qualifying first sale -> 2035, 0.880 / 2034, 0.885 (median on the 2035/2036 boundary, kept at 2036: judged P ~5%); earliest trial start 2027.25 / 2028.0 -> 2035 / 2036; P(A) 1/3 -> 2041, 0.859; 0.65 -> 2034; hazard 1.25% -> 2040, 0.660; 7.5% -> 2035, 0.912; programme 2.5/6 yr -> 2034/2037; recurrence 0.4/0.8 -> 2037/2035; common 0.03/0.15 -> 2035/2037. Reference class reformulation 5-8 yr from the 2027 flight + 0.5 -> 2032.5-2035.5 (path A if it succeeds). Proponent: no dated first product (Series D 'scale toward', 2026-09-30-varda-series-d-251m); background 'late 2020s' x1.5-3 -> 2031-2036. Return capacity grows (SpaceX Starfall demo launch and targeted splashdown, 1,000 kg payload capacity, 2026-06-23-spacex-starfall-demo; Redwire on Starfall 2028) but has no measured mapping to sale odds: context. Prev (2026-10-18, assess:20261018T191702Z): 0.877, 2031/2036/null, reproduced byte-identically; all inputs kept.
- Change
- Prev (2026-10-18-milestone-the-export, assess:20261018T191702Z, same made_at) reproduced byte-identically: its eta.py, run unchanged on the same ledger, gives the same output (0.877, 2031/2036/null). No new evidence: the ledger's rockets events and observations are unchanged since prev's run (last data merge f980611, which prev used), the launch-mass driver draws are identical (md5 5a3042a8...). Prev's inputs were re-checked against the event claims and the rubric and no error was found, so every input is kept (rule 7) and nothing moved; status yellow unchanged (rule 1: no newer evidence). Review-1: E1 the corrected Varda record's removed five-mission count is no longer cited (it fed no input); E2 the materials-hazard origin 2026.80 is relabelled an assumed state, with origin and alternative-state sensitivities (earlier origins -> 2035; kept 2036, disclosed as a boundary median). No number moved.
- Rubric rule
- Milestone yellow: measurable progress (repeated in-space production with return to Earth, funded drug-formulation programme) and median ETA 2036 <= made_at + 10 yr (RUBRIC.md §4); not green (p50 > made_at + 5 yr; no precursor at commercial scale; no non-proponent corroboration).
by assess:20261018T211044Z · 19 evidence event(s)Evidence and news: 19 cited by the assessment, 2 other event(s)
Cited evidence (19)
- NASA released its final Request for Proposals for commercial low Earth orbit space stations to follow the ISS on 9 October 2026, planning multi-award firm-fixed-price IDIQ contracts with proposals due 8 December 2026 and awards expected in spring 2027.
- Varda launched its W-8 and W-9 reentry vehicles together on SpaceX's Transporter-18 on 1 October 2026, the first time two commercial reentry vehicles went to orbit at the same time; they carry thermal-protection payloads and, on W-9, a navigation system for hypersonic flight, for commercial and government customers.
- Varda raised a $251 million Series D (total raised $598 million) on 30 September 2026 to scale toward the first pharmaceutical made in space for patients on Earth, having completed six successful reentry missions with over a dozen more launches and reentries planned through 2028.
- Redwire's SpaceMD contracted its first commercial mission on SpaceX's Starfall reentry spacecraft, planned for launch in 2028 with capacity for up to 32 PIL-BOX drug-crystallization units; since November 2023, 54 PIL-BOX systems have flown to the ISS and SpaceMD has crystallized 45 unique compounds for research.The Export [1] Redwire: SpaceMD announces its first commercial mission on SpaceX's Starfall (6 Aug 2026)
- Besxar's Flight-1 carried two wafer-carrying 'Clipper-class' canisters on a Falcon 9 first-stage booster to about 71 miles altitude on 5 July 2026 and recovered them intact (an 11-minute suborbital exposure test toward in-space compound-semiconductor manufacturing), the first of 12 contracted SpaceX missions.
- LambdaVision, which has flown nine ISS missions producing protein thin films for its pre-clinical artificial retina, signed a memorandum of understanding on 24 June 2026 with Vast for future research and manufacturing aboard Haven-1 and later commercial stations.
- SpaceX flew the first demonstration of Starfall, an uncrewed reentry capsule (3.1 m diameter, up to 1,000 kg of payload) aimed at returning in-space-manufactured products and other cargo, launching it on a Falcon 9 on 23 June 2026; SatNews reported a targeted Pacific splashdown, and a second demo flight is planned.
- Auxilium Biotechnologies bioprinted kidney and liver tissues on the ISS with its AMP-1 orbital bioprinter, the first of either tissue made in space, returned to Earth on mission AXLM-3, which splashed down in the Pacific on 17 June 2026, for research at the Wake Forest Institute for Regenerative Medicine (the printer has also produced 28 nerve-repair implants).
- Varda announced on 13 May 2026 a collaboration with United Therapeutics to process small-molecule medicines aboard multiple missions to low Earth orbit, seeking novel formulations for rare pulmonary disease; the research collaboration was billed as the first focused on space-based drug formulation that could result in patient therapies, with the first launch anticipated in early 2027, and Varda said it could not name anything manufactured in space, returned and sold.
- Space Forge reported that its ForgeStar-1 satellite has generated plasma more than 100 times in its in-house growth chamber in low Earth orbit, demonstrating on a free-flying commercial satellite the conditions needed for gas-phase semiconductor crystal growth (no material has been returned to Earth).
- Varda's W-5 capsule reentered and landed at the Koonibba Test Range in South Australia on 29 January 2026 after 9 weeks in orbit, the first full mission on Varda's in-house satellite bus; it carried a U.S. Navy payload collecting data for hypersonic reentry testing, funded through the AFRL Prometheus program.
- Vast began integration of its Haven-1 commercial space station on 20 January 2026 and moved its launch-readiness target to Q1 2027 on a Falcon 9, from a previously public mid-2026 launch date.
- Redwire bioprinted the first live human heart tissue in space and launched 14 PIL-BOX pharmaceutical crystallization experiments in 2025.
- MarketsandMarkets projects the in-space manufacturing market to reach $4.6B by 2030 and $62.8B by 2040.↳ corrects (or adds sources to) an earlier record:
2025: MarketsandMarkets projects the in-space manufacturing market to reach $4.6B by 2030 and… - Flawless Photonics has produced more than 12 km of ZBLAN optical fiber aboard the ISS, with quality verification ongoing.
- Axiom Space completed the first module’s final welds; testing of the pressure vessel was underway, and the revised station plan targets no earlier than 2028.The Export [1] Axiom Space: Year in review 2025↳ corrects (or adds sources to) an earlier record:
2025: Axiom Space completed welding of its first Axiom Station module, with the PPTM module… - Varda has completed a handful of missions and sees getting a drug into clinical trials as its next major milestone.The Export [1] TechCrunch: Varda says it has proven space manufacturing works [2] Wayback capture 20260102041149↳ corrects (or adds sources to) an earlier record:
2025-11-30: Varda Space Industries says it has completed five in-space pharmaceutical-crystal… - Space Forge returned a semiconductor-manufacturing payload from orbit to Earth in May 2025.The Export [1] Space Forge: payload returned to Earth
- Varda’s W-1 returned in February 2024, landing on U.S. soil under the FAA’s Part 450 licensing framework.The Export [1] TechCrunch: Varda says it has proven space manufacturing works. Now it wants to make it boring↳ corrects (or adds sources to) an earlier record:
2024-02: Varda Space Industries' W-1 capsule returned in February 2024, the first commercial…
Other events (2)
- Vast’s Haven-1 flight article is built; the company currently targets launch in 2027.The Export [1] Vast: Haven-1↳ corrects (or adds sources to) an earlier record:
2025: Vast's Haven-1 commercial space station flight article is complete, with launch… - Varda Space Industries launched its W-5 in-space pharmaceutical manufacturing capsule on 28 November 2025.
The Homestead: Permanent crewed base operates 1+ year using local resources (Moon or Mars).Distant: no crewed landing since 1972; space-grown food still experimental; ETA ~2069 · assessed 2026-10-18 · 61 event(s)
- Continuously crewed for 12+ months, and life-sustaining self-sufficient: the crew's air, water and food come from local resources and recycling for the whole year, with no resupply of consumables from Earth (equipment and spare parts may still come from Earth).
Space stations already recycle nearly all their water, and China's all its oxygen, but none grows a meaningful share of its crew's food, and a largely self-sufficient year has been shown only in a ground test; the first crewed Moon landing since 1972 is planned for 2028 and NASA's permanent outpost from 2032. Allowing for a landing, then a permanent base, then a decision to grow the base's own food and a decade to achieve it gives a 68% chance by 2100 and a median of 2069 (10% chance by 2044), so red. Whether anyone commits to growing a base's own food is the main uncertainty: funded requirements for it at both the US and Chinese bases would pull the median into the late 2050s.- ETA
- ~2069 (80% range 2044–never/after 2100; P(by 2100) 68%)
- ETA cruxes
- First whether anyone pursues food self-sufficiency: a funded Lunar Agriculture Module build, a crop-production requirement in NASA's Moon Base outpost phase, or a Chinese lunar greenhouse with a calorie target -> earlier (0.5/0.6 -> 2058); outpost plans that keep full resupply -> later (US at the station rate -> 2081). Then programme survival to a base (Artemis IV landing in 2028-29; abandonment 0.67 -> 2091), the closure interval (a greenhouse supplying > 25% of a crew's calories -> 5 yr -> 2063) and a funded settlement programme. Anchor: manifold-infkbfl3unm9gnq9qgky-p moving above ~0.6 for 2049 would put the market > 0.35 from us, to be examined against the pursuit crux.
- Path
- Reading (d964534 details): a base continuously crewed for 12+ months during which all oxygen, water and food consumed are produced on site from local resources or recycling (food grown from recycled water/CO2/nutrients counts; Earth food stocked and eaten in the year does not); equipment, spares and hardware may come from Earth. Moon or Mars. Paths: NASA's Moon Base (Artemis IV landing early 2028, outpost assembly from 2032, fission reactor by 2030) and China's ILRS (crewed landing before 2030, basic model 2035), each needing a continuous base and then bioregenerative food production (~58x the EDEN analog); SpaceX's stated self-growing lunar city; later a Mars settlement or successor programme. Water and air recycling are near closure (ISS 98% water; Tiangong >96% water, 100% O2); ISRU O2 at 12 g/h on Mars and 70 g/h in a ground vacuum chamber.
- Blockers
- Technical: no crewed lander demonstrated; Starship HLS propellant transfer demo slipped to Q1 2027; Axiom suits possibly not ready before 2031 (OIG)
- Technical: food closure: no station grows a meaningful share of crew food (Tiangong wheat and ISS crops are experiments); best analog greenhouse ~2% of a 4-crew diet; full closure demonstrated only on the ground (Lunar Palace)
- Technical: ISRU at demo scale (MOXIE 12 g/h; MRE 70 g/h on the ground); in-situ ice extraction not demonstrated (PRIME-1 could not excavate); polar surface ice sparse
- Programmatic: no programme plans food self-sufficiency for a base (Lunar Agriculture Module only a ground demonstrator, build decision by 2032); FY2027 request cuts NASA to $18.8B; Gateway paused; 3 of 4 2026 Moon Base landers slipped
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- structural floor chain (crewed landing -> continuously crewed base -> air/water/food closure -> literal self-sufficient year) with discounted proponent landing dates, reference-class intervals, disclosed pursuit judgments, a later-programme branch and named failure scenarios; one Monte Carlo over the US and Chinese programmes (earliest wins); prev's model reproduced and updated for new evidence; compared with one similar anchor after the pre-anchor forecast
- ETA calculation
- Prev's chain (assess:20261018T191702Z: 0.677, 2044/2069/null; N=400k, same seeds) reproduced byte-identically; all inputs kept. Landing (sourced projections, discounted on the time remaining from 2026.80): US early 2028 = 2028.25 (Artemis IV, 2026-10-09-artemis-iv-early-2028-artemis-v-late-2028; was Sep 2028) x LN(1.8; 1.3-2.8) -> p50 2029.4; China ~2030 (2026-02-11-long-march-10-mengzhou-abort-test) x LN(1.15; 1.0-1.5). Landing -> continuous base LN(7.5 yr; 4.5-13.5) (NASA outpost assembly from 2032, 2026-05-26-nasa-moon-base-phases-outpost-2032, consistent); abandonment US 0.30 / China 0.25. Closure gap measured: water 98% (ISS) and >96% (Tiangong), O2 100% regenerated (Tiangong), ~50% (ISS); food 0% grown on stations; EDEN ISS ~187 kcal/d vs 10,800 for 4 crew ~58x; Tiangong wheat through a full cycle (2026-09-21-tiangong-14-plant-species, an experiment); Lunar Palace closed a year on the ground. Pursuit of food closure US 0.40 / China 0.50 (disclosed judgments, moved from 0.35/0.45 by new evidence: Tiangong crops and a planned lunar vegetable garden, CMSA's in-orbit bioregenerative goal 2026-03-12-cmsa-tiangong-rlss-water-self-sufficiency-goal, the NASA/CSA/DLR Lunar Agriculture Module demonstrator 2026-08-01-nasa-csa-dlr-lunar-agriculture-module, net of the HALO ECLSS termination and an ISRU-cost study; a funded requirement would be 0.5/0.6). Closure LN(10 yr; 5-20) + 1-yr literal floor. Later programme if neither closes, P 0.5 (class: 0 of 2 resolved national crewed Mars programmes -> 1/4 per attempt; ~2.4 attempts 2045-70): U(2045,2070) + LN(12; 7-20) + closure + 1. MC: p_achieved 0.677; p10 2044, p50 2069, p90 null; if achieved 2043/2052/2085 (bimodal; the median lies between the modes); P by end 2049 0.273, 2060 0.448, 2070 0.508, 2080 0.577. Failure 0.323: S-a lunar bases on Earth resupply, nobody pursues food closure, no later programme 0.168 = 0.89 x [(1-0.70x0.40)(1-0.75x0.50) - 0.30x0.25] x 0.5; counted outcome class 0 of 5 long-duration stations (Salyut 6, Salyut 7, Mir, ISS, Tiangong) produced a meaningful share of crew food; no pursuit class is countable (no surface base has existed), so pursuit is a disclosed judgment; station rate transferred (1/7) -> US 2081, 0.624; both 2098, 0.510. Common mode 0.110 = H1 0.04 + H2 0.03 + H3 0.01 + H4 0.03 (each < 10%); no base ever and no later programme 0.034; closure after 2100 0.012. Sensitivities (same draws): pursuit kept 0.35/0.45 -> 2075, 0.653; 0.5/0.6 -> 2058, 0.721; 0.2/0.3 -> 2087; later P 0.25 -> 2072, 0.582; abandonment 0.67/0.67 -> 2091, 0.548; closure 5/15 yr -> 2063/2076; common 0.06/0.20 -> 2064/2078; US discount 2.5 (OIG suits ~2031, 2026-04-20-oig-spacesuit-delays-artemis) -> 2069. Proponent: Musk's self-growing lunar city in <10 yr (2026-02-08-musk-spacex-moon-city-priority) x1.5-3 -> 2040-2054, matching the conditional lunar mode. Lift: shared driver launch-mass (used_by rockets/the-deluge only) is not binding for a base; its draws are not sampled. Anchor (read after the pre-anchor p50 2069 was written): manifold-infkbfl3unm9gnq9qgky-p (Manifold InfkbfL3uNM9GNq9qgKy, similar, 53 forecasters, 'relatively self-sustaining extraterrestrial colony before 2050') 0.24 on 2026-10-15 vs ours P(by end 2049) 0.273: no gap (threshold 0.35). Its wording on the same draws: no completed year -> 0.300; 'relatively' self-sustaining -> 0.434; any location -> earlier; a sustained 'colony' -> later. Loose criteria and a 2049 horizon weigh it lightly; forecast not moved.
- Change
- Prev (2026-10-18-milestone-the-homestead, assess:20261018T191702Z, same made_at) reproduced byte-identically: its eta.py, run unchanged on the same ledger, gives the same output (0.677, 2044/2069/null). No new evidence: the ledger's rockets events and observations are unchanged since prev's run (last data merge f980611, which prev used), the launch-mass driver draws are identical (md5 5a3042a8...); the anchor manifold-infkbfl3unm9gnq9qgky-p has no point after 2026-10-15 (0.24). Prev's inputs were re-checked against the event claims and the rubric and no error was found, so every input is kept (rule 7) and nothing moved; status red unchanged (rule 1: no newer evidence).
- Rubric rule
- Milestone red: median ETA 2069 > made_at + 10 yr (RUBRIC.md §4: Distant).
by assess:20261018T211044Z · 40 evidence event(s)Evidence and news: 40 cited by the assessment, 21 other event(s)
Cited evidence (40)
- NASA’s Human Landing System paper for IAC 2026 (5–9 October) says the agency is working toward the Artemis IV crewed lunar landing in early 2028 and a second crewed landing, Artemis V, in late 2028; after Artemis III, SpaceX and Blue Origin are to demonstrate uncrewed precision lunar landings, with autonomous landing accuracy within 100 m of a target, and providers to be assigned to the crewed missions later.
- NASA named five winners (113 submissions from 33 countries) of its Deep Space Food Challenge: Mars to Table on 24 September 2026, awarding $650,000 in total for paper concepts of food systems for a 15-person crew over 500 sols on Mars; the $300,000 first prize went to the ANI concept, which combines crops, fermentation, fungi and bioreactor nutrient recycling to produce 50% of the food away from Earth.
- The Astronaut Center of China said on 21 September 2026 that more than 20 varieties of 14 plant species have been cultivated aboard the Tiangong space station, with wheat completing a full growing cycle and a second crop from those seeds close to harvest, as groundwork for station crop production and a planned "lunar vegetable garden".The HomesteadClosed-Loop Life Support [1] China cultivates 14 plant species aboard space station (Xinhua)
- At China's first Space Medical Engineering Conference (19–20 September 2026), attendees were told that Tiangong's regenerative life support system had operated in orbit for more than five years with 100% oxygen closure and over 96% water closure.Closed-Loop Life Support [1] 望宇登月服研制进展顺利 (People's Daily, 22 Sep 2026)
- NASA's Moon Base program manager said on 15 September 2026 that only one of the four lunar landers planned for 2026 (Astrobotic's Griffin, no earlier than November) is still expected to launch this year: Blue Origin's Blue Moon Mark 1 Endurance (delayed by a May New Glenn pad explosion) and Intuitive Machines' Nova-C Trinity slipped to the first part of 2027 and Firefly's Blue Ghost 2 to later in 2027.
- Elvis and McDowell (Frontiers in Space Technologies, 14 September 2026) estimate that even a generous 1 billion tons of polar lunar water would sustain a city of one million people for only about a century with ISS-level 98% water recycling (about 2.4 years with none), that current estimates are roughly 30 times lower, and that power is not the binding constraint, precluding large (>100,000-person) long-term lunar settlement unless recycling improves about tenfold or more water is found.
- NASA, the Canadian Space Agency and DLR signed a memorandum of understanding in August 2026 to start preparatory work on a Lunar Agriculture Module ground test demonstrator, linking a high-density crop-growing module to a habitat simulator to study crop growth at reduced pressure and oxygen, CO2 and water exchange, with a decision on building it due by 2032 at the latest.
- China's Chang'e-7 lunar south pole resource-survey mission, set to launch on a Long March 5 around 24 August 2026, did not meet launch conditions and cannot launch in this year's planned window, the China Manned Space Engineering website announced on 23 August 2026, delaying it to 2027 at the earliest.
- A NASA Kennedy paper for ICES 2026 (12-16 July) reports that the late-2024 vacuum test of Lunar Resources' LR-1 molten regolith electrolysis reactor processed 25 kg of lunar regolith simulant and produced oxygen at an average 0.07 kg per hour over 9 hours of sustained electrolysis (a ground demonstration in a vacuum chamber).
- The PRIME-1 team's operations summary from the IM-2 mission (Planetary Science Journal, 1 July 2026) reports that the TRIDENT drill and MSolo mass spectrometer both functioned on the Moon but that PRIME-1 could not fulfil its primary objectives because, with the lander on its side, TRIDENT lay parallel to the surface and could not excavate regolith.ISRU at Scale [1] The Polar Resources Ice Mining Experiment-1 (PRIME-1) Payload Operations Summary from the IM-2 Mission (PSJ 7:7, 2026)
- NASA awarded nearly $600 million on 30 June 2026 for four more Moon Base lunar landings in late 2028 (Astrobotic $297.9 million for two, Firefly $144.2 million and Intuitive Machines $148.3 million for one each), bringing its contracted CLPS surface deliveries to 17.
- A lifecycle-cost study (Johnson et al., Springer, June 2026) estimates that, against Earth resupply at an assumed $100,000/kg lunar delivery cost, lunar-derived oxygen is worth about $286,000 per kg and water about $121,000 per kg to early NASA Moon Base habitation elements with open-loop life support, though it finds ISRU lifecycle costs likely exceed Earth resupply in all but the maximum habitation-demand scenario.
- After NASA halted the lunar Gateway, Northrop Grumman terminated Paragon Space Development's contract to design and deliver the life support system for Gateway's HALO habitation module, a $100-million contract awarded in 2022, and Paragon cut about 70–77 jobs at its Tucson headquarters.Closed-Loop Life Support [1] Paragon Space cuts 77 jobs in Tucson after NASA contract ends (Phoenix Business Journal, 16 Jun 2026) [2] Paragon lays off staff after Gateway cancellation (SpaceNews, 13 Aug 2026) [3] NASA lunar program shutdown triggers layoffs at space contractor (AL.com, 13 Aug 2026)
- NASA's Moon Base roadmap (26 May 2026) schedules Phase One (build, test, learn) through 2029, Phase Two (early power, cargo, logistics and communications infrastructure) in 2029-2032, and Phase Three, assembly of a permanent South Pole outpost where astronauts live and work for extended periods, from 2032 onward.The Homestead [1] Moon Base Phases (NASA)
- SpaceX's IPO registration statement (S-1, filed 20 May 2026) states a goal of a sustainable lunar presence, saying a lunar economy first requires proving reliable extraction of water ice to sustain life and producing hydrogen-oxygen propellant, and that locally produced water and fuel would provide the foundation for a permanent human presence on the Moon, without giving a timeline.
- NASA's Divergent Deployable Wastewater Treatment Facility, a trailer with three bioreactors that turn separated urine, graywater, fecal and food waste into reusable water and hydroponic nutrient solution for a vertical garden, was connected to the University of North Dakota's Integrated Lunar/Martian Analog Habitat on 18 May 2026 for integrated testing with human subjects planned from summer 2026 through September 2028.Closed-Loop Life Support [1] NASA TechPort: Bioregenerative Life Support Systems (BLiSS) [2] At UND, a NASA trailer helping scientists learn how to live on other worlds (UND Today, June 2026) [3] NASA Testing Wastewater Treatment Facility for Future Moon Base (NASA article reposted by Waypoint 2 Space, 2 Jun 2026)
- NASA's CRS-34 mission overview (8 May 2026) said the ISS's automated Advanced Plant Habitat plant-growth facility would return to Earth on the June 2026 Dragon splashdown for eventual museum display; Ohalo III, a crop-production testbed housed in an EXPRESS rack, was targeted to launch to the ISS no earlier than August 2026 and is described as the first operational crop production system in space and a prototype for a Mars Transit Vehicle.
- A May 2026 study of Chandrayaan-2 dual-frequency radar observations found signatures consistent with subsurface ice in four doubly shadowed craters in the lunar south polar region, with the strongest radar evidence in crater F2 on the floor of Faustini.
- NASA's 2026 status paper on the ISS as an exploration life-support testbed (ICES-2026-379, released 29 April 2026) reports that the Brine Processor Assembly had recovered about 1,395 L of water over 1,079 days of operation by December 2025 (about 1.3 L/day), that no CO2-reduction (Sabatier) system is planned for the ISS after the Sabatier upgrade was cancelled (a companion NASA paper puts ISS oxygen-loop closure at about 50%), and that the upgraded oxygen generator is due in mid-2027 and the full exploration Air String in mid-2028.
- NASA's Office of Inspector General reported on 20 April 2026 that Axiom's lunar and microgravity spacesuits are at least a year and a half behind schedule, with demonstration readiness planned for late 2027 but, if testing follows historical averages, not until 2031, which could affect the Artemis landing timeline.
- Artemis II, NASA's first crewed lunar flyby, ended with a Pacific splashdown at 8:07 p.m. EDT on 10 April 2026 after a mission of 9 days 1 hour 32 minutes in which the four-person crew reached 252,756 miles from Earth, a record distance for human spaceflight.
- The President's FY2027 budget request (April 2026) would cut NASA to $18.8 billion while putting $8.5 billion into Artemis and Mars-focused capabilities, starting work toward a permanent lunar base camp near the lunar South Pole, and requesting no discretionary funding for Gateway, whose reconciliation-act money ($1.1 billion in FY2026 and $750 million in each of FY2027 and FY2028) would be transitioned to the lunar base camp.
- NASA's SLS launched the Artemis II Orion spacecraft with astronauts Reid Wiseman, Victor Glover, Christina Koch and CSA's Jeremy Hansen from Kennedy Space Center at 6:35 p.m. EDT on 1 April 2026 on NASA's first crewed lunar flyby in 50 years.
- At its 24 March 2026 "Ignition" event NASA announced a three-phase plan to build a lunar base (robotic CLPS and rover demonstrations, then semi-habitable infrastructure and recurring surface operations, then heavier infrastructure for a continuous human foothold and permanent base) and said it intends to pause the Gateway lunar station in its current form to focus on sustained surface operations.
- A University of Hawai'i-led ShadowCam study (Science Advances, 18 March 2026) found no evidence of widespread water ice at concentrations above 20-30% by weight on the surfaces of the Moon's permanently shadowed craters, identifying only a few areas 20-50 m across with optical signatures consistent with more than 10% ice (buried or lower-concentration ice not ruled out).
- China Manned Space said on 12 March 2026 that Tiangong's physico-chemical regenerative life support system has supported 10 crews for nearly 1,700 days (Shenzhou-12 to Shenzhou-21), that its design water closure is about 80% from condensate and urine recovery and over 90% with CO2 reduction, and that it plans membrane distillation to raise water recovery from urine to over 90% with the goal of 100% on-orbit water self-sufficiency, and that bioregenerative (plant, algae and biological waste-treatment) life support should be verified in orbit in due course.Closed-Loop Life Support [1] 空间站再生生保系统：把"生态循环"搬进浩瀚太空 (China Manned Space article republished by CCTV, 12 Mar 2026)
- The U.S. Senate Commerce Committee unanimously passed the NASA Authorization Act of 2026 on 4 March 2026, which would authorize $24.7 billion for FY2026 and $25.3 billion for FY2027, extend NASA's authority to operate the ISS from 2030 to 30 September 2032, require a continuous U.S. human presence in Earth orbit, and for the first time authorize a permanent Moon base (committee passage, not yet law).
- NASA restructured Artemis on 27 February 2026: Artemis III, now in 2027, becomes a low-Earth-orbit test of rendezvous and docking with one or both commercial landers (SpaceX, Blue Origin), the first crewed landing moves to Artemis IV in 2028, and NASA plans at least one lunar surface landing every year thereafter with a standardized SLS configuration.
- China flew a prototype Long March-10 crewed lunar rocket for the first time on 11 February 2026 in a low-altitude demonstration flight that also served as the first maximum-dynamic-pressure abort test of the Mengzhou crew spacecraft, with the first stage and return capsule splashing down at sea for recovery, CMSA said, as China aims to land astronauts on the Moon before 2030.
- Elon Musk wrote on 8 February 2026 that SpaceX has shifted its focus to building a "self-growing city" on the Moon, which he said it could potentially achieve in less than 10 years versus 20+ years for Mars, while SpaceX would also strive to begin building a Mars city in about 5 to 7 years (projections, not plans with dates).
- Congress rejected the administration's proposed cut of NASA's FY2026 budget to $18.8 billion, funding the agency at $24.438 billion (versus $24.838 billion in FY2025) in the Commerce-Justice-Science appropriations act (H.R. 6938) that President Trump signed into law on 23 January 2026.
- The US Department of Energy and NASA announced on 13 January 2026 a memorandum of understanding to develop a lunar surface fission reactor by 2030, a planned power system meant to operate for years without refuelling to enable sustained lunar missions (a target, not hardware).
- ISS food remains 100% resupply-dependent (0% produced on board), with plant growth limited to Veggie and Advanced Plant Habitat experiments.Closed-Loop Life Support [1] NASA: Astronauts plant seed pillows in space agriculture study
- China’s ILRS plan projects a basic model in the lunar south pole region by 2035.↳ corrects (or adds sources to) an earlier record:
2025: China's International Lunar Research Station (ILRS) plan targets a basic lunar base by… - An npj Microgravity review says China's Beijing Lunar Palace program demonstrated closed-system operations for a breathable atmosphere, water and nutritious food for a crew of four taikonauts for an entire year.Closed-Loop Life Support [1] npj Microgravity 2025: Lunar Palace 1 [2] Wayback capture 20260106124333↳ corrects (or adds sources to) an earlier record:
2025: China's Lunar Palace 1 bioregenerative life-support facility demonstrated 80% food… - China targets its first crewed lunar landing around 2030 (Reuters, April 2025).The Homestead [1] Reuters: China, Russia may build nuclear plant on moon to power lunar station, official says [2] Wayback capture 20251228183757↳ corrects (or adds sources to) an earlier record:
2025-04-23: China targets its first crewed lunar landing around 2030 (Reuters, April 2025). - NASA’s MOXIE completed its Mars mission in 2023 after 16 oxygen-production runs, reaching up to 12 g per hour at 98% purity or better.ISRU at Scale [1] NASA: MOXIE completes Mars mission↳ corrects (or adds sources to) an earlier record:
2024-09: NASA's MOXIE experiment on the Perseverance rover completed its mission (Sep 2024 per… - In June 2023 NASA achieved 98% water recovery on the ISS using the Brine Processor Assembly together with the Urine Processor and Water Processor assemblies, meeting the threshold for Mars missions.Closed-Loop Life Support [1] NASA: Water recovery milestone on ISS↳ corrects (or adds sources to) an earlier record:
2023-06: In June 2023 NASA achieved 98% water recovery on the ISS using the Brine Processor… - China’s space station can produce 100% of its oxygen supply through its onboard regeneration system, SCIO reported in April 2023.Closed-Loop Life Support [1] SCIO: China's space station achieves 100% oxygen regeneration↳ corrects (or adds sources to) an earlier record:
2023-04: China's Tiangong space station has achieved 100% oxygen regeneration and more than 95%… - EDEN ISS produced 268 kg of fresh edible biomass during its 286-day operational phase in 2018.Closed-Loop Life Support [1] Frontiers in Plant Science 2020: EDEN ISS↳ corrects (or adds sources to) an earlier record:
2020: The EDEN ISS greenhouse in Antarctica produced 268 kg of fresh produce from 12.5 m² of…
Other events (21)
- With Albania, Croatia, Côte d’Ivoire and San Marino signing, the Artemis Accords reached 76 signatory countries by 28 September 2026.
- NASA Glenn released a draft RFP (late August 2026) for the multiple-award Space Power and Reactor Capabilities (SPARC) contract whose main purpose is maturation, testing and flight design of Lunar Reactor-1, the first operational lunar surface fission reactor, with the final RFP planned around 15 November 2026 and ordering from April 2027 through 2031.The Homestead [1] NASA Seeks Comments on SPARC Draft RFP (GovCon Wire)
- NASA named the four-person Artemis III crew on 9 June 2026 and detailed the 2027 Earth-orbit mission, in which Orion will rendezvous and dock with test versions of one or both commercial landers (Blue Origin's pathfinder spending about two days docked), ahead of Artemis IV, the first planned crewed mission to the lunar South Pole, in 2028.
- NASA awarded Astrolab $219 million and Lunar Outpost $220 million for the first phase of Lunar Terrain Vehicles, to deploy crewed and uncrewed rovers to the lunar South Pole region by 2028, and Blue Origin $188 million (with a $280.4 million option) to deliver them.
- NASA named the first three Moon Base missions on 26 May 2026, all targeted for launch in 2026: Moon Base I on Blue Origin's Blue Moon Mark 1 Endurance lander (no earlier than fall 2026, to Shackleton Connecting Ridge), Moon Base II on Astrobotic's Griffin with Astrolab's FLIP rover, and Moon Base III on Intuitive Machines' Nova-C Trinity.
- +16 older event(s): the section's event log
Major Open Challenges
Radiation Shielding: Protecting crew from cosmic and solar radiation during months-long deep space transits.9 event(s)
Events (9)
- A Science Advances paper (12 August 2026) on the DLR-led MARE experiment, which flew two female phantoms on Artemis I with one wearing StemRad's ~26 kg HDPE AstroRad vest, estimates from extrapolated inner-belt measurements that the vest would cut effective dose by about 60% (58.2%) in an August 1972–like solar particle event and by nearly 40% in an October 1989–like event, sparing up to 193 and 131 days of deep-space radiation exposure, respectively.Radiation Shielding [1] Science Advances: First evaluation of wearable radiation protection for human deep space exploration, as flown on Artemis I (12 Aug 2026) [2] DLR: AstroRad vest can significantly reduce radiation exposure for humans in space (13 Aug 2026) [3] ESA Orion blog: Artemis I radiation results to protect astronauts (28 Aug 2026)
- NASA's Space Radiation Analysis Group reported at the 30 April 2026 Space Weather Workshop that the Artemis II crew active dosimeters (the dose of record) averaged about 10 mGy, with 80% from trapped Van Allen belt radiation and 20% from galactic cosmic rays; a small proton enhancement and an impulsive M1 flare on 4 April had no impact.
- On Artemis II's eighth flight day (8 April 2026), NASA opted to forgo the originally planned deployment demonstration of Orion's radiation shelter to prepare the cabin for entry; a shortened demonstration tested ventilation for an astronaut sheltering in a locker, while other elements were omitted.
- South Korea's K-RadCube, a 12U CubeSat deployed from Artemis II to measure radiation and its biological effects in and beyond the Van Allen belts with a tissue-equivalent dosimeter, returned only weak and abnormal signals and no observation data, KASA and KASI said on 3 April 2026.
- GSI/FAIR in Darmstadt, with ESA, put into operation Europe's first galactic cosmic ray simulator (a hybrid active-passive iron-ion beam method, up to 1 GeV/nucleon), making it the world's second GCR simulator after NASA's at Brookhaven, with results published in two Life Sciences in Space Research papers.
- +4 older event(s): the section's event log
Closed-Loop Life Support: Years-long operation without resupply from Earth.19 event(s)
Events (19)
- NASA named five winners (113 submissions from 33 countries) of its Deep Space Food Challenge: Mars to Table on 24 September 2026, awarding $650,000 in total for paper concepts of food systems for a 15-person crew over 500 sols on Mars; the $300,000 first prize went to the ANI concept, which combines crops, fermentation, fungi and bioreactor nutrient recycling to produce 50% of the food away from Earth.
- The Astronaut Center of China said on 21 September 2026 that more than 20 varieties of 14 plant species have been cultivated aboard the Tiangong space station, with wheat completing a full growing cycle and a second crop from those seeds close to harvest, as groundwork for station crop production and a planned "lunar vegetable garden".The HomesteadClosed-Loop Life Support [1] China cultivates 14 plant species aboard space station (Xinhua)
- At China's first Space Medical Engineering Conference (19–20 September 2026), attendees were told that Tiangong's regenerative life support system had operated in orbit for more than five years with 100% oxygen closure and over 96% water closure.Closed-Loop Life Support [1] 望宇登月服研制进展顺利 (People's Daily, 22 Sep 2026)
- NASA, the Canadian Space Agency and DLR signed a memorandum of understanding in August 2026 to start preparatory work on a Lunar Agriculture Module ground test demonstrator, linking a high-density crop-growing module to a habitat simulator to study crop growth at reduced pressure and oxygen, CO2 and water exchange, with a decision on building it due by 2032 at the latest.
- NASA concluded its $3 million LunaRecycle Challenge on 28 August 2026, naming MIT's CERBERUZ (which grinds mixed trash, including foam, into feedstock for injection-moulded parts or 3D-printing filament; $775,000 in total awards) first among prototype and digital-twin systems for recycling fabrics, plastics, foam and metals on lunar or deep-space missions.Closed-Loop Life Support [1] NASA Awards First Prize in Phase 2 of Agency's LunaRecycle Challenge (28 Aug 2026)
- +14 older event(s): the section's event log
ISRU at Scale: Production-level extraction of fuel, oxygen, and materials from planetary resources.24 event(s)
Events (24)
- NASA selected three PRISM science payload suites for CLPS delivery under its Moon Base program on 30 September 2026, including DISCO (Ames, PI Ariel Deutsch), which is to make the first direct on-surface measurements of ice in lunar micro-cold traps, pinpointing where ice exists and how much to inform turning ice into a usable resource for Moon Base (no flight date given).
- The UN COPUOS Legal Subcommittee's Working Group on Legal Aspects of Space Resource Activities circulated, with its 24 September 2026 intersessional meeting, a consolidated draft of non-binding Initial Recommended Principles for Space Resource Activities stating that extraction of space resources in compliance with the Outer Space Treaty 'does not inherently constitute national appropriation', for negotiation toward the April 2027 Legal Subcommittee session.
- Elvis and McDowell (Frontiers in Space Technologies, 14 September 2026) estimate that even a generous 1 billion tons of polar lunar water would sustain a city of one million people for only about a century with ISS-level 98% water recycling (about 2.4 years with none), that current estimates are roughly 30 times lower, and that power is not the binding constraint, precluding large (>100,000-person) long-term lunar settlement unless recycling improves about tenfold or more water is found.
- China's Chang'e-7 lunar south pole resource-survey mission, set to launch on a Long March 5 around 24 August 2026, did not meet launch conditions and cannot launch in this year's planned window, the China Manned Space Engineering website announced on 23 August 2026, delaying it to 2027 at the earliest.
- A Planetary Science Journal study from NASA's Mars Subsurface Water Ice Mapping (SWIM) team (published 31 July 2026) jointly interprets orbital data to estimate the likelihood of subsurface water ice between 60°S and 60°N, finding near-surface ice likely poleward of about 45° in both hemispheres but unconstrained by existing data closer to the equator.
- +19 older event(s): the section's event log
Beyond the framework: 54 event(s)
- Rocket Lab said on 14 October 2026 that its medium-lift reusable Neutron will not fly until 2027 after voids were found in carbon-composite propellant-tank domes of the first stage, traced to a design change made after a tank ruptured in a hydrostatic test in January; Neutron was announced in 2021 for a 2024 first flight.
- The U.S. DOT announced five proposed FAA rules to streamline Part 450 commercial launch and reentry licensing (electronic applications, more flexibility to show equivalent safety, combining three safety analyses into one submission), citing 217 of 329 commercial launches worldwide in 2025 by the U.S. and an expectation of 10,000 FAA-licensed launches and reentries a year by 2035.
- NASA said on 28 September 2026 that Boeing's uncrewed Starliner-1 engineering-evaluation mission could fly in December 2026 or January 2027 and that it plans to return astronauts on Starliner-2 by 2028.Beyond the framework [1] NASA, Boeing Share Update on Commercial Starliner Development Plans (NASA)
- With Albania, Croatia, Côte d’Ivoire and San Marino signing, the Artemis Accords reached 76 signatory countries by 28 September 2026.
- ESA's Space Environment Report 2026 (data to end-2025) says more than 300 launches put over 4,000 payloads in orbit in 2025, more than three intact satellites or rocket bodies now reenter daily, and projected LEO object and collision numbers are rising in a runaway (Kessler-type) effect even with the projection horizon cut from 200 to 100 years, making active debris removal necessary.Beyond the framework [1] ESA Space Environment Report 2026 (ESA)
- +49 older event(s): the section's event log
Latest news: the 8 most recently published of 189 events
- Rocket Lab said on 14 October 2026 that its medium-lift reusable Neutron will not fly until 2027 after voids were found in carbon-composite propellant-tank domes of the first stage, traced to a design change made after a tank ruptured in a hydrostatic test in January; Neutron was announced in 2021 for a 2024 first flight.
- On 14 October 2026 the FAA announced its intent to prepare a Programmatic EIS on SpaceX’s proposal to raise Starship operations at Starbase from the currently authorized 25 launches and 50 landings per year to up to 100 launches and 200 landings a year in Phase 1 and up to 456 launches and 912 landings a year in Phase 2.
- NASA's Human Landing System presentation at IAC 2026 (5-9 October) schedules SpaceX's Starship ship-to-ship propellant transfer demo for Q1 2027, after a first ship catch and first Cape flight in Q4 2026, whereas SpaceX's October 2025 update had targeted the in-space propellant transfer flight test for 2026.
- NASA released its final Request for Proposals for commercial low Earth orbit space stations to follow the ISS on 9 October 2026, planning multi-award firm-fixed-price IDIQ contracts with proposals due 8 December 2026 and awards expected in spring 2027.
- NASA’s Human Landing System paper for IAC 2026 (5–9 October) says the agency is working toward the Artemis IV crewed lunar landing in early 2028 and a second crewed landing, Artemis V, in late 2028; after Artemis III, SpaceX and Blue Origin are to demonstrate uncrewed precision lunar landings, with autonomous landing accuracy within 100 m of a target, and providers to be assigned to the crewed missions later.
- The U.S. DOT announced five proposed FAA rules to streamline Part 450 commercial launch and reentry licensing (electronic applications, more flexibility to show equivalent safety, combining three safety analyses into one submission), citing 217 of 329 commercial launches worldwide in 2025 by the U.S. and an expectation of 10,000 FAA-licensed launches and reentries a year by 2035.
- Varda launched its W-8 and W-9 reentry vehicles together on SpaceX's Transporter-18 on 1 October 2026, the first time two commercial reentry vehicles went to orbit at the same time; they carry thermal-protection payloads and, on W-9, a navigation system for hypersonic flight, for commercial and government customers.
- Varda raised a $251 million Series D (total raised $598 million) on 30 September 2026 to scale toward the first pharmaceutical made in space for patients on Earth, having completed six successful reentry missions with over a dozen more launches and reentries planned through 2028.
Fusion and Energy
237 events · latest news published 2026-10-16 (6 days ago)news per week
Energy is the key limiting factor of the material world. It does not have to be. If all electricity bills go to zero, how does that change our lives, from heating (homes, iron or saltwater) to food, to travel, to compute? There are many short-term fixes, but the long-term approach is unlikely to be anything but nuclear fusion (and superconductors make everything easier).
KPI
Behind pace: $113/MWh (2026-Q1), falling $2.90/MWh a year vs its own long-run pace of $3.51 · assessed 2026-10-18
The worldwide household price, in 2025 dollars, averaged $111/MWh from late 2023 to early 2026, down 6% from $119 over the previous two and a half years: a fall of about $2.90/MWh a year. That is clearly downward but slower than the $3.51/MWh a year it has fallen on average over its longer record, so it is behind pace. Caveat: the measure runs only to early 2026 and has been edging up over its most recent stretch, and more recent reports (record Japanese bills, rising retail prices in most EU markets, US electricity prices outpacing inflation) point to a further rise it has not yet captured.
The KPI's precise definition
The KPI, as defined: Worldwide average household (retail) electricity price ($/MWh): what people actually pay, population- or consumption-weighted where available, inflation-adjusted (real $)
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- elec-price-household-world-popw-real: worldwide population-weighted household retail electricity price, constant 2025 US$/MWh; goal_direction down. Framework KPI includes distribution, taxes and fees, not generation LCOE.
- Window
- Span
- 2021-Q2..2026-Q1, five calendar years; 19 observations, 2024-Q3 missing; equal time halves, not equal observation counts
- Current
- Span
- 2023-Q4..2026-Q1
- N
- 9
- Observations
- 2023-Q41112024-Q1108.52024-Q2105.12024-Q4113.22025-Q1109.72025-Q2113.12025-Q3112.82025-Q4114.42026-Q1112.9
- Mean
- 111.2
- SD
- 2.96
- Mean time
- 2025
- OLS
- Slope
- 2.478
- SE
- 1.061
- Previous
- Span
- 2021-Q2..2023-Q3
- N
- 10
- Observations
- 2021-Q21242021-Q3124.12021-Q4122.12022-Q1112.92022-Q2119.12022-Q3122.62022-Q4125.22023-Q1111.52023-Q2111.62023-Q3112.6
- Mean
- 118.6
- SD
- 5.767
- Mean time
- 2022
- OLS
- Slope
- -4.914
- SE
- 2.059
- Whole window OLS
- Slope
- -2.498
- SE
- 0.7339
- Mean difference
- -7.381
- Percent change
- -6.225
- Half mean time separation years
- 2.542
- Half mean pace per year
- -2.904
- Naive SE difference
- 2.074
- Noise note
- Half-mean fall 7.381 exceeds earlier-half SD 5.767; quarterly autocorrelation makes the naive SE descriptive, not a significance test. Benchmark shortfall is modest; recent-half OLS +2.478/yr is a watch item, not a whole-window reversal.
- Long run benchmark
- Span
- 2018-Q2..2026-Q1
- N
- 29
- Observations
- 2018-Q21322019-Q1134.62019-Q2136.82019-Q3129.52019-Q4129.72020-Q1128.42020-Q2130.42020-Q3127.32020-Q4128.12021-Q1121.92021-Q21242021-Q3124.12021-Q4122.12022-Q1112.92022-Q2119.12022-Q3122.62022-Q4125.22023-Q1111.52023-Q2111.62023-Q3112.62023-Q41112024-Q1108.52024-Q2105.12024-Q4113.22025-Q1109.72025-Q2113.12025-Q3112.82025-Q4114.42026-Q1112.9
- Mean
- 120.2
- SD
- 8.89
- Mean time
- 2022
- OLS
- Slope
- -3.507
- SE
- 0.3626
- Source
- uv run core/ledger.py series fusion elec-price-household-world-popw-real --as-of 2026-10-18; latest 2026-Q1 = 112.9, collector, published 2026-10-09; quarter midpoint time = year+(quarter-0.5)/4; same 29 observations as prev
- Benchmark
- Tier 3: own long-run OLS 2018-Q2..2026-Q1, 29 observations over 7.75 years, -3.5074077505 ±0.3625733515 $/MWh/year; current halves -2.9040437158, current-window OLS -2.4980145278 ±0.7339107168. No external required or milestone-implied household-price path exists.
- Benchmark source
- Fixed KPI specification: own series, tier 3; elec-price-household-world-popw-real (GlobalPetrolPrices country prices, World Bank population weights, CPI deflation).
- Data as of
- 2026-Q1
- Change
- Prev 2026-10-16-kpi (v6, yellow). No new observation of elec-price-household-world-popw-real (latest 2026-Q1 = 112.9, published 2026-10-09); window statistics reproduce prev exactly (halves 118.57 / 111.19, pace -2.904, OLS -2.498, benchmark -3.507). Data age 201 days from quarter end, below two quarterly cadences plus the ~192-day lag: not stale. Newly ingested context events (Eurostat EU H1-2026 €28.32/100 kWh, down from €28.96; ACER-CEER upward retail trend in 18 EU markets Jan-Aug 2026; Japan record household bills for Oct 2026; US CPI electricity +4.4% y/y vs +3.7% all items; GPP unweighted nominal average 174/176/179 for 2026-Q1..Q3) hint at a 2026 rise; they are other quantities, not imputed, and 2026-Q2 onward is unmeasured, not flat. Label: dropped "(tier 3)" per RUBRIC §6 (tier is in basis.benchmark). Yellow retained.
- Rubric rule
- Rubric §3: five-year half means clearly decline (-7.38 $/MWh, beyond the earlier-half SD 5.77), but the half-mean pace -2.90 and whole-window OLS -2.50 $/MWh/yr are below the tier-3 own long-run pace -3.51; yellow. §5: status unchanged.
by assess:20261018T192708Z · 7 evidence event(s)
Evidence and news: 7 cited by the assessment, 19 other KPI news
Cited evidence (7)
- Eurostat's household electricity price dataset (nrg_pc_204, updated October 16, 2026) puts the EU average household electricity price (incl. all taxes and levies, 2,500–4,999 kWh/yr band) at €28.32 per 100 kWh in H1 2026, down from €28.96 in H2 2025 and €28.79 in H1 2025 (excluding taxes and levies: €20.27, from €20.59 in H2 2025).
- The US BLS Consumer Price Index release for September 2026 (October 14, 2026) shows the CPI-U electricity index up 4.4% over the 12 months to September 2026 (not seasonally adjusted), versus a 3.7% rise in the all-items index.
- The ACER-CEER 2026 retail energy markets monitoring report (September 30, 2026) finds the EU-27 average household electricity price (ACER calculations on Eurostat data, all consumption bands) was EUR 294/MWh in 2025 versus EUR 224/MWh over 2019–2021, still above pre-crisis levels in both nominal and inflation-adjusted terms, that household electricity spending ranged from 1.2% (Luxembourg) to 3.9% (Czechia, Greece, Sweden) of national median income in 2025, and that more recent data show an emerging upward trend in retail electricity prices in 18 markets between January and August 2026.
- Japan's 10 major electric utilities announced on September 29, 2026 that their standard-household electricity bills for October 2026 usage (billed in November) will hit record highs at all 10 companies, rising by ¥835–1,578 a month after the end of the government's ¥3.50/kWh subsidy (plus a November 1 network-tariff revision), with Tokyo Electric Power's standard household at ¥9,561 a month (+¥1,286) and Okinawa Electric's at ¥10,752 (nominal yen).
- IEA's Electricity Mid-Year Update 2026 puts the average EU wholesale electricity price at around USD 105/MWh in H1 2026 (+15% year on year) amid the Strait of Hormuz LNG shock, and the US average at about USD 52/MWh (+10%).KPI [1] IEA Electricity Mid-Year Update 2026 (PDF), Prices chapter [2] IEA Electricity Mid-Year Update 2026 (report page, published 23 July 2026)↳ updates 2025: In H1 2025, EU wholesale electricity prices averaged around USD 90/MWh, about 30%… (corrected version)↳ updates 2025: In H1 2025, US wholesale electricity prices averaged around USD 48/MWh, about 40%… (corrected version)
- Berkeley Lab's Retail Electricity Price Trends and Drivers 2026 edition finds US national-average residential electricity prices rose 33% in nominal terms and 5.6% in inflation-adjusted terms from 2019 to 2025, with residential prices up 5.0% (nominal) in 2025 alone.
- IEA's Electricity 2026 report finds that between 2019 and 2024 household electricity prices rose by 36% on average in the European Union and 26% in the United States, outpacing general inflation of 22% and 23% respectively.
Other events (19)
- EIA's September 2026 Short-Term Energy Outlook puts the US average residential electricity price at 17.3 cents/kWh in 2025 and forecasts 18.2 cents/kWh in 2026 and 18.6 cents/kWh in 2027 (nominal).
- IRENA's Renewable Power Generation Costs in 2025 report (July 2026) gives the 2025 global weighted-average LCOE of newly commissioned projects as $44/MWh for solar PV (unchanged from 2024), $33/MWh for onshore wind and $78/MWh for offshore wind (2025 USD), with solar PV down 89%, onshore wind 71% and offshore wind 63% since 2010.
- Eurostat reports the consumption-weighted EU average household electricity price (incl. taxes, 2,500–5,000 kWh/yr band) at €28.96 per 100 kWh in H2 2025, up slightly from €28.79 in H1 2025 and still well above pre-2022-crisis levels.
- IEA's Household Energy Affordability report finds that global household spending on electricity exceeded USD 1 trillion for the first time in 2024, with electricity bills averaging 31% of total household energy spending (up from 26% in 2000).
- IEA's Household Energy Affordability report finds that EU average retail electricity prices reached USD 310/MWh in 2024 (about USD 130/MWh generation costs plus USD 180/MWh network charges and taxes), and that European household electricity prices in 2024 were about 10% higher in real terms than in 2019.
- +14 older event(s): the section's event log
9 more metric(s) we track
Household electricity price, simple country average (GlobalPetrolPrices)
179.0 $/MWh
Q3 2026 · source
▲ +9.0 vs Q3 2025
Milestone Countdown
A sequence: each milestone is expected after the one before it.
Expected year: dot = median (hollow: after 2100), bar = 80% range; arrows: each milestone is expected after the one before. The time axis is compressed: the farther out, the more years per step.
The First Watt: First commercial fusion plant delivers sustained electricity to a public grid.Progressing: no fusion plant on any grid; Helion's Orion licensed, 2028 goal softened, ETA ~2033 · assessed 2026-10-18 · 57 event(s)
- Sustained = net electricity exported under commercial operation for 3+ consecutive months.
No commercial fusion plant exports net electricity to a public grid, and no device has yet produced more electricity than it uses; Helion's Orion plant is licensed and being built, and Commonwealth Fusion Systems has applied to connect its first plant to the grid in the early 2030s. Discounting the companies' schedules for slippage and requiring three months of sustained export gives a 93% chance by 2100 and a median of 2033 (with a 10% chance each of 2029 or earlier and of after 2060), so yellow. The median moved a year later because Helion now says Orion only begins initial operations in 2028, with full power not before 2029-2030. Caveat: a prediction market puts commercial fusion electricity by 2035 at 37%, a looser test on which we are at 60%, and an industry survey's median is about 2037; both point later, and how much to trust Helion's fast schedule is what decides it.- ETA
- ~2033 (80% range 2029–2060; P(by 2100) 93%)
- ETA cruxes
- Helion route weight in driver fusion-first-watt first (.40/.25/.10 -> p50 2035/2037/2038; the 2035 Manifold anchor gap sits here); slip median 1.5-3x (2031-2035); Helion re-dated to 2029.5/2030.5 (2035/2036); Helion ramp 0-x2 (2032-2033); ARC date (2032-2033). Earlier: Polaris net electrical recovery, Orion generator installed and Chelan PUD interconnection signed, SPARC first plasma and Q>1 in 2027. Later: Orion grid connection past 2028, Polaris without the needed gain, SPARC slipping further, FOAK availability too low for 3 consecutive months.
- Path
- Reading (owner definition d964534): a plant in commercial operation exporting net electricity to a public grid for 3+ consecutive months, dated at the end of the third month; a pulse, test sync or PPA does not count. Driver start = start of commercial grid operation; Helion branch adds a ramp to net export (sourced from its full-power statement), all routes the 3-month window. Lead path: Helion Orion (licensed June 2026, $1.5B raised, Microsoft contract requiring grid connection by end-2028); then CFS ARC (PJM interconnection request, $4B raised, DOE-reviewed design, SPARC ~80% but behind its magnet plan), TAE Da Vinci 2031 (going-concern doubt), Type One Infinity Two 2034, General Fusion ~2035.
- Blockers
- Technical: no verified net electrical gain in any device; target gain is not electrical gain; a published critique questions Helion's D-He3 gain basis.
- Schedule: Helion softened its 2028 electricity goal (full power 2029-2030) and is still resolving FRC merging before locking Orion's design; SPARC turn-on slipped to 2027 and its magnet installation is behind plan.
- Reliability: a first-of-a-kind plant must hold 3 consecutive months of commercial net export; no plant availability data exist.
- Regulatory: US NRC fusion-machine rule still proposed (state licences issued in WA and TN); no structural block, no assumed gate.
- Gather: Polaris electrical results, Orion interconnection agreement, SPARC first plasma; driver fusion-first-watt refit (owner).
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- discounted proponent dates with construction floors in a route scenario mixture (shared driver fusion-first-watt), plus a sourced Helion ramp to net export and the definitional 3-month window; reference-class, survey and anchor cross-checks
- ETA calculation
- Gap: 0 vs 1 qualifying plant (net export under commercial operation for 3+ consecutive months); no net electrical gain anywhere (NIF record Q=4.13, 2025-04-07-nif-q-4-13-rv; 11th ignition 7.9 MJ gain ~3.8, 2026-06-20-nif-11th-ignition-7-9mj; Polaris D-T fusion at 150 M degC, no gain result, 2026-02-13-helion-polaris-d-t-fusion-150mc). No gain/plant series in any section. Model = prev 20261015T105338Z (reproduced exactly: P .9324, 2028.78/2032.43/2060.88): draw i of shared driver fusion-first-watt (driver-draws --as-of 2026-10-18, seed 20261015, n 10000; helion .50/arc .35/successor .08/never .07, slip lognormal(ln2,.40), helion start max(2028, asof+(2028-asof)F) from 2025-01-01-helion-2028-microsoft-target-rv, arc max(2030, asof+(2032.5-asof)F) from 2025-01-01-cfs-sparc-arc-timeline-rv) + milestone mapping. Mapping update on new evidence 2026-09-08-helion-orion-2028-goal-softened (2028 = 'begin initial operations'/grid connection; 50 MW full power not before 2029-2030): helion branch T = start + ramp tri(0.25,1.0,2.0) yr + 0.25 (3 months); arc/successor T = start + 0.25. Result: P(by2100) .9324, p10 2029.81, p50 2033.12, p90 2060.88 (calendar 2029/2033/2060); conditional median 2032.43; branch medians helion 2030.58, arc 2038.40. Sensitivities: ramp 0/x2 -> 2032.43/2033.93; helion weight .40/.25/.10 -> 2035.39/2037.63/2038.91, .55 -> 2032.28; slip 1.5x/3x -> 2031.78/2035.78; ARC 2031/2034 -> 2032.39/2033.70; Helion re-dated 2029.5/2030.5 (no ramp) -> 2035.26/2036.87; never .15/.03 -> P .849/.970. Cross-checks: discounted (1.5/2/3x) Helion full power 2030 -> 2031.6/2033.2/2036.4, ARC 2035.4/2038.2/2043.9, TAE Da Vinci 2031 (2026-08-05-tae-da-vinci-operations-2031; going-concern doubt) 2033.1/2035.2/2039.4, Type One 2034 2037.6/2041.2/2048.4; FIA 2026 survey (2026-07-13-fia-2026-global-fusion-industry-report, 45 firms): 42% by 2035, median ~2037.4, later than ours; fission 12-15 yr from precursor -> 2035-40. Pre-anchor p50 2033 (2029-2060, P .93). Anchors: manifold-iubcricrjrizushzxq97-p (engineering breakeven by 2040, weaker, 0.871, 56 forecasters) vs ours P(by 2040-12-31) .739: bound holds. manifold-pr0qngelph-p (commercial fusion electricity by 2035, weaker, 0.368, 11 forecasters) vs ours P(by 2035-12-31) .597: weaker-bound gap .23, explained not closed. Crux: Helion route weight (.50) and slip in driver fusion-first-watt; the market's .37 matches helion weight ~.25 (.411) or a 2030.5 Helion date (.430). Settles it: Polaris net electrical recovery and an Orion interconnection agreement (toward ours) or Orion grid connection re-dated past 2028 (toward the market). Thin market, loose, weaker event: soft bound. No move: shared-driver inputs; driver finding reported (refit warranted on 2026-09-08-helion-orion-2028-goal-softened, 2026-05-08-helion-tiny-merge-testbed, 2026-09-22-cfs-sparc-8-of-18-tf-magnets; weight .25-.40 -> p50 2035-2037). Prev 2028/2032/2060 P .93 -> 2029/2033/2060 P .93. Working: notes/fusion/20261018T192708Z.md.
- Change
- Prev 2026-10-15-milestone-the-first-watt-3 (assess:20261015T105338Z; yellow, 2028/2032/2060, P .93), reproduced exactly (2028.78/2032.43/2060.88, P .9324); unchanged on driver draws as of 2026-10-18. Trigger: newly ingested 2026 events (gather 2026-10-18). (a) New evidence 2026-09-08-helion-orion-2028-goal-softened: helion-branch ramp tri(0.25,1.0,2.0) yr between start of operations and sustained net export (prev 0): p10 2028 -> 2029 (2028.78 -> 2029.81), p50 2032 -> 2033 (2032.43 -> 2033.12); p90 2060 and P .9324 unchanged. The other new events (SPARC delays, CFS PJM request and $1B, TAE 2031 and going-concern doubt, Type One 2034, General Fusion 2035, FIA survey, Polaris D-T) bear on driver inputs and cross-checks: reported as a driver finding, no move. Status yellow unchanged.
- Rubric rule
- RUBRIC v6 §4/§4.1 on the d964534 wording: no verified achievement; measurable construction and licensing progress; unconditional p50 2033 <= made_at+10 yr gives yellow; not green (p50 after 2031-10, no net-electricity precursor at relevant scale). Status unchanged.
by assess:20261018T192708Z · 48 evidence event(s)Evidence and news: 48 cited by the assessment, 9 other event(s)
Cited evidence (48)
- China's BEST tokamak project in Hefei had its project campus delivered for use and its main-machine assembly more than halfway complete as of early October 2026, with construction expected to finish by the end of 2027 and deuterium-tritium fusion power-generation demonstrations planned around 2030.
- Trump Media's Form S-4 for its merger with TAE Technologies (filed September 30, 2026) states that TAE plans to start building its Da Vinci plant in 2026, reach first plasma in 2029, demonstrate fusion energy in 2031 and begin initial operations thereafter, and discloses that TAE's Norm device has demonstrated sustained temperatures of only 75 million °C, that its experimental reactors generally run on hydrogen with occasional small amounts of deuterium, and that failure with hydrogen-boron could force a shift to deuterium-tritium.
- TAE Technologies' financial statements in the TMTG merger Form S-4 filed September 30, 2026 show an accumulated deficit of $1.8 billion and cash of $114.3 million at June 30, 2026 (down from $151.9 million at March 31), state that recurring losses raise substantial doubt about TAE's ability to continue as a going concern, and now expect the all-stock merger, announced in December 2025 for mid-2026, to close in late 2026.
- On September 30, 2026, Commonwealth Fusion Systems agreed to buy more than 10,000 km (3 GA-m) of high-temperature superconducting tape from Fujikura, the largest such order, for its ARC power plants starting with the first at the Fall Line Fusion Power Station in Virginia, with Fujikura expanding its manufacturing capacity to supply it.The First WattD-T Fusion [1] CFS — CFS places largest single purchase order of HTS tape with Fujikura (2026-09-30)
- On September 22, 2026, Commonwealth Fusion Systems reported that 8 of the 18 toroidal-field magnets of its SPARC tokamak were installed, behind its January 2026 expectation of installing all 18 by the end of summer 2026.
- Axios Pro reported on September 8, 2026 that Helion had removed from its website the prediction that its Orion plant would produce electricity by 2028 (the Orion page now says it will 'begin initial operations in 2028'), and that a Helion executive said Orion's 50 MW would in no case reach full operation in 2028, with full power not before 2029 or 2030, while the Microsoft contract requires grid connection and start of operations by end-2028.The First WattAneutronic Fusion [1] Axios Pro — Helion quietly shifts goalpost on fusion energy (2026-09-08) [2] Helion — Orion
- On August 31, 2026, Tennessee issued Type One Energy the first fusion-specific byproduct material license in the US under the new state framework for fusion machines, for its Project Infinity at TVA's Bull Run site, where Type One intends to break ground on the first phase in 2026 and expects full startup by 2034 of what is forecast as the 400 MWe Infinity Two stellarator plant.
- General Fusion, which became the first publicly listed fusion company on Nasdaq in July 2026 with about US$150 million of cash to fund its LM26 programme through key milestones by end-2028, said on August 18, 2026 that it aims to deploy its first-of-a-kind fusion plant by approximately 2035.
- TAE Technologies said on August 5, 2026 that it expects its first fusion power plant, the 50 MWe Da Vinci, to begin operations in 2031, with a site to be selected later in 2026.The First WattAneutronic Fusion [1] TAE — TAE and Black Moon Energy enter helium-3 supply and commercialization agreement (2026-08-05)
- On August 5, 2026, Commonwealth Fusion Systems said a DOE-assembled panel of 14 independent experts had approved its ARC power plant preconceptual design and technology roadmap as milestones of DOE's Milestone-Based Fusion Development Program, for a plant designed for 400 MW net electrical output, and that SPARC was almost 80% complete.
- On July 30, 2026, Commonwealth Fusion Systems announced $1 billion of additional equity financing, the largest fusion round since its own 2021 Series B, bringing its total capital raised to $4 billion to finish assembling SPARC and develop its first ARC plant at the Fall Line site, which it says is on track to put power on the grid in the early 2030s.The First WattD-T Fusion [1] CFS — CFS raises another $1 billion, bringing total capital raised to $4 billion (2026-07-30)
- The Fusion Industry Association's 2026 global fusion industry report (56 companies surveyed) counts a record $4.48 billion raised in the 12 months to July 2026 and $14.24 billion in total funding to date, and of 45 companies answering when the first fusion plant will deliver commercial electricity, 3 said 2026–2030, 16 said 2031–2035 (21 a year earlier), 13 said 2036–2040, 9 said 2041–2045 (5 a year earlier) and 4 said later.The First Watt [1] Fusion Industry Association — The global fusion industry in 2026 (report PDF, 2026-07-13) [2] FIA — Fusion industry attracts record annual funding of $4.48bn, raising total to $14.24bn (2026-07-13) [3] Fusion Industry Association — The global fusion industry in 2026 (report PDF, revised 2026-07-15) [4] Power Technology — Nuclear fusion industry attracts record annual funding of $4.48bn (2026-07-15)
- On June 20, 2026, the National Ignition Facility achieved fusion ignition for the 11th time, with a measured yield of 7.9 ± 0.4 MJ and a target gain of about 3.8, below the 8.6 MJ / gain 4.13 record of April 2025.
- On June 16, 2026, Helion received a Radioactive Materials License and a Radioactive Air Emissions License from the Washington Department of Health for its Orion fusion plant in Malaga, Washington, which lets it proceed with the generator building (earthwork began in spring 2026), while a transmission interconnection agreement with Chelan County PUD was still being worked toward.
- On June 4, 2026, Helion announced a $465 million Series G led by Thrive Capital at a $15.5 billion post-money valuation, bringing total investment in the company to $1.5 billion, to scale manufacturing and deploy its Orion fusion plant.The First WattAneutronic Fusion [1] Helion — Helion raises $465 million Series G (2026-06-04, updated September 2026)↳ follows up on 2025-01: Helion announced a $425 million Series F.
- GeekWire reported on May 8, 2026 that Helion is building Tiny Merge, a fusion testbed less than one-eighth the size of its Polaris prototype, to investigate plasma-ring (FRC) formation and merging questions that its larger machines had not fully resolved before Orion's design is locked, with operation planned by the end of summer 2026.
- On April 28, 2026, Commonwealth Fusion Systems became the first fusion developer to apply to connect a power plant to a major US grid operator, filing a PJM Interconnection request for its first 400 MW ARC plant at the newly named Fall Line Fusion Power Station in Chesterfield County, Virginia, which it targets to deliver electricity in the early 2030s (interconnection studies to generation can take four to six years).The First WattD-T Fusion [1] CFS — CFS becomes first fusion company to apply to PJM Interconnection (2026-04-28)
- Helion announced on February 13, 2026 that its seventh-generation Polaris prototype became the first privately developed fusion machine to demonstrate measurable deuterium-tritium fusion (in January 2026) and reached plasma temperatures of 150 million °C; Helion said it will keep raising Polaris plasma temperatures to show reliable operation with deuterium-helium-3, the fuel relevant for its future commercial operations.The First WattAneutronic Fusion [1] Helion — Helion achieves new industry-first fusion energy milestones (2026-02-13)
- Comments by Max Planck IPP's Karl Lackner and co-authors in the Journal of Fusion Energy (published February 7, 2026) argue that Helion's D-He3 pulsed-FRC gain projections (Kirtley & Milroy, 2023), which rest on an ion-to-electron temperature ratio of about 10, neglect collisional ion-to-electron power transfer that exceeds the D-He3 fusion power by more than an order of magnitude at 100 keV ion temperature (over two orders at 40 keV), implying much tighter requirements on plasma lifetime, losses and direct-conversion efficiency.Aneutronic Fusion [1] Lackner et al., Comments on the Paper 'Fundamental Scaling of Adiabatic Compression of FRC Thermonuclear Fusion Plasmas', J. Fusion Energy (2026-02-07) [2] Scientific American — Helion Energy wants to build fusion power on a start-up timeline (2026-05-19; includes Helion's response)
- China's BEST (Burning Plasma Experimental Superconducting Tokamak) project in Hefei, presented at the Fusion Energy Technology and Industry Conference 2026, aims to demonstrate net fusion power gain and fusion-based electricity generation by around 2030 (a government experimental device, not a commercial plant).
- On January 6, 2026, Commonwealth Fusion Systems installed the first of the 18 toroidal-field magnets of its SPARC demonstration tokamak, said it expected to install all 18 by the end of summer 2026, and said it hoped to turn SPARC on in 2027.
- In 2025, Zap Energy's FuZE-3 device reached a plasma pressure of 1.6 GPa (about 10,000 times atmospheric pressure), and the company is scaling to FuZE-A.↳ corrects (or adds sources to) an earlier record:
2025: In 2025, Zap Energy's FuZE-3 device reached a plasma pressure of 1.6 GPa (about 10,000… - The UK government committed £2.5 billion to fusion energy over five years.Beyond the framework [1] Kleinman Center: Bringing fusion energy to the grid↳ corrects (or adds sources to) an earlier record:
2025: The UK government committed £2.5 billion to fusion energy over five years. - In 2025, TAE Technologies achieved a world-first formation of a field-reversed configuration plasma using neutral beam injection only.↳ corrects (or adds sources to) an earlier record:
2025: In 2025, TAE Technologies achieved a world-first formation of a field-reversed… - TAE’s hydrogen-boron power-plant design aims to deliver commercial fusion by the early 2030s.The First WattAneutronic Fusion [1] TAE Technologies: TAE shortens device roadmap, prepares for commercial era↳ follows up on 2025-11-17: On November 17, 2025, TAE Technologies announced it would skip its sixth-generation… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025: TAE Technologies targets a commercial p-B11 fusion power plant in the early 2030s. - By 2025, Helion Energy's Polaris D-He3 field-reversed-configuration fusion prototype was operational and the company had over 500 employees.Aneutronic FusionThe First Watt [1] Helion: Starting to build the world's first fusion power plant in Malaga, WA
- Helion Energy targets delivering fusion electricity to Microsoft by 2028 under a power purchase agreement.The First WattAneutronic Fusion [1] Helion: Starting to build the world's first fusion power plant in Malaga, WA↳ updates 2024: Helion Energy missed its previously stated target of demonstrating net electricity from… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025: Helion Energy targets delivering fusion electricity to Microsoft by 2028 under a power… - In 2025, First Light Fusion published its FLARE concept, a path to fusion gain of ~1,000 at a facility cost of $100–200M (versus $5.3B for NIF), and pivoted to a technology-licensing model.↳ corrects (or adds sources to) an earlier record:
2025: In 2025, First Light Fusion published its FLARE concept, a path to fusion gain of… - In 2025, Commonwealth Fusion Systems had its SPARC magnets validated by the US DOE (an $8M milestone payment), took delivery of the SPARC vacuum vessel and had SPARC assembly underway.
- Commonwealth Fusion Systems targets SPARC operation in 2026 and its ARC commercial fusion power plant in the early 2030s.↳ corrects (or adds sources to) an earlier record:
2025: Commonwealth Fusion Systems targets SPARC operation in 2026 and its ARC commercial… - In 2025, Google agreed to buy half the power produced at the CFS plant.↳ corrects (or adds sources to) an earlier record:
2025: In 2025, Google agreed a deal to buy 200 MW of fusion power from Commonwealth Fusion… - In December 2025, Tokamak Energy reported three performance records on its ST40 spherical tokamak, including 1 MA plasma current and the highest triple product achieved by a private company (~8×10¹⁸ m⁻³·keV·s).↳ corrects (or adds sources to) an earlier record:
2025-12: In December 2025, Tokamak Energy reported three performance records on its ST40… - Tokamak Energy targets commercial spherical-tokamak fusion power in the 2030s.
- CATF reported on December 18, 2025, that DOE announced the return of a standalone Office of Fusion.↳ corrects (or adds sources to) an earlier record:
2025: The US Department of Energy established an Office of Fusion, its first standalone… - Germany signed a memorandum of understanding for a laser fusion power plant at the former Biblis nuclear site, targeted by 2035.Beyond the framework [1] ANS Nuclear News: Industry update May 2025 [2] Wayback capture 20250429152151↳ corrects (or adds sources to) an earlier record:
2025: Germany signed a memorandum of understanding for a laser fusion power plant at the… - In November 2025, Tokamak Energy validated its Demo4 high-temperature-superconducting magnet system at 11.8 T and −243°C.Beyond the frameworkD-T Fusion [1] Tokamak Energy announces fusion power plant magnet technology breakthrough↳ corrects (or adds sources to) an earlier record:
2025-11: In November 2025, Tokamak Energy validated its Demo4 high-temperature-superconducting… - On November 17, 2025, TAE Technologies announced it would skip its sixth-generation device (Copernicus) and accelerate directly toward a commercial fusion power plant.The First WattAneutronic Fusion [1] TAE Technologies: TAE shortens device roadmap, prepares for commercial era↳ corrects (or adds sources to) an earlier record:
2025-11-17: On November 17, 2025, TAE Technologies announced it would skip its sixth-generation… - Helion Energy secured a Conditional Use Permit from Chelan County to build the fusion generator building for its Orion plant near Malaga.The First WattAneutronic Fusion [1] GeekWire: Helion lands permit for constructing reactor building at Washington site [2] Wayback capture 20260105113659↳ follows up on 2025-07-30: On July 30, 2025, Helion Energy began construction of its Orion fusion power plant in… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-10-15: On October 15, 2025, Helion Energy secured a building permit from Chelan County,… - On August 28, 2025, Commonwealth Fusion Systems announced an $863M Series B2 round, bringing its total funding to about $3B, the largest deep-tech raise of 2025.↳ corrects (or adds sources to) an earlier record:
2025-08-28: On August 28, 2025, Commonwealth Fusion Systems announced an $863M Series B2 round,… - The Fusion Industry Association’s 2025 report says total fusion investment surpassed $9.7 billion, while surveyed companies estimate more than $77 billion in total capital required to bring every surveyed company to commercialization.↳ corrects (or adds sources to) an earlier record:
2025-07: The Fusion Industry Association's 2025 Global Fusion Industry Report found cumulative… - On July 30, 2025, Helion Energy began construction of its Orion fusion power plant in Malaga, Washington, described as the first commercial fusion power plant site.The First WattAneutronic Fusion [1] Helion: Starting to build the world's first fusion power plant in Malaga, WA↳ corrects (or adds sources to) an earlier record:
2025-07-30: On July 30, 2025, Helion Energy began construction of its Orion fusion power plant in… - In June 2025, TAE Technologies raised $150M, bringing its total funding to over $1.8B.Aneutronic Fusion [1] TAE shortens device roadmap, prepares for commercial era
- On June 22, 2025, a Los Alamos National Laboratory-led team working with LLNL achieved ignition using NIF.D-T Fusion [1] LLNL: Achieving Fusion Ignition↳ corrects (or adds sources to) an earlier record:
2025-06: In June 2025, LLNL's National Ignition Facility achieved ignition on its THOR… - On April 7, 2025, the National Ignition Facility at LLNL achieved a target gain of Q=4.13, producing 8.6 MJ of fusion yield from 2.08 MJ of laser energy input, the highest fusion gain ever recorded.↳ corrects (or adds sources to) an earlier record:
2025-04-07: On April 7, 2025, the National Ignition Facility at LLNL achieved a target gain of… - On February 12, 2025, CEA's WEST tokamak set a world record for plasma duration of 22 minutes 17 seconds.The First Watt [1] CEA: WEST beats the world record for plasma duration↳ corrects (or adds sources to) an earlier record:
2025-02-12: On February 12, 2025, CEA's WEST tokamak set a world record for plasma duration of 22… - Helion announced a $425 million Series F.↳ corrects (or adds sources to) an earlier record:
2025-01: In January 2025, Helion Energy raised a $425M Series F, bringing its total funding to… - ITER's 2024 baseline reset moved first plasma to 2034 and deuterium-tritium operations to 2039, a delay of about nine years.↳ corrects (or adds sources to) an earlier record:
2024: ITER's 2024 baseline reset moved first plasma to 2034 and deuterium-tritium operations… - Helion Energy missed its previously stated target of demonstrating net electricity from fusion in 2024.↳ corrects (or adds sources to) an earlier record:
2024: Helion Energy missed its previously stated target of demonstrating net electricity from…
Other events (9)
- Japan's METI conditionally selected Helical Fusion as an operator in its fusion power-generation demonstration grant programme, under which about JPY 60 billion (USD 370 million) in subsidies is expected to go to the selected operators by the end of February 2029 toward power-generation demonstrations in the 2030s.
- Japan's fusion roadmap adopted on July 21, 2026 by the Headquarters for Japan's Growth Strategy targets a fusion power-generation demonstration in the 2030s, estimates JPY 3.1 trillion of public-private investment through FY2040 (assuming one demonstration plant), and notes that fusion power generation has yet to be demonstrated anywhere in the world.
- On June 4, 2026, the US Department of Energy released its finalized Fusion Science and Technology Roadmap, a national strategy aimed at fusion pilot plants and commercialization by the mid-2030s, implemented through DOE's new Office of Fusion, while stating that the roadmap does not commit DOE to specific funding levels.
- Type One Energy, Tokamak Energy and AECOM formed the UK Infinity Fusion Consortium on May 6, 2026 to develop a private-sector-led UK project based on Type One's 400 MWe Infinity Two stellarator plant design, with no site or date announced.
- TAE Technologies completed on April 7, 2026 a site-evaluation tour of Alabama, Ohio and Texas (begun March 26) for its first fusion power plant, which targets about 50 MWe of electricity generation in the early 2030s; no site was selected.
- +4 older event(s): the section's event log
Grid Parity: Fusion Levelized Cost Of Electricity falls below $100/MWh, competitive with gas.Distant: fusion cost unmeasured; first plants estimated >$150/MWh vs <$100 (2025$) goal, ETA ~2071 · assessed 2026-10-18 · 22 event(s)
- $100/MWh in constant 2025 US dollars is the test ("competitive with gas" explains it, it is not a second threshold).
No fusion plant runs, so fusion's cost of electricity is unmeasured; the only independent estimate puts early plants above $150/MWh, against the under-$100/MWh (2025 dollars) goal. A 2026 Nature Energy study finds fusion plants likely to get cheaper by only 2-8% for each doubling of capacity built, far slower than the roughly 20% of offshore wind that the previous forecast borrowed. Starting the cost decline from the first plant's arrival (see The First Watt) and allowing a 20% chance that costs never fall (as with fission), the chance by 2100 is 66% and the median 2071 (previously 2046): decades away, so red. How fast fusion costs fall with experience is what decides it.- ETA
- ~2071 (80% range 2038–never/after 2100; P(by 2100) 66%)
- ETA cruxes
- Mapping first: fusion experience rate (Tang 2-8%) vs the offshore analogue (~20%): mapping weight 0/.5/1 -> 2079/2062/2046, LR 2% -> null, 8% -> 2058, offshore LR 15% -> 2063; non-learning (P .53-.81, p50 2061-2090); FOAK cost and its dollar basis ($150-250, 2021 $: 2063-2095); deployment pace (recent offshore rate -> null); First Watt timing (2070-2075). Earlier: audited FOAK LCOE <$120 (2025 $), repeat orders with measured capex declines >8% per doubling, a cheap mass-manufactured design. Later/never: FOAK overruns, low availability, redesigns, no repeat orders.
- Path
- Reading (owner definition d964534): fusion-plant LCOE strictly below $100/MWh in constant 2025 USD; "competitive with gas" is explanatory, not a second threshold; not a subsidised PPA or market price. Proxy chain (no fusion cost series): First Watt start of commercial operation (shared driver fusion-first-watt) -> FOAK cost (Lindley >$150) -> cost decline = offshore wind's deployment-driven decline (shared driver irena-lcoe) rescaled to fusion's experience rate (Tang et al. 2026; transfer assumed, rate sourced), weight .25 on the unscaled offshore rate. Gas LCOE ($77-102 benchmarks, 2025-26) is context only.
- Blockers
- Technical/economic: no operating fusion plant, so no measured capital, O&M, availability or LCOE.
- Economic: independent early FOAK estimate >$150/MWh vs <$100 (2025 $); a Nature Energy study puts fusion experience rates at 2-8%, so learning closes the gap slowly.
- Data: no fusion deployment series to measure Wright learning; offshore-wind capacity (for the transfer) not in the ledger; driver irena-lcoe not yet refit with IRENA 2025.
- Gather: offshore wind cumulative capacity series, Lindley's dollar year, a second independent fusion experience-rate estimate, real plant cost/availability data.
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- learning curve (Wright-law transfer: offshore-wind deployment pace from shared driver irena-lcoe, fusion experience rate from Tang et al. 2026, mixed with the prior 1:1 offshore mapping) after the shared First Watt timing (driver fusion-first-watt), with a counted non-learning branch; reference-class and discounted-survey cross-checks
- ETA calculation
- Gap: fusion LCOE unmeasured; 2023-01-01-lindley-fusion-foak-lcoe-rv >$150/MWh early magnetic FOAK (sourced); assumed median C0 $170 (2024 $) = $174.5 (2025 $, CPI-U 321.943/313.689) -> 1.74x vs $100. C0 consistency: 2026-07-10-whyte-lo-fusion-economic-q-framework ($100/MWh <-> ~$12/W at 2% real) and 2026-07-10-tokamak-energy-st-e1-capex-target ($12-14k/kW commercial phase -> ~$100-111 at 2% real, pilot above). Analogue: driver irena-lcoe r_offshore (lognormal(ln .075651,.55); offshore 208/155/114/91/81/79 2010-24, 2024-01-01-irena-lcoe-offshore-wind-2024-rv; 2025 = 78 in 2025 $ (-3.8% real), 2026-07-02-irena-lcoe-2025-solar-wind, not yet in the fit: driver finding). Model = prev 20261016T063036Z (reproduced exactly: P .7475, 2033.36/2046.04/null): G_i = start_i + ln(C0/thr)/r_i, start_i = draw i of shared driver fusion-first-watt, r_offshore_i = draw i of shared driver irena-lcoe (driver-draws --as-of 2026-10-18, seed 20261015, n 10000); thr $97.436 (2024 $); 20% non-learning; FW never -> never. Mapping update on new evidence 2026-03-23-nature-energy-fusion-experience-rates-overestimated (fusion experience rates 2-8%, adopted 5%, vs 8-20% assumed): offshore's decline embeds LR ~20% (b_off = .075651 / ln(83/3.1)/14 = .075651/.2348 = .322; capacity 3.1 -> 83 GW 2010-24, GWEC, sourced); fusion r_i = r_offshore_i x b_fus/b_off with LR_fus tri(2,5,8)% (ratio at 5% .229, r ~.0174/yr; 0.5566/.0174 = 32 yr after the first plant), weight .75; prev 1:1 mapping weight .25. Failure scenario non-learning 20%: class = capital-intensive generators commercialised since 1950 (fission, CCGT, onshore, offshore, PV, CSP): 1 of 6 (fission) no real cost decline -> .17, Laplace .25; .20 kept. FW never .068 (driver); slow learning missing 2100 .093 (model outcome). MC: P(by2100) .6555, p10 2038.00, p50 2071.30, p90 null (calendar 2038/2071/null); conditional median 2054.42; P(by end-2050) .284. Sensitivities: Tang only 2079.41 (P .625), weight .5 2062.05, prev mapping 2046.03 (.748); LR_fus 2%/8%/12% -> null (.423)/2058.12/2051.50; LR_off 15%/25% -> 2062.70/2080.61; recent offshore rate -> null (.463); C0 150/190/250 -> 2063.15/2078.11/2095.05; C0 in 2021 $ 2080.25; non-learning .10/.35/0 -> 2065.07/2090.09/2061.18; follow-on +3 yr 2074.28; FW slip 3x 2074.84, helion weight .25 2073.61. Cross-checks: first commercial plant -> LCOE <$100 took fission ~10-15, onshore ~15-20, offshore ~28, PV ~30 yr: ~20 yr from ~2032 -> ~2052 = our conditional median 2054; FIA viability survey (2026-07-13-fia-2026-survey-cost-viability-timing) median ~2038.9, discounted 1.5/2/3x -> 2044.9/2050.9/2063.0 (looser event; our P(by end-2063) .44). No anchors. Prev 2033/2046/null P .75 -> 2038/2071/null P .66. Working: notes/fusion/20261018T192708Z.md.
- Change
- Prev 2026-10-16-milestone-grid-parity (assess:20261016T063036Z; red, 2033/2046/null, P .75), reproduced exactly (2033.36/2046.04/null, P .7475); on 2026-10-18 driver draws 2033.35/2046.03/null. Trigger: newly ingested 2026 events. (a) New evidence 2026-03-23-nature-energy-fusion-experience-rates-overestimated (fusion experience rates 2-8%, adopted 5%) contradicts prev's 1:1 offshore mapping, whose implied experience rate is ~20% (b_off .322 from GWEC capacity 3.1 -> 83 GW 2010-24): replaced by r = r_offshore x b_fus/b_off, LR_fus tri(2,5,8)%, weight .75 (prev mapping .25, assumed): p10 2033 -> 2038 (2033.35 -> 2038.00), p50 2046 -> 2071 (2046.03 -> 2071.30), p_achieved .75 -> .66 (.7475 -> .6555); p90 null. Tang-only would give 2044/2079/.625. Other new evidence (IRENA 2025 points, Whyte-Lo, Tokamak Energy capex target, gas LCOE rises, FIA viability survey) informs cross-checks and C0 consistency; no input moved. Status red unchanged.
- Rubric rule
- RUBRIC v6 §4/§4.1: no verified fusion LCOE below $100/MWh (2025 USD); unconditional p50 2071 > made_at+10 yr gives red/Distant. Status unchanged.
by assess:20261018T192708Z · 22 evidence event(s)Evidence and news: 22 cited by the assessment
Cited evidence (22)
- Axios Pro reported on September 8, 2026 that Helion had removed from its website the prediction that its Orion plant would produce electricity by 2028 (the Orion page now says it will 'begin initial operations in 2028'), and that a Helion executive said Orion's 50 MW would in no case reach full operation in 2028, with full power not before 2029 or 2030, while the Microsoft contract requires grid connection and start of operations by end-2028.The First WattAneutronic Fusion [1] Axios Pro — Helion quietly shifts goalpost on fusion energy (2026-09-08) [2] Helion — Orion
- Lazard's 2026 Levelized Cost of Energy+ (version 19.0, released July 13, 2026) gives unsubsidized US LCOE ranges of $51–129/MWh (average $90) for new gas combined cycle, which it calls a 15-year high, $175–255/MWh (average $215) for new nuclear, $40–98/MWh (average $69) for utility-scale solar PV and $37–99/MWh (average $68) for onshore wind.
- In the Fusion Industry Association's 2026 global fusion industry report survey, 42 fusion companies answered when the first fusion plant will demonstrate a low enough cost/high enough Q to be considered commercially viable: 2 said 2026–2030, 11 said 2031–2035, 14 said 2036–2040, 8 said 2041–2045, 5 said 2046–2050 and 2 said after 2050.
- A Journal of Fusion Energy paper by MIT's Dennis Whyte, Andrew Lo and co-authors (published online 10 July 2026) proposes a concept-agnostic 'economic gain' (Q_econ ≥ 1) framework for fusion power plants and, in its illustrative base case (net electricity value $100/MWh, 2% real interest rate, 40% conversion efficiency, 2025 US dollars), finds economic breakeven at an overnight cost of about $12/W and, at Q_econ = 1.5 (about $6/W), an effective levelized cost of electricity of $67/MWh.
- Tokamak Energy's pre-concept design of its ST-E1 spherical-tokamak fusion power plant, published in Nuclear Fusion on 10 July 2026, targets in its commercial phase a normalised capital cost of $12,000–14,000 per kW of net electric power, 800–1,000 MW net electric output and availability above 80%, after a pilot phase not expected to reach commercially competitive capital cost or availability.
- IRENA's Renewable Power Generation Costs in 2025 report (July 2026) gives the 2025 global weighted-average LCOE of newly commissioned projects as $44/MWh for solar PV (unchanged from 2024), $33/MWh for onshore wind and $78/MWh for offshore wind (2025 USD), with solar PV down 89%, onshore wind 71% and offshore wind 63% since 2010.
- IRENA's Renewable Power Generation Costs in 2025 report finds that a gas-turbine shortage roughly doubled the capital cost of a new combined-cycle gas plant to about USD 2,400/kW in the United States, pushing the LCOE of new gas-fired generation towards USD 100/MWh in higher-gas-price markets such as Italy, Germany and Japan, while it stayed in the USD 50–60/MWh range where gas is cheaper.
- EIA's Levelized Costs of New Generation Resources in the Annual Energy Outlook 2026 (Counterfactual Baseline case) projects an average levelized cost of $77.46/MWh (2025 dollars) for new US natural gas combined-cycle plants entering service in 2031, versus $87.81/MWh for advanced nuclear, $58.33/MWh for solar PV and $56.75/MWh for onshore wind.
- An ETH Zurich analysis in Nature Energy (Tang, Noll, Panda and Schmidt, published 23 March 2026) finds that fusion power plants' large unit size, extreme design complexity and intermediate need for customization are empirically linked to experience rates of 2–8% (it adopts 5%) rather than the 8–20% assumed in existing fusion cost projections, compiles first-of-a-kind fusion plant capital cost estimates ranging from US$1,400 to US$43,000 per kW, and concludes that fusion's cost trajectory will probably remain far above competing energy technologies.
- BloombergNEF's Levelized Cost of Electricity 2026 report finds that the global benchmark levelized cost of electricity of new combined-cycle gas turbine plants rose 16% in 2025 to $102/MWh, the highest on record, driven by gas-turbine equipment prices, while the benchmark for fixed-axis solar was $39/MWh, onshore wind $40/MWh and offshore wind $100/MWh.
- Helion Energy targets delivering fusion electricity to Microsoft by 2028 under a power purchase agreement.The First WattAneutronic Fusion [1] Helion: Starting to build the world's first fusion power plant in Malaga, WA↳ updates 2024: Helion Energy missed its previously stated target of demonstrating net electricity from… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025: Helion Energy targets delivering fusion electricity to Microsoft by 2028 under a power… - Commonwealth Fusion Systems targets SPARC operation in 2026 and its ARC commercial fusion power plant in the early 2030s.↳ corrects (or adds sources to) an earlier record:
2025: Commonwealth Fusion Systems targets SPARC operation in 2026 and its ARC commercial… - Lazard's LCOE+ 2025 (June 2025) gives LCOE ranges of $27–73/MWh (avg $50) for onshore wind, $29–92/MWh (avg $58) for solar PV, $41–108/MWh (avg $78) for gas combined cycle and $142–222/MWh for new nuclear fission.↳ corrects (or adds sources to) an earlier record:
2025: Lazard's LCOE+ 2025 (June 2025) gives LCOE ranges of $27–73/MWh (avg $50) for onshore… - On August 28, 2025, Commonwealth Fusion Systems announced an $863M Series B2 round, bringing its total funding to about $3B, the largest deep-tech raise of 2025.↳ corrects (or adds sources to) an earlier record:
2025-08-28: On August 28, 2025, Commonwealth Fusion Systems announced an $863M Series B2 round,… - On July 30, 2025, Helion Energy began construction of its Orion fusion power plant in Malaga, Washington, described as the first commercial fusion power plant site.The First WattAneutronic Fusion [1] Helion: Starting to build the world's first fusion power plant in Malaga, WA↳ corrects (or adds sources to) an earlier record:
2025-07-30: On July 30, 2025, Helion Energy began construction of its Orion fusion power plant in… - TIME reports that CFS executives anticipate good business at $100/MWh and say fusion takes over the world at $50/MWh.↳ corrects (or adds sources to) an earlier record:
2025: Commonwealth Fusion Systems, as reported by TIME in 2025, projects that nth-of-a-kind… - IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned utility-scale solar PV as $43/MWh in 2024 (2024 USD), down from $417 (2010), $135 (2015), $79 (2018), $61 (2020) and $51/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned onshore wind as $34/MWh in 2024 (2024 USD), down from $113 (2010), $76 (2015), $56 (2018), $41 (2020) and $34/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned offshore wind as $79/MWh in 2024 (2024 USD), down from $208 (2010), $155 (2015), $114 (2018), $91 (2020) and $81/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - ITER's 2024 baseline reset moved first plasma to 2034 and deuterium-tritium operations to 2039, a delay of about nine years.↳ corrects (or adds sources to) an earlier record:
2024: ITER's 2024 baseline reset moved first plasma to 2034 and deuterium-tritium operations… - Helion Energy missed its previously stated target of demonstrating net electricity from fusion in 2024.↳ corrects (or adds sources to) an earlier record:
2024: Helion Energy missed its previously stated target of demonstrating net electricity from… - Lindley et al. (Energy Policy, 2023) estimate that first-of-a-kind fusion power plants would have a levelized cost of electricity above $150/MWh.↳ corrects (or adds sources to) an earlier record:
2023: Lindley et al. (Energy Policy, 2023) estimate that first-of-a-kind fusion power plants…
Tech Tree
D-T Fusion: (easiest ignition, but 14 MeV neutrons + tritium breeding required)75 event(s)
- General challenges: tritium breeding (TBR>1) and neutron-resistant materials at scale
- Magnetic confinement (CFS, ITER, Zap Energy) — challenges: plasma stability, heat exhaust, materials survival
- Inertial confinement (NIF, First Light) — challenges: rep-rate (need 10 Hz), target cost ($50k→$0.10), driver efficiency
Events (75)
- General Fusion and UKAEA announced on October 8, 2026 that the LM26 magnetized-target-fusion machine reached electron temperatures of about 1.1 keV by Thomson scattering just before peak compression (about 1.2 keV by AXUV at peak) with ion heating to about 0.46 keV, the company's 1 keV milestone (a heating result, not net energy; paper submitted for peer review).
- On October 7, 2026, Zap Energy said DOE approved two milestones for its in-house pulsed-power capacitor: a 35 kV, 50,000 A design projected to last 6 billion pulses (versus about 100,000 for typical commercial units) and its first manufactured prototype (lifetime is a design projection).
- China's BEST tokamak project in Hefei had its project campus delivered for use and its main-machine assembly more than halfway complete as of early October 2026, with construction expected to finish by the end of 2027 and deuterium-tritium fusion power-generation demonstrations planned around 2030.
- On October 6, 2026, Xcimer Energy finalized an agreement for up to $30.5 million of additional milestone-contingent funding in the second budget period of DOE's Milestone-Based Fusion Development Program, bringing its potential DOE milestone payments to $39.5 million.
- Stellarator developer Type One Energy announced on October 6, 2026 the completion of a $200 million Series B financing round.
- +70 older event(s): the section's event log
Aneutronic Fusion: (D-He3, p-B11 — no neutrons/tritium, but 10× hotter, lower cross-section)29 event(s)
- Helion (D-He3, 2028 target), TAE (p-B11, early 2030s)
- Challenge: net gain at 1+ billion °C
Events (29)
- Trump Media's Form S-4 for its merger with TAE Technologies (filed September 30, 2026) states that TAE plans to start building its Da Vinci plant in 2026, reach first plasma in 2029, demonstrate fusion energy in 2031 and begin initial operations thereafter, and discloses that TAE's Norm device has demonstrated sustained temperatures of only 75 million °C, that its experimental reactors generally run on hydrogen with occasional small amounts of deuterium, and that failure with hydrogen-boron could force a shift to deuterium-tritium.
- TAE Technologies' financial statements in the TMTG merger Form S-4 filed September 30, 2026 show an accumulated deficit of $1.8 billion and cash of $114.3 million at June 30, 2026 (down from $151.9 million at March 31), state that recurring losses raise substantial doubt about TAE's ability to continue as a going concern, and now expect the all-stock merger, announced in December 2025 for mid-2026, to close in late 2026.
- ENN announced on September 28, 2026 that its EXL-50U spherical tokamak had achieved proton-boron (p-11B) fusion reactions in late September (first signals on September 19), measuring about 10^8 alpha particles per second by using neutral-beam injection plus RF waves to accelerate fast protons into the 162–200 keV cross-section resonance over a ~2 keV thermal plasma — the first p-B11 fusion on a commercial company's own device, a reaction demonstration far from energy gain.Aneutronic Fusion [1] ENN Research — 新奥氢硼聚变商业化取得重大突破 “玄龙-50U”装置实现氢硼聚变反应 (2026-09-28) [2] China Daily (Chinese) — 新奥“玄龙-50U”装置实现氢硼聚变反应 (2026-09-28) [3] China Newsweek — 刚刚，可控核聚变又一重大突破 (2026-10-15; Sep 19 signal, ~2 keV background) [4] SCMP — Chinese commercial reactor achieves nuclear fusion with clean hydrogen-boron fuel (2026-09-28)
- On September 15, 2026, ENN broke ground in Langfang on EHL-2 (Helong-2), its third-generation spherical-torus proton-boron fusion device, which it targets to finish and start experiments by end-2027 and with which it aims to 'light the first lamp' of p-B11 fusion by 2030, followed by low-cost power generation and entry into the demonstration-reactor stage before 2035.Aneutronic Fusion [1] ENN Research — “和龙-2”建设项目开工，向氢硼聚变发电迈进 (2026-09-16)
- Axios Pro reported on September 8, 2026 that Helion had removed from its website the prediction that its Orion plant would produce electricity by 2028 (the Orion page now says it will 'begin initial operations in 2028'), and that a Helion executive said Orion's 50 MW would in no case reach full operation in 2028, with full power not before 2029 or 2030, while the Microsoft contract requires grid connection and start of operations by end-2028.The First WattAneutronic Fusion [1] Axios Pro — Helion quietly shifts goalpost on fusion energy (2026-09-08) [2] Helion — Orion
- +24 older event(s): the section's event log
Deep Geothermal: (most viable mid-term fusion alternative)31 event(s)
- Quaise, GA Drilling — millimeter-wave/plasma drilling to 10-20 km
Events (31)
- Fervo Energy declared commercial operation of Cape Station's first GeoBlock on September 30, 2026, one day ahead of its contractual date, with 33 MW of net power production and revenue now generated under its PPA; the other two 33 MW Phase I GeoBlocks are expected online by January 1, 2027 and the 400 MW Phase II in 2028.
- On September 24, 2026, the first 33 MW GeoBlock of Fervo Energy's Cape Station enhanced geothermal project in Beaver County, Utah, synchronized to the grid and achieved first power, which Fervo describes as the first time a utility-scale enhanced geothermal development has reached this milestone anywhere in the world.Deep Geothermal [1] Fervo Energy Achieves First Power at Cape Station
- On September 21, 2026, the US DOE announced more than $99 million for 21 geothermal projects, five field-scale enhanced geothermal tests and 16 exploration-drilling projects, including up to $25M for Quaise Energy's superhot demonstration at Newberry Volcano, Oregon (bottomhole 265–365°C, targeting a >25 MW-gross well triplet) and two Fervo Energy awards totaling about $20M in Idaho and Nevada.Deep Geothermal [1] DOE: Energy Department Announces $99 Million for 21 Projects to Advance U.S. Geothermal Energy Development [2] DOE: Pilots and Demonstrations: Next-Generation Field Tests and Drilling for New Resources [3] Quaise Energy Selected for Up to $25 Million in U.S. Department of Energy Support [4] Fervo Energy Secures Two U.S. Department of Energy Awards
- On September 17, 2026, Mazama Energy announced $135M in new capital (Series B led by Centaurus Capital and Doerr Capital, with ConocoPhillips and Shell Ventures) and said its second Newberry, Oregon well, Project Athena, reached 10,350 ft in 15 days (about 80% faster than its 2025 well that created a 331°C engineered geothermal system) and is drilling toward 400°C, with Project Ceres targeting 15 MWe per well and a power demonstration in 2027.
- On September 1, 2026, Fervo Energy and Google signed a 396 MW power purchase agreement for Cape Station enhanced geothermal capacity expected online in 2028, with an option for Google to add about 600 MW (nearly 1 GW in total) by June 2030.Deep Geothermal [1] Fervo Energy and Google Sign 396 MW PPA
- +26 older event(s): the section's event log
Beyond the framework: 67 event(s)
- On October 15, 2026, the IEA Governing Board said about 330 million barrels had been released so far under the March 2026 collective action over the Strait of Hormuz crisis, agreed to accelerate the remaining ~100 million barrels including over 20 million barrels of diesel, and noted total releases will exceed the original 400 million barrels.Beyond the framework [1] IEA Governing Board statement on energy markets and security
- On October 7, 2026, Germany's breakthrough-innovation agency SPRIND launched a milestone-based fusion funding programme for start-ups on behalf of the research ministry, with €500 million available through 2029 and total public funding to rise to as much as €2.5 billion in the longer term.Beyond the framework [1] SPRIND launches a Milestone Program for Nuclear Fusion
- On September 30, 2026, California Governor Gavin Newsom signed SB 925, requiring a state strategic plan for the development of fusion energy in California.Beyond the framework [1] Governor Newsom signs legislation to accelerate California's fusion industry
- The US NRC issued the Tennessee Valley Authority a construction permit for a GE Vernova Hitachi BWRX-300 small modular reactor at Clinch River, Tennessee, the first US construction permit for a BWRX-300, announced September 29, 2026.
- On September 24, 2026, US Representatives Zoe Lofgren and Jay Obernolte introduced the bipartisan American Leadership in Fusion Act, which would provide $10 billion in direct investment to accelerate fusion commercialization and establish an Office of Fusion at DOE (a bill, not yet law).Beyond the framework [1] Lofgren and Obernolte Lead Introduction of American Leadership in Fusion Act
- +62 older event(s): the section's event log
Latest news: the 8 most recently published of 237 events
- Eurostat's household electricity price dataset (nrg_pc_204, updated October 16, 2026) puts the EU average household electricity price (incl. all taxes and levies, 2,500–4,999 kWh/yr band) at €28.32 per 100 kWh in H1 2026, down from €28.96 in H2 2025 and €28.79 in H1 2025 (excluding taxes and levies: €20.27, from €20.59 in H2 2025).
- IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned utility-scale solar PV as $43/MWh in 2024 (2024 USD), down from $417 (2010), $135 (2015), $79 (2018), $61 (2020) and $51/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned onshore wind as $34/MWh in 2024 (2024 USD), down from $113 (2010), $76 (2015), $56 (2018), $41 (2020) and $34/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - IRENA's Renewable Power Generation Costs in 2024 report gives the global weighted-average LCOE of newly commissioned offshore wind as $79/MWh in 2024 (2024 USD), down from $208 (2010), $155 (2015), $114 (2018), $91 (2020) and $81/MWh (2022).KPIGrid Parity [1] IRENA: Renewable Power Generation Costs in 2024, datafile [2] IRENA: Renewable Power Generation Costs in 2024 [3] IRENA Renewable Power Generation Costs in 2024↳ corrects (or adds sources to) an earlier record:
2024: IRENA's Renewable Power Generation Costs in 2024 report gives the global… - On October 15, 2026, the IEA Governing Board said about 330 million barrels had been released so far under the March 2026 collective action over the Strait of Hormuz crisis, agreed to accelerate the remaining ~100 million barrels including over 20 million barrels of diesel, and noted total releases will exceed the original 400 million barrels.Beyond the framework [1] IEA Governing Board statement on energy markets and security
- The US BLS Consumer Price Index release for September 2026 (October 14, 2026) shows the CPI-U electricity index up 4.4% over the 12 months to September 2026 (not seasonally adjusted), versus a 3.7% rise in the all-items index.
- In an April 2026 field test at NORCE's Ullrigg well in Stavanger, Norway, GA Drilling ran two coordinated NexTitan downhole drive units that drilled 66 ft (~20 m) of hard phyllite (22,000 psi UCS) over 7.5 hours and demonstrated a downhole turbine generator as a battery alternative.
- General Fusion and UKAEA announced on October 8, 2026 that the LM26 magnetized-target-fusion machine reached electron temperatures of about 1.1 keV by Thomson scattering just before peak compression (about 1.2 keV by AXUV at peak) with ion heating to about 0.46 keV, the company's 1 keV milestone (a heating result, not net energy; paper submitted for peer review).
Health and Lifespan
304 events · latest news published 2026-10-16 (6 days ago)news per week
Some people long for immortality, others would just prefer not to die (or live a low-quality life) yet. The biggest symbol for fixing this is our journey to cure cancer, but the fight against other diseases such as Alzheimer's (or potential pandemics), as well as perfected early diagnostics, are just as important. And to achieve not just plain longevity, but long healthy lifespan, eventually we need to move further — to evolve our biology, not just fight off diseases.
KPI
Behind pace: 63.8 yrs (2023); +0.03 yrs/yr excluding the COVID dip vs +0.13 pre-2019 trend · assessed 2026-10-18
Healthy life expectancy in the median country was 63.8 years in 2023 (WHO Global Health Estimates), back above its 2019 level after the COVID-19 dip. Leaving the 2020–21 pandemic years aside, the other three years from 2019 to 2023 average 63.6 years against 63.4 in 2014–18, a gain of about 0.03 years per year — a quarter of the pace before 2019 (0.13 years per year): moving the right way but too slowly, so yellow. The main caveat is that a pace measured from only three recent years is imprecise.
The KPI's precise definition
The KPI, as defined: Median country Healthy Life Expectancy (HALE) at birth (years) — currently ~63y (WHO GHE 2021; ~72y is median life expectancy, not HALE)
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- hale-median-country: unweighted median across 185 countries of WHO GHE HALE at birth, both sexes, years; goal_direction up; fixed five-year window.
- Window
- Current
- Span
- 2019–2023
- Values
- 201963.6202062.5202162.1202263.4202363.8
- N
- 5
- Mean
- 63.08
- OLS slope yr per yr
- 0.13
- Residual SD
- 0.8204
- Scored ex transient
- Span
- 2019, 2022, 2023
- Values
- 201963.6202263.4202363.8
- N
- 3
- Mean
- 63.6
- Slope
- 0.02308
- Residual SD
- 0.2746
- Previous
- Span
- 2014–2018
- Values
- 201463.3201563.4201663.1201763.5201863.8
- N
- 5
- Mean
- 63.42
- Slope
- 0.11
- Residual SD
- 0.2214
- Comparison
- Mean difference ex covid
- 0.18
- Centre year difference
- 5.333
- Between centres rate
- 0.03375
- Descriptive noise SE
- 0.08685
- Descriptive z
- 2.073
- Pace share of benchmark
- 0.265
- Mean difference raw
- -0.34
- Benchmark fit
- Span
- 2000–2018
- N
- 19
- Slope
- 0.1274
- Residual SD
- 0.1189
- Values
- 200061.3200161.5200261.7200361.8200461.9200562200662.2200762.3200862.5200962.6201062.9201162.7201262.9201363.1201463.3201563.4201663.1201763.5201863.8
- Through 2019 sensitivity
- 0.1241
- Source
- uv run core/ledger.py series health hale-median-country --as-of 2026-10-18 (collector, GHO WHOSIS_000002, published 2026-10-01); statistics in notes 20261018T211426Z Appendix B.
- Noise caveat
- The historical OLS residual SD is a descriptive noise scale, not a measurement-error model; the raw window slope (+0.13) includes the pandemic rebound and the raw mean difference (−0.34) the trough; neither substitutes for the ex-transient comparison.
- Benchmark
- Tier 3, own pre-2019 long-run OLS 2000–2018: +0.127 yr/yr (2000–2019: +0.124). Ex-transient centre-to-centre pace +0.034 yr/yr = 26.5% of benchmark. WHO global population-weighted +~0.28 yr/yr 2000–2019 is context only.
- Benchmark source
- WHO Global Health Estimates, GHO WHOSIS_000002 (hale-median-country), fixed KPI spec tier 3.
- Transients discounted
- COVID-19 2020–2021: 63.6 (2019) → 62.5 → 62.1, then 63.4 (2022) and 63.8 (2023); two years < one window, named cause, reversal measured; the trough is omitted from secular scoring and the rebound is not scored as acceleration.
- Data as of
- 2023
- Change
- Yellow unchanged. No new observation of hale-median-country (effective series identical to prev, published 2026-10-01; 2026-10-02-who-ghe-2023-hale-release restates the same 2019/2023 values), so every statistic reproduces prev exactly. Record moves from rubric v4 to v6 (KPI rules unchanged v2–v6, no re-judgment). Label drops '(tier 3)' per §6. New context events (WHS 2026 global HALE 63 in 2023, GBD 2023 morbidity gap, ONS UK healthy life expectancy fall 2022–24, PAHO) do not measure the assessed metric.
- Rubric rule
- §3: yellow — the ex-transient window moves upward beyond descriptive noise (z 2.07) but below the tier-3 own-trend pace (26.5% of it); COVID-19 2020–21 is a transient with its reversal measured, so it moves the status in neither direction. Not stale: data to 2023, 2.8 years before made_at, irregular cadence threshold 3 years.
by assess:20261018T211426Z · 7 evidence event(s)
Evidence and news: 7 cited by the assessment, 10 other KPI news
Cited evidence (7)
- WHO's Global Health Estimates 2000–2023, released 2 October 2026, put global healthy life expectancy (HALE) at birth (WHO global aggregate) at 62.8 years in 2023, still 0.4 years below its 2019 level, while the unweighted median HALE at birth across 185 countries (GHO WHOSIS_000002, both sexes) reached 63.8 years in 2023, above its 2019 value of 63.6 years.
- A GBD 2023 study published in The Lancet Public Health (IHME release, 21 July 2026) found that global healthy life expectancy at birth rose from 55.9 to 63.1 years between 1990 and 2023 while life expectancy rose from 64.6 to 73.8 years, widening the global morbidity gap (years lived in poor health) from 8.8 to 10.7 years.
- WHO's World Health Statistics 2026 reported that global healthy life expectancy (HALE) at birth, after rising from 58 years in 2000 to 63 years in 2019, fell to 61 years in 2021 and recovered to 62 years in 2022 and 63 years in 2023, with female (64 years) and male (62 years) HALE in 2023 still slightly below their 2019 levels.
- The UK Office for National Statistics reported that healthy life expectancy at birth (expected years in 'good' self-rated general health) in the UK for 2022–2024 was 60.7 years for males and 60.9 years for females, down 1.8 and 2.5 years from 2019–2021 and the lowest since the series began in 2011–2013.
- WHO World Health Statistics 2025 reported that global healthy life expectancy (HALE) at birth fell by 1.6 years between 2019 and 2021 during the COVID-19 pandemic, rolling it back to 2012 levels.
- The UK Office for National Statistics (Dec 2025) reported life expectancy at birth for 2022–2024 of 83.0 years for females (recovered to pre-COVID 2019 levels) and 79.1 years for males (still about 10 weeks below pre-COVID levels).↳ corrects (or adds sources to) an earlier record:
2025-12: The UK Office for National Statistics (Dec 2025) reported life expectancy at birth for… - WHO reports global healthy life expectancy (HALE) at birth increased 9%, from 58.1 years in 2000 to 63.5 years in 2019, a 5.3-year increase.↳ corrects (or adds sources to) an earlier record:
2019: According to WHO, global healthy life expectancy (HALE) at birth rose by 5.4 years…
Other events (10)
- PAHO's Health in the Americas 2026 reported that life expectancy in the Region of the Americas reached 77.1 years in 2023, exceeding pre-pandemic levels, but that between 2000 and 2023 healthy life expectancy rose only about 1.5 years versus nearly 4 years for life expectancy.
- Japan's Ministry of Health, Labour and Welfare 2025 abridged life table put life expectancy at birth at 81.35 years for men and 87.33 years for women in 2025, up 0.25 and 0.20 years from 2024.
- China's State Council issued its national health plan for the 15th Five-Year Plan period (2026–2030), targeting an average life expectancy of 80 years by 2030, up from 79.25 years in 2025.
- The UN Population Division's programme plan note to the Commission on Population and Development (E/CN.9/2026/CRP.1, 26 March 2026) states that the release of the World Population Prospects has been postponed from 2026 to 2027, amid liquidity constraints that also prevented the subprogramme from recruiting staff.
- China's National Health Commission head Lei Haichao said on 7 March 2026 that the average life expectancy of Chinese people reached 79.25 years in 2025, an increase of 1.32 years from 2020.
- +5 older event(s): the section's event log
Milestone Countdown
Independent fronts: these can fall in any order.
now20282030203520402050206020752100later
Expected year: dot = median (hollow: after 2100), bar = 80% range. The time axis is compressed: the farther out, the more years per step.
The Long Middle Age: Validated interventions reverse ≥10 years of biological age (epigenetic clocks + functional biomarkers), with corresponding reductions in age-related disease incidence.Distant: best measured reversal ~2.5–4 yrs vs ≥10 plus function and less disease; ETA ~2042 · assessed 2026-10-18 · 41 event(s)
The largest biological-age reversals measured in people are about 2.5–4 years (semaglutide and an experimental lung-fibrosis drug in 2026; earlier small diet and hormone studies), against at least 10 years plus matching functional gains and a demonstrated drop in age-related disease. A chain of clinical stages — first a qualifying intervention, then an outcome study — puts the median forecast at 2042 (10% chance by 2034; 75% by 2100), more than 10 years out, so red. The biggest uncertainty is how soon any intervention reaches a 10-year reversal at all.- ETA
- ~2042 (80% range 2034–never/after 2100; P(by 2100) 75%)
- ETA cruxes
- Crux 1 is the stage-A length (A median 5 vs 12 yr vs the 21.7-yr max-effect envelope: p50 2038 vs 2047 vs 2060), moved by controlled human ≥5-yr multi-clock results with functional benefit or by ER-100/NewLimit reprogramming data. Crux 2 is the reading of 'corresponding reductions in incidence' (inferred from biomarkers: 2035, P .88; proportional: 2046, P .62; the anchor manifold-60n0viotdzfv1f9jmm1q-p resolves on inferred-like endpoints). Then surrogate validity (F1), systemic reprogramming safety (F2 .25: 2045), and outcome designs under 4 years or an accepted surrogate (B 3: 2038).
- Path
- Natural reading (framework.yaml gives no details): one validated intervention reverses ≥10 years on epigenetic clocks and functional biomarkers in humans, with a reduction in age-related disease incidence shown in an outcome study; no fixed reduction size. Alternatives (owner): proportional reduction (2046) and incidence inferred from validated biomarkers (2035). No series measures the quantity, so the ETA rests on a clinical-transition reference class anchored at 2025 first-in-human dosing. Candidate paths: partial reprogramming (now in humans), GLP-1 and other geroprotective drugs, plasma exchange, senolytics, immune rejuvenation.
- Blockers
- Clinical: no human ≥10-year clock-and-function reversal; best effects 2.5–4 years on unlike clocks and short trials.
- Safety/translation: systemic reprogramming safety unknown (ER-100 is local intravitreal dosing, 3 patients); UNITY ceased operations; the NIA ITP found 0 of 11 new regimens extended mouse lifespan.
- Validation: linked incidence evidence needs an outcome trial (assumed ≥4-year design) unless regulators accept an aging surrogate (ARPA-H PROSPR, FDA focus areas pending).
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- Reference class (clinical-transition chain) with a trial-design floor and named failure scenarios whose disclosed probabilities are bracketed by counted classes; unconditional Monte Carlo (40,000 draws, common random numbers); compared after the fact with a similar Manifold anchor
- ETA calculation
- Gap (measured): prev TRIIM 2.5 yr (legacy), Fitzgerald 3.23, Buck 2.61; new 2026: semaglutide DNAm PhenoAge −4.90 yr/yr over 32 weeks ≈ −3.0 yr (2026-05-19-semaglutide-epigenetic-clocks-nature-communications), proteomic organ clocks −1 to −4 yr in 10,052 trial participants (2026-09-22-semaglutide-proteomic-organ-age-five-rcts-preprint), rentosertib −3 to −4 (one −6) at week 4 (2026-09-07-rentosertib-proteomic-aging-clocks-phase-2a). Epigenetic clocks (the wording's family): best 3.23 (2021) and ≈3.0 (2026), 3.1–3.3× short of 10, unchanged; proteomic clocks 3–4 yr, 2.5–3.3× short; unlike studies, no fittable series. Model kept (prev reproduced exactly, .7473/2034.89/2042.33/null): anchor 2025 first human dosing (RTR242); A=max(3,LN(8,.65)) yr to a qualifying ≥10-yr clock+function intervention, B=max(4,LN(6,.30)) yr linked incidence validation (assumed design prior; BP trial 4 yr); T=2025+A+B. Named failures (independent, total .252): F1 surrogate decoupling .15 (counted class: 8 of 9 mouse lifespan tests coupled to clocks, raw .111 to Laplace .182; disclosed judgment inside); F2 safety ceiling .12 (0 of 7 resolved platforms halted by safety, Laplace .111, survivorship-biased low). Result: P .7473, p10 2034.89, p50 2042.33, p90 null; conditional median 2039.60; P(by end-2039) .394. New-evidence cross-checks: XPRIZE Healthspan (10-yr function restoration, grand prize 2030) discounted 1.5–3× → 2031.6–2036.4, bracketing stage A's central 2033; crude max-effect envelope (2.5 epigenetic 2019.9 → 3.5 proteomic 2026.4, 5.18%/yr) needs 20.28 more years: capability 2046.68, i.e. stage A median 21.68 yr from 2025 → p10 2041.73, p50 2060.55, P .7192 (stage A alone by end-2039 .251) (sensitivity, not adopted: it mixes clock families and unlike drugs, and on epigenetic clocks alone the best effect is flat, log-OLS +0.4%/yr, so no trend dates arrival; the plausible route is a step change; partial reprogramming ER-100 first dosed 2026-06-09, safe in 3 at Dose Level 1; ER-100 as stage-A anchor → 2043.33). Sensitivities (p50, P): A 5/12 → 2038.46/2047.64; B 3 no floor/9 → 2038.98/2045.75; F1 .05/.30/.60 → 2041.08 (.837)/2046.06 (.616)/null (.349); F2 .05/.25 → 2041.49 (.804)/2045.21 (.635). Definition: proportional reading 2046.06, P .616; inferred-incidence reading 2035.58, P .883. Pre-anchor p50 2042 (p10 2034, p90 null, P .747) written before reading the anchor. Anchor manifold-60n0viotdzfv1f9jmm1q-p (similar; 0.6012 on 2026-10-15 by 2039-12-31, forecasters=27): ours P(by end-2039) .394, gap .207 < .35, no gap; the market's event is our stage A under stricter study terms without the incidence clause (our stage A alone .733, more optimistic than the market); the difference is our stage B + F1. Crux vs the market: whether 'corresponding reductions in incidence' needs an outcome study; settled by regulator acceptance of an aging surrogate (ARPA-H PROSPR, FDA focus areas) or the owner's reading. Thin, stale market: low weight, no move. Cross-section context: robots-software metr-50-horizon-by-release last 1044.78 min (2026-04-07); no measured mapping to clock-years, no METR multiplier. No shared driver.
- Change
- Red unchanged; P .747, p10 2034, p50 2042, p90 null unchanged (prev reproduced exactly on the same draws). New evidence (ledger top-up, 2026: semaglutide and rentosertib clock effects of ~3–4 yr, ER-100 first human dosing and interim safety, XPRIZE Healthspan finalists, ARPA-H PROSPR, FDA aging focus, ITP 0 of 11, mouse lifespan studies) was tested against each input: none measures stage A, stage B or a failure probability, so none moves. It added proteomic-clock effects of 3–4 yr (2.5–3.3× short; epigenetic gap unchanged at 3.1–3.3×), resolved prev's 'ER-100 dosing status' crux (dosed June 2026), and added the crude max-effect envelope (stage A 21.68 yr: p50 2060.55, P .719) and ER-100 anchor (2043) sensitivities and the XPRIZE discounted-proponent cross-check. Anchor unchanged (0.6012), no gap.
- Rubric rule
- §4 and §4.1: red, because the unconditional median ETA 2042 is more than 10 years after made_at. No verified achievement meets the literal wording.
by assess:20261018T211426Z · 20 evidence event(s)Evidence and news: 20 cited by the assessment, 21 other event(s)
Cited evidence (20)
- On 8 October 2026 Life Biosciences reported interim Phase 1 data showing that a single intravitreal Dose Level 1 of the OSK reprogramming therapy ER-100 with 56 days of oral doxycycline was well tolerated through Day 56 in 3 open-angle glaucoma participants with no dose-limiting toxicities or serious adverse events, with visual-field improvements in 2 of 3 and a recommendation to escalate to Dose Level 2.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Life Biosciences Announces First-in-Human Data from Ongoing Phase 1 Trial Evaluating ER-100 in Optic Neuropathies
- At ARDD 2026 on 1 October 2026 FDA Chief Scientist Steven Kozlowski said the agency plans to include aging and longevity among the topics of its updated Focus Areas of Regulatory Science report due early in fiscal year 2027, while FDA's Jeffrey Siegel described benefit across two or three age-related diseases sharing a mechanism as one possible approval route for aging therapies.
- A Novo Nordisk-coauthored preprint posted on 22 September 2026 analyzing proteomics from 10,052 participants in five placebo-controlled semaglutide trials (up to 156 weeks) reported that semaglutide reduced proteomic heart biological age by about 2-4 years versus placebo and kidney, brain, pancreas, lung and multi-organ clock ages by 1-4 years, while the authors stated this does not establish biological rejuvenation or surrogate validity.
- A Nature Biotechnology analysis published on 7 September 2026 of serum proteomics from 42 idiopathic pulmonary fibrosis patients in Insilico's randomized placebo-controlled 12-week phase 2a trial of rentosertib found all six proteomic aging clocks trending toward lower predicted biological age, which Insilico put at about 3–4 years (up to 6 years on one clock) at week 4 in the 30 mg twice-daily group.The Long Middle AgePredictive Biomarkers [1] Integration of proteomic aging clocks in a phase 2a clinical trial supports simultaneous geroprotective assessment (Nature Biotechnology) [2] Nature Biotechnology | Insilico’s AI-Driven IPF Candidate Rentosertib Shows Potential for Biological Age Reversal, as Assessed by Six Proteomic Aging Clocks
- A Nature Medicine study published on 21 August 2026 harmonized 51 human intervention studies with DNA-methylation data (TranslAGE) and computed 16 epigenetic clocks plus 94 other DNAm biomarkers, finding DunedinPACE showed the largest effect sizes and PCGrimAge the most significant results, with pharmacological interventions producing larger effects than lifestyle, supplement or procedure interventions.The Long Middle AgePredictive Biomarkers [1] Responsiveness of epigenetic aging biomarkers to longevity interventions in humans (Nature Medicine)
- On 11 August 2026 XPRIZE named 20 finalist teams in the $101 million XPRIZE Healthspan, ten of which received $1 million Milestone 2 awards ($10 million in total), to run clinical trials of up to one year in 2026-2029 aiming to restore muscle, cognitive and immune function by at least 10 years in adults aged 50-90, with a grand prize of up to $81 million in 2030.
- On 9 June 2026 Life Biosciences announced that the first participant had been dosed in its Phase 1 trial of ER-100 (NCT07290244), an OSK (OCT4, SOX2, KLF4) epigenetic reprogramming therapy for optic neuropathies including open-angle glaucoma and NAION.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Life Biosciences Announces First Patient Dosed in Phase 1 Trial of ER-100 for Optic Neuropathies
- On 2 June 2026 epigenetic reprogramming company NewLimit closed a $435 million Series C led by Founders Fund and said it plans to bring its first aging reprogramming medicine, a liver therapy, into human clinical trials in 2027.The Long Middle AgeSafe In-Vivo Rejuvenation [1] NewLimit raises $435M led by Founders Fund to bring longevity medicines to human trials
- A post hoc analysis of a 32-week randomized placebo-controlled phase 2b trial in HIV-associated lipohypertrophy, published in Nature Communications on 19 May 2026 (84 participants with clock data), found semaglutide associated with a 0.09 lower DunedinPACE pace of aging (about 9% slower) and a PhenoAge reduction of 4.90 years per year relative to placebo.
- The RAPA-EX-01 randomised double-blind trial in 40 sedentary adults aged 65–85 (Journal of Cachexia, Sarcopenia and Muscle, 15 April 2026) found that once-weekly 6 mg sirolimus (rapamycin) for 13 weeks did not enhance gains from home exercise: chair-stand improvement was 2.13 repetitions lower than placebo in the primary analysis (p = 0.089) and 3.44 lower per protocol (p = 0.007), with more adverse events (99 vs 63) and one possibly drug-related pneumonia.Causal Model of AgingThe Long Middle Age [1] Exercise and Weekly Sirolimus (Rapamycin) in Older Adults: RAPA-EX-01 Randomised, Double-Blind, Placebo-Controlled Trial [2] PubMed abstract 41985884 (RAPA-EX-01) [3] Exercise and Weekly Sirolimus (Rapamycin) in Older Adults: RAPA‐EX‐01 Randomised, Double‐Blind, Placebo‐Controlled Trial (PMC)
- The NIA Interventions Testing Program reported in GeroScience on 17 March 2026 that none of 11 regimens tested in its C2022 cohort of genetically heterogeneous UM-HET3 mice (including astaxanthin, meclizine and mitoglitazone at new doses or start ages, pioglitazone, alpha-ketoglutarate, mifepristone, methotrexate and atorvastatin-telmisartan) increased lifespan in either sex, and that pioglitazone reduced female median lifespan by 6.2%.
- Longeveron's Phase 2b trial of allogeneic bone-marrow MSCs (laromestrocel) in 148 ambulatory adults with aging frailty, published in Cell Stem Cell on 25 February 2026, reported a 63.4 m improvement in 6-minute walk distance versus placebo at month 9 (p=0.0077) and a non-significant 41.3 m at month 6 (p=0.0635).
- On 24 February 2026 ARPA-H announced the seven performer teams of its PROSPR healthspan programme, investing up to $144 million over five years to identify early biomarkers of aging usable as surrogate endpoints and to run the first trials aimed at extending healthspan, including a UT Health San Antonio phase 3 hybrid trial repurposing an SGLT2 inhibitor, rapamycin and semaglutide to establish a regulatory path for aging therapeutics.The Long Middle AgePredictive Biomarkers [1] ARPA-H: Research teams to add more healthy years to Americans' lives as they age
- On 28 January 2026 Life Biosciences announced that the US FDA had cleared its IND for ER-100, a therapy based on controlled expression of the OCT4, SOX2 and KLF4 (OSK) reprogramming factors, allowing a Phase 1 first-in-human trial (NCT07290244) in open-angle glaucoma and NAION, which the company called the first cellular rejuvenation therapy using partial epigenetic reprogramming to reach human clinical trials.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Life Biosciences Announces FDA Clearance of IND Application for ER-100 in Optic Neuropathies
- In 2025, Retro Biosciences dosed the first participant in a clinical trial of its autophagy-focused drug candidate RTR242, becoming a clinical-stage company.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Longevity.Technology: Retro Bio commences first-in-human trial [2] Wayback capture 20260101151419↳ corrects (or adds sources to) an earlier record:
2025-12: In December 2025 Retro Biosciences dosed the first patient in a trial of RTR242, an… - A 2025 Nature Medicine study in 33,995 Chinese adults found intensive blood pressure control reduced dementia risk by 15% (RR 0.85) over 4 years.↳ corrects (or adds sources to) an earlier record:
2025: A 2025 Nature Medicine study in 33,995 Chinese adults found intensive blood pressure… - UNITY Biotechnologies ceased operations following a clinical trial that was not impressive enough for the company to continue operating.The Long Middle AgeSafe In-Vivo RejuvenationBeyond the framework [1] Fight Aging!: A look back at 2025 progress towards the treatment of aging as a medical condition↳ corrects (or adds sources to) an earlier record:
2025-09: In September 2025 UNITY Biotechnology's liquidation was approved after its Phase 2b… - In May 2025 a Buck Institute-led trial reported that biweekly therapeutic plasma exchange with IVIG reduced biological age by an average of 2.61 years.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Buck Institute: Clinical Trial and Multi-omics Analysis Demonstrates the Impact of Therapeutic Plasma Exchange on Biological Age↳ corrects (or adds sources to) an earlier record:
2025-05: In May 2025 a Buck Institute randomized trial of therapeutic plasma exchange plus IVIG… - A 2021 randomized controlled clinical trial found that an 8-week diet and lifestyle treatment group scored 3.23 years younger than controls on the Horvath DNAmAge clock.The Long Middle Age [1] Aging: Potential reversal of epigenetic age using a diet and lifestyle intervention: a pilot randomized clinical trial↳ corrects (or adds sources to) an earlier record:
2021: A 2021 randomized controlled trial (Fitzgerald et al., Aging) of an 8-week diet and… - The 2019 TRIIM trial (rhGH + metformin + DHEA) in 9 participants reported a 2.5-year reversal of biological age across 4 epigenetic clocks.The Long Middle Age [1] Aging Cell 2019: TRIIM thymus regeneration trial
Other events (21)
- A Wyss Institute/Church-lab study published in Nature Aging on 14 October 2026 used a transcriptomic-age screen in old human fibroblasts to identify the splicing factor SRSF1, whose overexpression reversed several aging hallmarks without loss of fibroblast identity or pluripotency-program activation, and whose liver-tropic AAV delivery to aged mice slowed frailty and extended median lifespan by 14% versus controls.Safe In-Vivo RejuvenationThe Long Middle Age [1] Transcriptomic reprogramming screen identifies SRSF1 as a regulator of aging (Nature Aging) [2] Reversing age by reprogramming gene expression networks (Wyss Institute)
- Rubedo Life Sciences reported on 12 October 2026 that four weeks of topical RLS-1496 in its Phase 1b atopic dermatitis study reduced measured skin epigenetic age (MitraBio clock) by a mean 2.1 years across 11 paired photo-aged skin samples (p=0.032), and that final Phase 1b/2a data in 22 actinic keratosis patients showed a mean 49% lesion reduction after 28 days with minimal irritation.
- A double-blind randomized proof-of-principle trial of the senolytics dasatinib plus quercetin in 31 patients with biopsy-proven fibrotic MASH (21 weeks; Nature Metabolism, 1 October 2026) met its primary endpoint, with ≥1-point fibrosis improvement without MASH worsening in 8 of 17 (47%) D+Q participants versus 1 of 14 (7%) on placebo.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Senolytics dasatinib and quercetin in metabolic dysfunction-associated steatohepatitis: a proof-of-principle randomized, controlled trial (Nature Metabolism)
- On 30 September 2026 the KHL Foundation disclosed that a small number of people had received an experimental 11-gene therapy containing the OSK reprogramming factors, developed through a Triple Helix Science initiative, one of them dosed submucosally on 11 March 2026 (before Life Biosciences' 9 June 2026 announcement of the first participant dosed in its ER-100 trial), while stating that this did not prove OSK reverses cognitive decline.The Long Middle AgeSafe In-Vivo Rejuvenation [1] KHL Foundation: Madrid 2026 Press Release [2] Fight Aging!: KHL Foundation Reports the Likely First Human Partial Epigenetic Reprogramming Treatments
- A Nature study published on 2 September 2026 found that daily semaglutide started in 20-month-old female C57BL/6 mice improved physical and cognitive function, attenuated hallmarks of ageing including inflammation, senescence and stem-cell decline, and, continued for life, extended lifespan by about 100 days on average versus vehicle, acting as a calorie-restriction mimetic with some benefits beyond matched 24% calorie restriction.
- +16 older event(s): the section's event log
Precise Screening: >50% of major non-communicable diseases (cancer, cardiovascular, metabolic, neurodegeneration) caught pre-symptomatic.Distant: share caught before symptoms unmeasured, modelled ~23% vs >50%; ETA ~2057 · assessed 2026-10-18 · 41 event(s)
- The share of new diagnoses in a year across the four classes (case-weighted), not an average of the class shares. Each person counts once per class, at the first diagnosis of that class's disease (cancer: once per primary cancer); a small, poorly screened class can be outweighed.
- Caught = the disease itself is diagnosed before its first symptom or event (e.g. atherosclerosis on imaging, silent AF, type 2 diabetes by routine HbA1c, cancer by screening). Risk factors (hypertension, prediabetes, high cholesterol) do not count.
- Judged at the frontier: the whole resident population of at least one country or region of 50M+ (not a selected cohort or one insurer's members) gets there; worldwide diffusion is not required.
No study measures, for any population of 50 million or more, the share of these diseases diagnosed before symptoms; a modelled estimate puts it around 23% against the more than 50% needed. Assumed curves for how fast screening spreads in each disease class cross 50% around 2057 (10% chance by 2041; 94% by 2100), more than 10 years out, so red. The biggest uncertainty is how fast heart disease starts being found before a first event such as a heart attack or stroke.- ETA
- ~2057 (80% range 2041–2091; P(by 2100) 94%)
- ETA cruxes
- Crux 1 is the CVD pre-symptomatic first-diagnosis rate (r_cvd .03 → 2070, .09 → 2052), now watched through AI coronary-calcium uptake on routine CT. Then the frontier baseline shares (a routes-to-diagnosis audit with first-diagnosis counts; UK-like cancer share → 2061), FDA's Galleri decision and uptake under Medicare from 2028 (r_cancer .04–.12: 2054–2065), and screening retrenchment (USPSTF).
- Path
- Natural reading: share of first diagnoses per person per class (cancer per primary), case-weighted across four classes, made before the first symptom or event, in one whole 50M+ population; risk scores, hypertension, prediabetes and cholesterol excluded. Path: organised and MCED cancer screening, routine HbA1c, imaging (incl. AI coronary calcium on routine CT) and ECG detection of silent CVD/AF, blood biomarkers for preclinical Alzheimer's. The model is an assumed diffusion prior, not a fit to the target.
- Blockers
- Measurement: no population routes-to-diagnosis series across the four classes (only the UK cancer screening share, 6–7%); gather first-diagnosis counts with pre-symptom flags.
- Regulatory/evidence: NHS-Galleri showed no stage III–IV reduction; FDA's Galleri decision pending; Medicare MCED coverage starts only in 2028.
- Deployment: CVD and neuro classes have no population screening programme; p-tau217 tests are cleared for symptomatic people only.
- Institutional: USPSTF disruption (leaders dismissed May 2026; no meeting since March 2025).
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- S-curve (per-class logistic diffusion, assumed reference-class priors) with named failure scenarios; unconditional Monte Carlo (40,000 draws, common random numbers)
- ETA calculation
- Gap: literal share unmeasured. ASSUMED baseline weights cancer/CVD/metabolic/neuro .30/.40/.22/.08 × pre-symptomatic shares .20/.10/.60/.01 = .2328 vs >.50 (2.148×). s_i(t)=K_i/[1+(K_i/s0_i−1)e^{−r_i(t−2025)}], K .70/.65/.90/.50, r .08/.06/.04/.09; central path .27145 (2030), .35842 (2040), .44738 (2050), .52682 (2060), crossing 2056.43. Draws as prev (Dirichlet(40w); s0 U[.10,.30]/[.04,.20]/[.40,.80]/[.002,.03]; K U[.50,.85]/[.40,.80]/[.75,.98]/[.15,.70]; rates ×LN(1,.45) common ×LN(1,.20) per class). Named failures: F1 ceiling <50% .0015; F2 diffusion too slow (retrenchment as after the 2012 PSA downgrade; USPSTF disruption 2025–26) .062; both <10%. Result (prev model reproduced exactly): P .9365, p10 2041.54, p50 2057.91, p90 2091.09; conditional median 2056.53; calendar-year floors. New evidence tested against inputs: UK screen-detected cancers 6–7% (2026-04-23-cruk-uk-screen-detected-cancer-share-6-7pct) is a non-frontier population at the low end of the cancer s0 range (sensitivity U[.06,.14] → 2061.92, P .913); NHS-Galleri IRR 1.03 for stage III–IV (2026-09-22-nejm-nhs-galleri-primary-results) vs FDA panel 7–2 and Medicare MCED coverage from 2028 (r_cancer .04 → 2065.97, .871; .12 → 2054.34, .951); incidental/AI coronary calcium made a class-1 input and paid by CMS (r_cvd .09 → 2052.10, .971; .03 → 2070.84, .797). None measures a rate, so no input moves. Other sensitivities (prev, same draws): all s0 low/high 2076.16/2044.55; all rates ×0.5/×1.5 2090.82 (.597)/2046.94; K_cvd .45/.80 2062.44/2054.31; CVD weight ×1.5/×0.67 2061.53/2054.80. Cross-section context: robots-software metr-50-horizon-by-release last 1044.78 min (2026-04-07); no measured mapping, no METR multiplier. No shared driver, no anchor.
- Change
- Red unchanged; model and inputs unchanged (prev reproduced exactly: .9365/2041.54/2057.91/2091.09), so no quantile or P moves. New evidence (ledger top-up: MCED trial results and regulatory steps, UK screen-detected share, CAC guideline and payment, USPSTF disruption) measures no model input; it added sensitivities (UK-like cancer share, cancer and CVD rates) and updated the blockers.
- Rubric rule
- §4 and §4.1: red, because the unconditional median ETA 2057 is more than 10 years after made_at. No verified achievement meets the literal wording.
by assess:20261018T211426Z · 19 evidence event(s)Evidence and news: 19 cited by the assessment, 22 other event(s)
Cited evidence (19)
- On 23 September 2026 the FDA's Molecular and Clinical Genetics advisory panel, reviewing the first PMA for a multi-cancer early detection test, voted 7–2 (one abstention) that the benefits of GRAIL's Galleri outweigh its risks for screening adults 50 and older, unanimously on safety and 6–4 on effectiveness; FDA's final decision was still pending.
- The peer-reviewed PATHFINDER 2 results, published in Nature Medicine on 22 September 2026, reported that among 32,007 performance-analyzable adults aged 50+ without suspected cancer, GRAIL's Galleri test had a 0.54% cancer detection rate, 60.3% positive predictive value, 99.64% specificity, 91.3% cancer-signal-origin accuracy and episode sensitivity of 39.3% for all cancers and 69.8% for 12 prespecified deadly cancers, with 0.6% of 35,335 participants undergoing invasive procedures.Predictive BiomarkersPrecise Screening [1] Performance and safety of a multi-cancer early detection test: the PATHFINDER 2 study (Nature Medicine)
- The peer-reviewed NHS-Galleri primary results, published in NEJM on 22 September 2026 (142,250 randomised participants), confirmed no difference in stage III–IV incidence of 12 prespecified cancers after three screening rounds (incidence rate ratio 1.03, 95% CI 0.92–1.14) and a stage IV rate ratio of 0.86 (95% CI 0.74–1.00), with a companion Nature Medicine paper reporting per-round test specificity of about 99.5–99.6%.Precise ScreeningPredictive Biomarkers [1] NEJM: Effect of Screening with Multicancer Early-Detection Test on Late-Stage Cancer Diagnosis [2] PubMed: Effect of Screening with Multicancer Early-Detection Test on Late-Stage Cancer Diagnosis (N Engl J Med, Epub 2026 Sep 22) [3] Nature Medicine: Performance of a multi-cancer early detection test in the randomized controlled NHS-Galleri trial
- On 17 September 2026 HHS appointed eight new members to the US Preventive Services Task Force, joining the eight remaining members of a panel that had not met since March 2025.
- An international consensus statement published in Diabetologia on 10 September 2026 and endorsed by multiple diabetes professional associations set out principles and minimum requirements for general-population screening for islet autoantibodies to diagnose early-stage (presymptomatic) type 1 diabetes, noting that 20–67% of children and adolescents currently present with diabetic ketoacidosis at clinical diagnosis.Predictive BiomarkersPrecise ScreeningProgrammable Immunity [1] International consensus guidance for general population screening for islet autoantibodies to diagnose early-stage type 1 diabetes (Diabetologia) [2] PubMed: International consensus guidance for general population screening for islet autoantibodies to diagnose early-stage type 1 diabetes: a nominal group technique process (Diabetologia abstract)
- On 24 August 2026 the FDA cleared Roche's Elecsys pTau217 blood test, developed with Lilly, as the first single-biomarker blood test supporting both rule-in and rule-out of Alzheimer's amyloid pathology across primary and specialty care, for people 55 and older with signs, symptoms or complaints of cognitive decline (not asymptomatic screening).Precise ScreeningUnbroken MindPredictive Biomarkers [1] Roche receives FDA clearance for Elecsys pTau217
- Full NHS-Galleri results presented at ASCO on 30 May 2026 showed no reduction in combined stage III–IV cancers, but stage IV diagnoses of 12 prespecified cancers fell 9%, 22% and 26% in screening rounds one to three (14% overall), stage I–II diagnoses rose 16%, and the test had 52.0% PPV, 99.55% specificity and 54.7% episode sensitivity for the 12 cancers (30.7% for all cancers).Precise ScreeningPredictive Biomarkers [1] GRAIL Reports Full Results From NHS-Galleri Trial ... at 2026 ASCO Annual Meeting [2] NHS-Galleri: trial reports fewer diagnoses of the most advanced cancers
- HHS Secretary Robert F. Kennedy Jr. dismissed US Preventive Services Task Force leaders John Wong and Esa Davis effective 11 May 2026, escalating disruption of the body whose A/B screening grades set no-cost-sharing coverage and which had not met since March 2025.
- Cancer Research UK's Cancer in the UK 2026 report (23 April 2026) stated that roughly 6% to 7% of UK cancer patients are diagnosed through screening, and estimated England could raise this to around 9% (about 11,000 extra screen-detected cancers) by expanding and optimising its breast, bowel, cervical and lung screening programmes.Precise ScreeningTaming Cancer [1] Cancer Research UK: Cancer in the UK — what's needed as cases exceed 400,000 a year
- Bunkerhill Health announced on 15 April 2026 that CMS had established a new national billing code and payment under the Hospital Outpatient Prospective Payment System, effective 1 April 2026, for algorithmic analysis of coronary artery calcium and aortic valve calcium on chest CT scans, and that FDA had cleared its Contrast CAC and Contrast AVC algorithms, the first AI algorithms cleared to detect and quantify these on contrast-enhanced, routine non-gated chest CT.
- The 2026 ACC/AHA multisociety dyslipidemia guideline (13 March 2026) recommends selective coronary artery calcium scoring to detect subclinical atherosclerosis in borderline/intermediate-risk adults (initiating lipid-lowering therapy if CAC >0) and, as a class 1 recommendation, that incidental CAC found on noncardiac CT, including by validated AI algorithms, inform treatment decisions.
- CDC's National Diabetes Statistics Report (updated 11 March 2026) estimated that 27.6% of US adults with diabetes, about 11.0 million people, were undiagnosed (NHANES 2021–2023 data).Precise Screening [1] CDC: National Diabetes Statistics Report
- On 19 February 2026 GRAIL announced topline results of the 142,000-participant randomised NHS-Galleri trial: annual Galleri multi-cancer blood testing did not achieve the primary endpoint of a statistically significant reduction in stage III–IV diagnoses of 12 prespecified deadly cancers, although stage IV diagnoses fell by more than 20% in the second and third screening rounds and the overall cancer detection rate was four-fold that of England's standard screening alone.Precise ScreeningPredictive Biomarkers [1] GRAIL: Landmark NHS-Galleri Trial Demonstrates a Substantial Reduction in Stage IV Cancer Diagnoses ... (topline results)
- England's National Cancer Plan (4 February 2026) premised its survival ambition on raising the share of cancers diagnosed at stage 1–2 by at least 20 percentage points above the 2019 level by 2035, committed to completing national lung cancer screening rollout by 2030 and increasing bowel screening sensitivity, and pledged an implementation plan to offer multi-cancer early detection tests at scale if the UK NSC judges them effective.
- On 3 February 2026 the Nancy Gardner Sewell Medicare Multi-Cancer Early Detection Screening Coverage Act was signed into law as part of the FY2026 spending package, establishing Medicare coverage of FDA-approved multi-cancer early detection screening tests (one per beneficiary per year) beginning in 2028, with payment at parity with multi-target stool DNA tests through 2030.
- The NHS-Galleri trial is a randomised controlled trial of 140,000 participants in the NHS in England.Precise ScreeningPredictive Biomarkers [1] NHS NDRS: GRAIL - NDRS [2] Wayback capture 20251212170056↳ corrects (or adds sources to) an earlier record:
2025: The NHS-Galleri trial of multi-cancer early detection enrolled 140,000 participants,… - GRAIL reported that PATHFINDER 2 found adding Galleri to USPSTF A and B recommended screenings increased cancer detection more than seven-fold, 53.5% of new cancers detected were stage I or II, and episode sensitivity was 73.7% for the 12 cancers responsible for two-thirds of U.S. cancer deaths.Precise ScreeningPredictive Biomarkers [1] GRAIL: PATHFINDER 2 Results Show Galleri Multi-Cancer Early Detection Blood Test Increased Cancer Detection More Than Seven-Fold When Added to USPSTF A and B Recommended Screenings [2] Wayback capture 20251018165916↳ follows up on 2025-06-18: On 18 June 2025 GRAIL announced positive top-line PATHFINDER 2 results; PATHFINDER 2… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-10-17: On 17 October 2025 GRAIL presented full PATHFINDER 2 results at ESMO: adding the… - On 2 June 2025 Clairity Breast received FDA De Novo authorization as the first AI-powered platform predicting five-year breast cancer risk from a standard mammogram.Precise ScreeningPredictive Biomarkers [1] BCRF: A Breakthrough in Breast Cancer Prevention: FDA Clears First AI Tool to Predict Risk Using a Mammogram↳ corrects (or adds sources to) an earlier record:
2025-06-02: On 2 June 2025 Clairity Breast received FDA De Novo authorization as the first… - In May 2025 the FDA cleared the first blood test for diagnosing Alzheimer's disease, Lumipulse pTau217/Aβ1-42, with 91.7% sensitivity and 97.3% specificity.Precise ScreeningUnbroken MindPredictive Biomarkers [1] FDA clears first blood test used in diagnosing Alzheimer's disease↳ corrects (or adds sources to) an earlier record:
2025-05: In May 2025 the FDA cleared the first blood test for diagnosing Alzheimer's disease,…
Other events (22)
- On 9 October 2026 Natera announced that the FDA had granted Breakthrough Device designation to its cell-free-DNA multi-cancer early detection assay after assessing data from 12 cancer types; the test is not FDA approved.Predictive BiomarkersPrecise Screening [1] Natera: FDA Grants Breakthrough Device Designation for Natera's Multi-Cancer Early Detection Assay (9 Oct 2026)
- IARC Handbooks of Cancer Prevention Volume 21 (special report in NEJM, 24 September 2026) concluded that low-dose CT screening of high-risk individuals reduces lung cancer mortality and the incidence of late-stage lung cancer.Precise ScreeningTaming Cancer [1] IARC: IARC Handbooks of Cancer Prevention Volume 21 — the IARC perspective on lung cancer screening
- On 16 September 2026 the US House Ways and Means Committee unanimously advanced the bipartisan Alzheimer's Screening and Prevention (ASAP) Act, which would remove the legal barrier keeping Medicare from covering early Alzheimer's detection tests (committee passage, not enactment).
- On 20 August 2026 the FDA cleared C2N Diagnostics' mass-spectrometry PrecivityAD2 blood test (Aβ42/40 plus %p-tau217) to aid identification of Alzheimer's amyloid pathology in adults aged 40 and older with cognitive symptoms; in a 1,142-person validation study its dual cutoffs gave a 97.6% positive and 93.1% negative predictive value, while the 17.3% of results in a third 'likely positive' category had a 77.3% positive predictive value.Predictive BiomarkersUnbroken MindPrecise Screening [1] C2N: FDA Clears C2N Diagnostics' PrecivityAD2 — First Alzheimer's Blood Test for Adults with Cognitive Symptoms as Young as 40 (20 Aug 2026)
- On 27 July 2026 the FDA approved Freenome's SimpleScreen CRC blood test as a colorectal cancer screening option for average-risk adults 45 and older; in a Census-adjusted prespecified analysis of the more-than-48,000-participant PREEMPT CRC study it detected colorectal cancer with 81.1% sensitivity but advanced precancerous lesions with 13.7% sensitivity, with 90.4% specificity for advanced colorectal neoplasia.Precise ScreeningPredictive Biomarkers [1] FDA Approves Freenome's SimpleScreen CRC Blood-Based Screening Test; Abbott to Commercialize in the U.S.
- +17 older event(s): the section's event log
Taming Cancer: 80% of all solid tumor cases reach 10-year survival — metastatic cancer becomes a chronic, manageable condition.Distant: US 70.5% 5-yr survival (2016–22) vs 80% at 10 yrs; 10-yr non-survival −1.5%/yr, ETA ~2073 · assessed 2026-10-18 · 76 event(s)
- Survival is relative (net) survival, the standard cancer statistic. The metastatic clause explains what the 80% means; it is not a separate test.
- Judged at the frontier: the whole resident population of at least one country or region of 50M+ (not a selected cohort or one insurer's members) gets there; worldwide diffusion is not required.
The newest US figure is 70.5% five-year relative survival for all cancers diagnosed in 2016–2022 (NCI SEER registries), about the same for solid tumours alone and an estimated 63% at ten years, against a target of 80% at ten years. The share dying within five years of diagnosis has been falling about 1.7% a year, and in England and Wales the share dying within ten years has fallen only a little more slowly; at that slightly slower pace, plus the ten years of follow-up needed to confirm ten-year survival, the median forecast is 2073 (10% chance by 2056; 85% by 2100), more than ten years out, so red. Updating to this newer US figure moved the forecast only slightly; South Korea already reports 73.7% five-year survival, but its population is projected to fall below the 50-million bar around 2040, before it could qualify. The main uncertainty is how US ten-year survival tracks five-year survival.- ETA
- ~2073 (80% range 2056–never/after 2100; P(by 2100) 85%)
- ETA cruxes
- Crux 1 is the 5→10-yr mapping: ε .887 measured in England & Wales, US transfer assumed; constant q gives p50 2092 (P .58), ε 1.0 gives 2068; gather US 10-yr relative survival by diagnosis year. Then the 5-yr rate (1.69%/yr vs long-run 1.0–1.2%: 2073 vs 2090–null), conditional q (2066–2080), the definition of '50M+' for shrinking countries (Korea counted as 50M+ today: 2069; owner's call), publication lag (0: 2070), and hard-site plateau signals (F1 .31: P .72).
- Path
- Reading from framework.yaml and its details: in a whole 50M+ population, ≥80% of solid-tumour cases are alive at 10 years in relative (net) survival; the metastatic clause explains the level. Proxy: US all-cancer 5-yr relative survival (newest SEER release) mapped to solid tumours, with an assumed conditional 5→10 factor and an England & Wales-measured elasticity carrying the 5-yr trend to the 10-yr endpoint; it tracks the target because solid tumours are ~91% of cases and 10-yr and 5-yr non-survival move together (ε .887). The population must be 50M+ when its qualifying cohort is diagnosed (Korea, 51.7M today, falls below ~2040). Path: targeted agents (RAS inhibitors), ADCs, immunotherapy, mRNA neoantigen therapy, earlier detection and lung-incidence decline.
- Blockers
- Measurement: no all-solid-tumour 10-year or period-survival series in the ledger; register SEER 5- and 10-yr relative survival by diagnosis year, Korea's 10-yr survival, and the England & Wales Cancer Survival Index.
- Structural: 10-year follow-up after the qualifying cohort (literal floor) plus an assumed ~2.9-year publication lag.
- Clinical: metastatic and hard-site survival remains low (distant stage 35% at 5 years; pancreas 13%), which could plateau the aggregate.
- Funding: the proposed −37.3% NCI cut was not enacted (FY2026 NCI $7.35B, +$128M); disparities persist.
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- Trend (exponential decline of residual non-survival; sourced SEER 5-yr rate mapped to the 10-yr endpoint by an elasticity measured in England & Wales) with structural floor (10-yr follow-up) plus assumed publication lag and named plateau and slow-decline scenarios; unconditional Monte Carlo (40,000 draws, common random numbers)
- ETA calculation
- Anchor (measured; replaces prev's older central anchor, rule 7 (b)): 2026-04-22-seer-april-2026-release-five-year-survival, US 5-yr relative survival 70.5% all sites, 2016–2022 diagnoses (midpoint 2019.0), ±.05 pp rounding band (prev's rule: the rounding width of the published figure; ±.5 pp band → p10 2055.98, p50 2073.09, P .8467); check 2026-01-13-acs-cancer-statistics-2026-five-year-survival-70pct 70% (2015–21). Solid-only (sourced, Siegel 2025: heme 9.41% at 69.8%): (.705−.0941×.698)/.9059 = .7057. Assumed q=S10/S5 ~U(.85,.95): S10 ≈ .635, 10-yr non-survival .365 → .20 = 1.824× (ln .6012). Rate (sourced SEER joinpoint 2012–23): 5-yr non-survival −1.686%/yr, r ~ LN(.0169,.38) (long-run 1.04% near the slow decile; spread assumed); ε = dlnNS10/dlnNS5 measured .887 in England & Wales (Coleman 2025), US transfer assumed, N(.887,.075). T = 2019 + ln[(1−S5q)/.2]/(εr) + 10 (literal floor) + 2.9 (assumed lag); central .6012/(.887×.0169) = 40.10 → 2059.1 + 12.9 = 2072.0. Named failures (same draws): F1 aggregate plateau drawn .10, realised .062 — counted class US 1975–2020 and E&W 1971–2018: 0 of 2 plateaued, 0 decade stalls in ~9.2 series-decades, Laplace .089/decade → ≥1 in 4 decades .31 upper bound; permanence not countable, .10 disclosed judgment; F2 slow decline realised .091 (<10%; ≥40-yr national spans 0 of 2 below the ~1.03%/yr threshold). Result: P .8471, p10 2056.004, p50 2073.11, p90 null (2115.7); conditional median 2069.80; calendar-year floors. Prev reproduced exactly first (.8329/2055.70/2073.93). Sensitivities (p50, P): constant q 2092.99, .579; ε .77 2079.45; ε 1.0 2068.10, .894; rate 1.04% null, .496; 1.20% 2090.93; 2.43% 2060.18, .941; q .85/.95 2080.01/2066.10; lag 0/4 2070.21/2074.21; F1 0/.31 → P .909/.720. Frontier: Korea 73.7% all cancers 2019–23 (2026-01-20-korea-2023-cancer-registry-five-year-survival; 69.6% excl. thyroid) at the US rate → 2069.34 (P .884), but Statistics Korea's 2023 projection puts Korea at 50.06M in 2040 and 42.30M in 2060, so with its qualifying cohort required while Korea is 50M+ (cohort ≤2040) it adds nothing (best-of US/Korea 2073.10, P .847); at Korea's own 3.08%/yr (thyroid-inflated; the US-rate runs assume rate transfer) 2053 unlimited, null with the population limit (P .494), best-of with the US per draw 2071.60, P .865, p10 2045.13. Cross-checks: M/I mapping ~2069; distant-stage 5-yr non-survival 83→65% ≈ 1.1%/yr; England's National Cancer Plan (75% 'living well' 5 yr after a 2035 diagnosis) is a different endpoint, low weight. Cross-section context: robots-software metr-50-horizon-by-release last 1044.78 min (2026-04-07), no measured mapping to survival, no METR multiplier. No shared driver and no anchor.
- Change
- Red unchanged. Prev reproduced exactly (.8329/2055.70/2073.93/null). One input changed, rule 7 (b) error in prev: prev already knew SEER's 70.5% for 2016–22 (a sensitivity, executed at the wrong cohort date 2017) yet kept the older ACS 69% cohort (2014–20, 2017.0) as central, contrary to §4.1 step 3 (most current comparable series of the outcome); its ledgering as 2026-04-22-seer-april-2026-release-five-year-survival verifies it but is not new information. Anchor → 70.5% at 2019.0 with prev's own band rule (rounding width of the published figure: ±.5 pp for an integer, so ±.05 pp for one decimal). Effects: level+date at a ±.5 band P .833 → .8467, p10 2055.70 → 2055.98, p50 2073.93 → 2073.09; band ±.5 → ±.05 pp p10 → 2056.004, p50 2073.11, P .8471. The calendar p10 move 2055 → 2056 is a knife-edge (.004 yr past the year boundary, .02 yr of it from the band); p50 2073 and p90 null unchanged. Other new evidence (Korea 73.7%, 2026 approvals and setbacks, NCI funding enacted, England's plan) informs sensitivities, the 50M+ definition crux and blockers only. Every other input unchanged.
- Rubric rule
- §4 and §4.1: red, because the unconditional median ETA 2073 is more than 10 years after made_at. No verified achievement meets the literal wording.
by assess:20261018T211426Z · 19 evidence event(s)Evidence and news: 19 cited by the assessment, 57 other event(s)
Cited evidence (19)
- BioNTech announced on 13 October 2026 that the phase 2/3 AHEAD-MERIT trial of its mRNA cancer immunotherapy BNT113 plus pembrolizumab in HPV16-positive metastatic head and neck cancer missed its first dual primary endpoint of progression-free survival versus pembrolizumab alone, with the trial continuing to its overall-survival endpoint.Taming CancerProgrammable Immunity [1] BioNTech — Provides Update on Phase 2/3 Trial for BNT113 in HPV16-positive Head and Neck Cancer (13 Oct 2026)
- On 8 October 2026 the FDA approved atezolizumab with fluoropyrimidine and oxaliplatin chemotherapy as adjuvant treatment of stage III mismatch repair-deficient colon cancer, based on ATOMIC, where adding atezolizumab to mFOLFOX6 improved disease-free survival versus mFOLFOX6 alone with a hazard ratio of 0.50 (3-year DFS 86.3% vs 76.2% in NEJM).Taming CancerProgrammable Immunity [1] FDA approves atezolizumab in combination with chemotherapy for Stage III mismatch repair deficient colon cancer [2] Atezolizumab plus FOLFOX for Stage III Mismatch Repair–Deficient Colon Cancer — NEJM 2026 [3] Atezolizumab plus FOLFOX for Stage III Mismatch Repair-Deficient Colon Cancer — N Engl J Med 2026;394(12):1155-1166 (PubMed)
- BioNTech announced on 28 August 2026 that its phase 2 trial of the individualized mRNA cancer vaccine autogene cevumeran in resected ctDNA-positive colorectal cancer was terminated on the data monitoring committee's recommendation after it found a numerical imbalance in overall survival between arms and judged further continuation unlikely to change the efficacy outcome.Taming CancerProgrammable Immunity [1] BioNTech — Provides Update on Phase 2 Clinical Trial of Autogene Cevumeran in Resected Colorectal Cancer (28 Aug 2026)
- On 26 August 2026 the FDA approved daraxonrasib (Rasonque), an inhibitor of the RAS GTPase family, for previously treated metastatic pancreatic adenocarcinoma, based on RASolute 302, where median overall survival was 13.2 versus 6.7 months with standard chemotherapy (HR 0.40).
- Merck and Moderna announced on 19 August 2026 that the phase 3 INTerpath-001 trial in 1,137 patients with resected stage IIB–IV melanoma met its primary endpoint of recurrence-free survival and key secondary endpoint of distant metastasis-free survival for the individualized mRNA neoantigen therapy intismeran autogene plus pembrolizumab versus pembrolizumab alone, the first positive phase 3 for an mRNA-based cancer therapy (effect sizes not yet disclosed; overall survival still being evaluated).Taming CancerProgrammable Immunity [1] Merck/Moderna — Phase 3 INTerpath-001 trial of intismeran autogene plus KEYTRUDA met endpoints of RFS and DMFS
- On 22 June 2026 China's NMPA approved CARsgen's satricabtagene autoleucel (satri-cel), an autologous Claudin18.2-targeted CAR T-cell therapy, for Claudin18.2-positive, HER2-negative advanced gastric or gastroesophageal junction adenocarcinoma after at least two prior lines of therapy, the world's first approved CAR-T therapy for a solid tumor.Taming CancerProgrammable Immunity [1] CARsgen — Announces Approval of Satri-cel, the World's First CAR T-Cell Therapy Product for Solid Tumors (22 June 2026)
- With its 22 April 2026 data release (incidence through 2023, US mortality through 2024), NCI's SEER Cancer Stat Facts put US 5-year relative survival for cancer of any site at 70.5% for patients diagnosed in 2016–2022.
- England's National Cancer Plan (DHSC and NHS England, 4 February 2026) set the goal that three-quarters of people diagnosed with cancer in 2035 are cancer-free or living well with cancer after 5 years, up from a projected 60% in 2022, and aims to save 320,000 more lives by 2035.
- The US Consolidated Appropriations Act, 2026, signed on 3 February 2026, gave the National Cancer Institute $7.35 billion for FY2026, $128 million more than its FY2025 enacted level, instead of the 37.3% cut the administration had requested.Taming CancerBeyond the framework [1] NCI — Budget and Appropriations (FY26) [2] KFF — FY 2026 Labor, Health and Human Services, Education, and Related Agencies (Labor HHS) Global Health Funding in the Consolidated Appropriations Act (updated: signed 3 Feb 2026) [3] OncLive — Federal Funding Cuts Hit Cancer Research Despite Public Support (31 Oct 2025)
- Korea's Ministry of Health and Welfare and Central Cancer Registry reported (January 2026) a nationwide 5-year relative survival rate of 73.7% for all cancers combined among patients diagnosed in 2019–2023 (69.6% excluding thyroid cancer, per the accompanying registry paper), up from 54.2% for 2001–2005, with 92.7% for localized-stage and 27.8% for distant-stage cancers.Taming Cancer [1] MOHW (korea.net) — Korea Records 2.73 Million Cancer Survivors, with 7 in 10 Living More Than Five Years [2] Park et al., Cancer Statistics in Korea: Incidence, Mortality, Survival, and Prevalence in 2023 — Cancer Res Treat 2026 [3] Korea Biomedical Review — Korea's cancer patient count hits 2.73 million (20 Jan 2026)
- The American Cancer Society's Cancer Statistics, 2026 (13 January 2026) reported that US 5-year relative survival for all cancers combined reached 70% for people diagnosed during 2015–2021, that 5-year survival for distant-stage disease (all cancers combined) doubled from 17% in the mid-1990s to 35%, and that the cancer mortality rate fell 34% from its 1991 peak through 2023, averting 4.8 million deaths.
- SEER cancer statistics give 5-year survival for stage IV (metastatic) disease of 31% for breast cancer, 8.2% for lung cancer and 15.6% for colorectal cancer in the US.Taming Cancer [1] SEER Cancer Stat Facts
- In a phase 2b trial, personalized mRNA-4157 (V940) plus pembrolizumab reduced post-resection melanoma recurrence by 44% and distant metastasis risk by 65% compared with anti-PD-1 alone.Programmable ImmunityTaming Cancer [1] Cromos Pharma: Cancer Vaccines 2025: The Rise of mRNA Therapies [2] Wayback capture 20260122202650↳ corrects (or adds sources to) an earlier record:
2025: The personalized mRNA cancer vaccine mRNA-4157/V940 plus pembrolizumab for melanoma,… - Black patients' 5-year overall survival was 1% to 2% lower than White patients' for localized-stage cancers after stratifying by urbanicity, and 7% lower for regional-stage cancers in large metropolitan areas.Taming Cancer [1] CancerNetwork: Five-Year Cancer Survival Rates Lower Among Rural Cancer Populations [2] Wayback capture 20251018173020↳ corrects (or adds sources to) an earlier record:
2025: Black cancer patients in the US have 5-year survival rates 4–7% lower than White… - ACS Cancer Facts & Figures 2025 reports an 86% 10-year survival rate for invasive breast cancer in the US.Taming Cancer [1] American Cancer Society: Cancer Facts & Figures 2025↳ corrects (or adds sources to) an earlier record:
2025: ACS Cancer Facts & Figures 2025 reports an 86% 10-year survival rate for invasive… - ACS Cancer Facts & Figures 2025 reports a 5-year relative survival rate of 13% for all stages of pancreatic cancer.Taming Cancer [1] American Cancer Society: Cancer Facts & Figures 2025↳ corrects (or adds sources to) an earlier record:
2025: ACS Cancer Facts & Figures 2025 reports a 5-year survival rate of 8% for pancreatic… - ACS Cancer Facts & Figures 2025 reports a 69% 5-year relative survival rate for all cancers combined in the US.Taming Cancer [1] American Cancer Society: Cancer Facts & Figures 2025↳ corrects (or adds sources to) an earlier record:
2025: ACS Cancer Facts & Figures 2025 reports a 69% 5-year relative survival rate for all… - The US FY2026 budget request proposed a 37.3% decrease in the National Cancer Institute budget from FY2025.↳ follows up on 2025-03: Federal funding for cancer research through March 2025 was 31% lower than in the same… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025: The US FY2026 budget request for NIH cancer research funding is 37% below the FY2025… - On 27 April 2025, a trial of neoadjuvant PD-1 blockade in early-stage mismatch repair-deficient solid tumors found that 84 of 103 patients who completed treatment had a clinical complete response and 82 did not undergo surgery.Taming CancerProgrammable Immunity [1] NEJM: Nonoperative Management of Mismatch Repair–Deficient Tumors↳ corrects (or adds sources to) an earlier record:
2025-04-27: On 27 April 2025 MSK researchers reported in NEJM that immunotherapy alone produced…
Other events (57)
- IARC Handbooks of Cancer Prevention Volume 21 (special report in NEJM, 24 September 2026) concluded that low-dose CT screening of high-risk individuals reduces lung cancer mortality and the incidence of late-stage lung cancer.Precise ScreeningTaming Cancer [1] IARC: IARC Handbooks of Cancer Prevention Volume 21 — the IARC perspective on lung cancer screening
- At the IASLC World Conference on Lung Cancer on 13 September 2026, the China-only phase 3 HARMONi-2 trial in 398 patients with treatment-naive PD-L1-positive advanced NSCLC reported that ivonescimab monotherapy improved median overall survival to 30.8 versus 22.6 months with pembrolizumab (HR 0.73).Taming CancerProgrammable Immunity [1] IASLC — Late-Breaking HARMONi-2 Analysis Shows Ivonescimab Significantly Improves Overall Survival Versus Pembrolizumab
- On 4 September 2026 the FDA granted accelerated approval to camizestrant (Etcamah) with a CDK4/6 inhibitor for HR-positive, HER2-negative advanced breast cancer upon detection of an ESR1 mutation in blood ctDNA during aromatase inhibitor plus CDK4/6 inhibitor therapy, based on SERENA-6, where switching gave median progression-free survival of 16 versus 9.2 months (HR 0.44).Taming CancerPredictive Biomarkers [1] FDA grants accelerated approval to camizestrant with a CDK4/6 inhibitor for ESR1-mutated HR-positive, HER2-negative breast cancer
- On 25 August 2026 the FDA approved first-line zanidatamab plus tislelizumab and chemotherapy for HER2-positive advanced gastric, gastroesophageal junction or esophageal adenocarcinoma, based on HERIZON-GEA-01, where median overall survival was 26.4 versus 19.2 months with trastuzumab plus chemotherapy (HR 0.72).
- A Nature Medicine phase 1 trial published 6 August 2026 of intracranially infused autologous B7-H3-targeting CAR-T cells (TX103) in 15 patients with recurrent glioblastoma found no dose-limiting toxicities, a 12-month overall survival rate of 66.7% and median overall survival of 19.1 months from first infusion (uncontrolled).Programmable ImmunityTaming Cancer [1] Intracranial delivery of B7-H3-targeting CAR-T cells for recurrent glioblastoma: a phase 1 trial [2] PubMed 42562965
- +52 older event(s): the section's event log
Unbroken Mind: Dementia prevalence (CDR ≥1) after age 80 falls below 5%.Distant: 80+ dementia ~26% (US 2002) vs <5%; falling ~2.3%/yr among 65+, ETA ~2079 · assessed 2026-10-18 · 45 event(s)
- Prevalence among all residents aged 80+ of a country or region of 50M+, age-standardized within 80+ (WHO standard population) so that an older 80+ mix does not decide it; worldwide diffusion is not required.
- CDR ≥1 or a clinical-consensus dementia diagnosis, whichever the population study uses; they are treated as equivalent.
About a quarter of US adults aged 80 and over had dementia in the last national clinical study (2002), against a target below 5%. Dementia prevalence among Americans 65 and over has been falling about 2.3% a year (adjusted for age); extending that decline — with the part due to rising education fading out, a chance that the decline stalls and a chance that prevention drugs add to it — gives a median forecast of 2079 (10% chance by 2052; 68% by 2100), more than 10 years out, so red. The main uncertainty is whether prevalence among the over-80s falls as fast as among all over-65s.- ETA
- ~2079 (80% range 2052–never/after 2100; P(by 2100) 68%)
- ETA cruxes
- Crux 1 is mapping M1, whether 80+ standardised prevalence falls as fast as 65+ age-adjusted (×0.75 makes p50 null); gather 80–84/85–89/90+ waves (HRS-HCAP, NHATS). Then whether the decline continues or stalls (r0 1.6%: null; stall .23 as if the flat GBD Europe estimate were a measured series: 2083), the 2002 baseline (2075–2086), the education share and fade (2073–2084), and prevention readouts (AHEAD 3-45, TRAILBLAZER-ALZ 3, PrevenTRON ~2031–33; regime .20–.60: 2072–2089).
- Path
- Reading (framework.yaml details): CDR≥1 or clinical-consensus dementia among all residents 80+ of a 50M+ population, WHO-standardised within 80+, below 5%. Proxy trend: US HRS 65+ age-adjusted prevalence, calibrated on ADAMS, assumed to apply 1:1 to 80+ (same outcome, overlapping mostly-80+ case population). Path: continued vascular, education and smoking gains, plus possible disease-modifying prevention (anti-amyloid in preclinical disease, tau lowering, vascular and metabolic risk control).
- Blockers
- Measurement: no registered 80+ CDR≥1 prevalence series and no US national clinical prevalence study since ADAMS; register HRS-HCAP/NHATS age-specific waves.
- Mapping: the 65+ age-adjusted trend may overstate the 80+ decline; modelled European estimates show a near-flat age-standardised trend.
- Clinical: no preclinical prevention trial has read out positive; semaglutide (EVOKE) failed and blarcamesine was withdrawn; anti-amyloid effects are partial with ARIA risk.
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- Trend (exponential prevalence decline, sourced HRS series) with named, counted stall and no-regime slow-decline scenarios and a prevention-regime branch; unconditional Monte Carlo (40,000 draws, common random numbers)
- ETA calculation
- Anchor: ADAMS 2002 24.19% (80–89), 37.36% (90+) (measured); WHO weights within 80+ .874/.126 (sourced) → .2585, prior U(.2419,.28) (assumed); gap ×5.17 (ln 1.643). Trend (sourced): HRS 65+ age-adjusted 12.2%→8.5% 2000–2016 = 2.26%/yr; NHATS 72+ 2.8%/yr 2011–19; CFAS 1.22%/yr 1991–2011; Wolters 2020 incidence −13%/decade. M1 (assumed): 65+ rate applies 1:1 to 80+. Model (prev, reproduced exactly): r0=.0226×LN(1,.30); education share U(.2,.4) fades from 2040; F-stall .15 from 2030; prevention regime p .40 adding LN(.020,.55)/yr from U(2035,2050). Constant-HRS check 2074.7. Named failures (total .317): F-stall drawn .15, realised .112 — class 1 of 10 high-income trend series not declining (Hisayama), 1 of 9 deduplicated: raw .10–.11, Laplace .167–.182; .15 disclosed judgment. No-regime slow decline .200 = .60×.85×.39 (1 of 3 national prevalence series below the needed ~1.68%/yr, Laplace .40; regime .40 within the counted bracket .36–.49 from 0 of 3 preclinical anti-amyloid prevention readouts, Laplace .20 per family, 2–3 families). Result: P .6829, p10 2052.52, p50 2079.61, p90 null; conditional median 2067.95; calendar-year floors. New evidence tested: GBD 2023 EU age-standardised prevalence −1.7% since 1990 (2026-07-31-gbd-2023-eu-dementia-prevalence-by-age) is a modelled synthesis, not a measured trend series; if counted as a second non-decliner (2 of 11, Laplace .23) → p50 2083.04, P .641 (sensitivity, not adopted). PrevenTRON (1,600 people, 4–6 yr) readout ~2031–33 is consistent with the regime start (U(2032,2045) → 2078.65). Sensitivities (p50, P): M1 ×.75 → null, .474; r0 1.6%/2.8% → null (.440)/2065.66; P0 .2419/.30 → 2075.97/2086.63; edu .2/.4 → 2075.87/2084.42; stall 0/.30 → 2074.70/2086.73; regime .20/.60 → 2089.03/2072.40; regime rate .010/.030 → 2085.96/2075.85; start U(2045,2060) → 2082.41. Cross-section context: robots-software metr-50-horizon-by-release last 1044.78 min (2026-04-07); no measured mapping, no METR multiplier. No shared driver and no anchor.
- Change
- Red unchanged; model and every input unchanged (prev reproduced exactly: .6829/2052.52/2079.61/null), so no quantile or P moves. New evidence (ledger top-up: GBD 2023 EU prevalence, Alzheimer's Association 2026 facts, COSMIC incidence levels, PrevenTRON launch, tau-lowering phase 2/1 data, zoster and ACTIVE observational/trial signals, EVOKE and blarcamesine failures) measures no model input; it added the GBD-EU stall and PrevenTRON start sensitivities and updated the blockers.
- Rubric rule
- §4 and §4.1: red, because the unconditional median ETA 2079 is more than 10 years after made_at. No verified achievement meets the literal wording.
by assess:20261018T211426Z · 19 evidence event(s)Evidence and news: 19 cited by the assessment, 26 other event(s)
Cited evidence (19)
- Arrowhead reported on 14 October 2026 interim phase 1/2a data in healthy volunteers showing that subcutaneously injected ARO-MAPT, a tau-silencing siRNA, crossed the blood-brain barrier and reduced CSF total tau by a mean 73% at day 105 (maximum 85%) at the 12.8 mg/kg dose, supporting quarterly dosing; patient cohorts in early Alzheimer's disease are still enrolling, with data expected in 2027.
- A JAMA Neurology individual-participant meta-analysis of 22 population-based COSMIC cohorts (38,418 people in 18 countries or territories; published 12 October 2026) found an age- and sex-standardised all-cause dementia incidence of 9.7 per 1,000 person-years (8.3 in high-income vs 13.6 in low- and middle-income countries) and projected 24.7 million new cases in 2050 if age-specific incidence stays constant.
- Roche's phase 3 PrevenTRON trial, whose design was first presented at AAIC in July 2026 and which had begun recruiting in the UK by 10 September 2026, will recruit 1,600 cognitively unimpaired people aged 55–80 worldwide and randomise those with high plasma p-tau217 to the brain-shuttle anti-amyloid antibody trontinemab or placebo, following them for four to six years to test prevention or delay of symptomatic Alzheimer's disease.
- A GBD 2023 analysis in The Lancet Regional Health – Europe (31 July 2026) estimated 7.76 million people with dementia in the EU-27 in 2023 (age-standardised prevalence 651.3 per 100,000, down 1.7% since 1990 while case numbers rose 88.5%), with age-specific prevalence of 17.3% in men and 21.4% in women aged 85–89 and 29.0% and 36.4% at ages 90–94.
- WHO published on 15 July 2026 the second edition of its guidelines on risk reduction of cognitive decline and dementia, adding recommendations to reduce air-pollution exposure and implement tailored multidomain interventions alongside physical activity, tobacco cessation, reduced alcohol and management of hypertension, diabetes and high cholesterol, stating that up to 45% of dementia risk is attributable to modifiable risk factors.
- Full phase 2 CELIA data presented at AAIC on 14 July 2026 showed that diranersen 60 mg intrathecally every six months (n=60) slowed decline versus placebo at 18 months by 26% on CDR-SB, 42% on ADAS-Cog13 and 50% on MMSE, with 50–65% mean reductions in CSF total tau and lower tau-PET signal across doses, but no separation on the ADCS-ADL-MCI daily-function scale; Biogen plans phase 3.
- A USC-led microsimulation based on the Health and Retirement Study, published in Alzheimer's & Dementia on 24 June 2026, estimated that 5.7 million US adults aged 51+ (95% CI 5.6–6.0 million) live with dementia in 2026, with total costs of $818 billion including quality-of-life losses, unpaid care and medical and long-term care.
- An observational matched-cohort study of US Medicare beneficiaries aged 65+ (502,845 two-dose recombinant zoster vaccine recipients vs 1,005,690 unvaccinated comparators), published in Alzheimer's & Dementia, found a lower risk of new-onset dementia after vaccination, with hazard ratios of 0.67 (95% CI 0.65–0.68) through 3 years and 0.74 (0.69–0.79) after 3 years.
- The Alzheimer's Association's 2026 Alzheimer's Disease Facts and Figures (21 April 2026) estimates 7.4 million Americans aged 65+ live with clinical Alzheimer's dementia (not all-cause dementia), with prevalence of 5.2% at ages 65–74, 13.8% at 75–84 and 35.8% at 85+, and projects up to 13.8 million by 2060.
- On 25 March 2026 Anavex withdrew its EU marketing authorisation application for blarcamesine in Alzheimer's disease before EMA completed re-examination of its December 2025 recommendation to refuse approval, which had found that the main study failed to demonstrate effectiveness.Unbroken Mind [1] EMA: Blarcamesine Anavex — withdrawal of application
- Full evoke and evoke+ phase 3 results published in The Lancet (3,808 randomised participants with early Alzheimer's disease) confirmed that oral semaglutide up to 14 mg daily did not slow progression: CDR-SB change at week 104 differed from placebo by −0.08 (95% CI −0.35 to 0.20) in evoke and 0.10 (−0.17 to 0.38) in evoke+.Unbroken Mind [1] Efficacy and safety of oral semaglutide 14 mg (flexible dose) in early-stage symptomatic Alzheimer's disease (evoke and evoke+): two phase 3, randomised, placebo-controlled trials (Lancet, Europe PMC record) [2] The Lancet: evoke and evoke+ full results [3] Efficacy and safety of oral semaglutide 14 mg (flexible dose) in early-stage symptomatic Alzheimer's disease (evoke and evoke+) (PubMed abstract, Lancet 2026; Epub 2026 Mar 19)
- A 20-year follow-up of the NIH-funded ACTIVE randomized trial (Medicare claims 1999–2019 for 2,021 participants aged 65+ at enrolment), published in Alzheimer's & Dementia: TRCI on 9 February 2026, found that only participants randomized to adaptive speed-of-processing training who completed booster sessions had a lower risk of claims-diagnosed dementia (HR 0.75, 95% CI 0.59–0.95; a 25% lower rate), while memory and reasoning training showed no effect.Unbroken Mind [1] NIH: Cognitive speed training over weeks may delay the diagnosis of dementia over decades [2] Impact of cognitive training on claims-based diagnosed dementia over 20 years: evidence from the ACTIVE study (Alzheimer's & Dementia: TRCI) [3] Impact of cognitive training on claims-based diagnosed dementia over 20 years: evidence from the ACTIVE study (PubMed abstract, Alzheimers Dement (N Y) 2026 Feb 9)
- According to the NHS, 1 in every 6 people over the age of 80 have Alzheimer’s disease.Unbroken Mind [1] NHS: Alzheimer's disease↳ corrects (or adds sources to) an earlier record:
2025: According to the NHS, about 1 in 6 (~17%) people aged over 80 in the UK have dementia. - The US ADAMS study estimated dementia prevalence at 24.19% among people aged 80–89 and 37.36% among those aged 90 and older.Unbroken Mind [1] PMC: Prevalence of Dementia in the United States: The Aging, Demographics, and Memory Study↳ corrects (or adds sources to) an earlier record:
2025: The US ADAMS study estimated dementia prevalence at about 24% among people aged 80–89… - Eisai reported that modeling estimated continued lecanemab treatment could delay progression from MCI due to Alzheimer’s disease to moderate disease by 3.5 years overall and 8.3 years in the low-amyloid group.↳ follows up on 2025-01: In January 2025 the FDA approved once-every-four-weeks intravenous maintenance dosing… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-12: Eisai reported at CTAD in December 2025 that continued lecanemab treatment may delay… - A 2025 Nature Medicine study in 33,995 Chinese adults found intensive blood pressure control reduced dementia risk by 15% (RR 0.85) over 4 years.↳ corrects (or adds sources to) an earlier record:
2025: A 2025 Nature Medicine study in 33,995 Chinese adults found intensive blood pressure… - On 30 July 2025 Lilly reported three-year TRAILBLAZER-ALZ 2 data: CDR-SB decline was reduced by 1.2 points at 36 months versus the matched ADNI group, and earlier Kisunla initiation reduced progression to the next disease stage by 27% on CDR-G versus delayed initiation.Unbroken Mind [1] Eli Lilly via PR Newswire: Lilly’s Kisunla (donanemab-azbt) showed growing benefit over three years in early symptomatic Alzheimer’s disease [2] Wayback capture 20250730152354↳ corrects (or adds sources to) an earlier record:
2025-07-30: On 30 July 2025 Lilly reported 3-year TRAILBLAZER data for Kisunla (donanemab) showing… - The 2024 Lancet Commission on dementia reported that age-standardized dementia incidence has declined by about 20% in high-income countries over 2–3 decades, attributed to better cardiovascular health and education, while absolute case numbers grow with population aging.
- The SHARE 2022 study found dementia prevalence among people aged 65+ in Europe ranging from 4.5% to 22.7% depending on the country.Unbroken Mind [1] PMC: Harmonized prevalence estimates of dementia in Europe vary strongly with childhood education↳ corrects (or adds sources to) an earlier record:
2022: The SHARE 2022 study found dementia prevalence among people aged 65+ in Europe ranging…
Other events (26)
- On 16 September 2026 the US House Ways and Means Committee unanimously advanced the bipartisan Alzheimer's Screening and Prevention (ASAP) Act, which would remove the legal barrier keeping Medicare from covering early Alzheimer's detection tests (committee passage, not enactment).
- Eisai and Biogen announced on 3 September 2026 that China's NMPA had approved the lecanemab subcutaneous autoinjector (500 mg once weekly as two 250 mg injections) as an initiation treatment for early Alzheimer's disease, allowing at-home treatment in a country where Eisai estimates 17 million people had MCI or mild dementia due to Alzheimer's in 2024.
- On 24 August 2026 the FDA cleared Roche's Elecsys pTau217 blood test, developed with Lilly, as the first single-biomarker blood test supporting both rule-in and rule-out of Alzheimer's amyloid pathology across primary and specialty care, for people 55 and older with signs, symptoms or complaints of cognitive decline (not asymptomatic screening).Precise ScreeningUnbroken MindPredictive Biomarkers [1] Roche receives FDA clearance for Elecsys pTau217
- On 20 August 2026 the FDA cleared C2N Diagnostics' mass-spectrometry PrecivityAD2 blood test (Aβ42/40 plus %p-tau217) to aid identification of Alzheimer's amyloid pathology in adults aged 40 and older with cognitive symptoms; in a 1,142-person validation study its dual cutoffs gave a 97.6% positive and 93.1% negative predictive value, while the 17.3% of results in a third 'likely positive' category had a 77.3% positive predictive value.Predictive BiomarkersUnbroken MindPrecise Screening [1] C2N: FDA Clears C2N Diagnostics' PrecivityAD2 — First Alzheimer's Blood Test for Adults with Cognitive Symptoms as Young as 40 (20 Aug 2026)
- After a fourth committee meeting on 8 July 2026, NICE paused its appraisals of lecanemab (29 July 2026) and donanemab (30 July 2026) for NHS use in England because the manufacturers and NHS England had commenced commercial discussions.
- +21 older event(s): the section's event log
Major Open Challenges
Causal Model of Aging: We lack a unified, predictive model explaining how molecular aging processes cause specific diseases — and which interventions safely extend healthspan without tradeoffs.20 event(s)
Events (20)
- A Nature Aging study published on 6 October 2026 used cross-species, multi-tissue lipidomics to identify lengthening of lipid acyl chains as a conserved ageing signature in mice, C. elegans, Drosophila and humans that also occurs in progressing human heart disease, with genetic analyses supporting a causal role of the lipid remodeller PLB1 in human frailty and Plb1 knockdown reversing lipid elongation and extending C. elegans lifespan.Causal Model of Aging [1] Longer lipids mark aging and constrain lifespan (Nature Aging)
- A Nature Aging study published on 5 October 2026 showed that the age-related decline of chaperone-mediated autophagy makes senescent cells more pro-senescent and inhibits macrophages from engulfing them, so that macrophage-specific CMA loss increased senescent-cell accumulation in mice, while pharmacological CMA activation reduced senescent-cell burden in aged mice and disease severity in a mouse pulmonary fibrosis model.Causal Model of AgingSafe In-Vivo Rejuvenation [1] Decline of chaperone-mediated autophagy in aging impairs macrophage clearance of senescent cells (Nature Aging)
- A Nature study published on 2 September 2026 found that daily semaglutide started in 20-month-old female C57BL/6 mice improved physical and cognitive function, attenuated hallmarks of ageing including inflammation, senescence and stem-cell decline, and, continued for life, extended lifespan by about 100 days on average versus vehicle, acting as a calorie-restriction mimetic with some benefits beyond matched 24% calorie restriction.
- An NIH-funded Science study published on 23 July 2026, profiling hippocampi from 40 neurologically healthy adults aged 20–95, found that between ages 50 and 75 the brain's embryonic yolk-sac-derived microglia are depleted and replaced by cells resembling blood monocyte-derived microglia with elevated inflammatory signatures, while hippocampal astrocytes and blood-brain-barrier-maintaining cells decline and 3D genome architecture erodes across cell types.Brain PreservationCausal Model of Aging [1] NIH: Brain immunity may undergo a major midlife overhaul [2] Epigenetic and 3D genome reprogramming during the aging of the human hippocampus (Science)
- A Science study published on 16 July 2026 reported that in aged mice prostaglandin E2 receptor EP2 signalling in tissue-resident macrophages impairs their clearance of senescent neutrophils, and that reducing macrophage EP2 signalling preserved mitochondrial fitness and prevented cognitive decline, frailty, sarcopenia, adiposity, cardiac impairment and systemic inflammation, while pharmacological EP2 inhibition restored youthful neutrophil clearance.Causal Model of AgingProgrammable Immunity [1] Restored clearance of senescent neutrophils by tissue-resident macrophages limits organ aging (Science) [2] Breakdown of immune cells’ interaction is key driver in aging, study finds (Stanford Medicine) [3] PubMed abstract 42462036: Restored clearance of senescent neutrophils by tissue-resident macrophages limits organ aging (Science)
- +15 older event(s): the section's event log
Predictive Biomarkers: Identifying reliable molecular signatures that detect disease years before symptoms across cancer, cardiovascular, metabolic, and neurodegenerative conditions — with sufficient specificity to avoid overwhelming healthcare with false positives.45 event(s)
Events (45)
- On 9 October 2026 Natera announced that the FDA had granted Breakthrough Device designation to its cell-free-DNA multi-cancer early detection assay after assessing data from 12 cancer types; the test is not FDA approved.Predictive BiomarkersPrecise Screening [1] Natera: FDA Grants Breakthrough Device Designation for Natera's Multi-Cancer Early Detection Assay (9 Oct 2026)
- On 23 September 2026 the FDA's Molecular and Clinical Genetics advisory panel, reviewing the first PMA for a multi-cancer early detection test, voted 7–2 (one abstention) that the benefits of GRAIL's Galleri outweigh its risks for screening adults 50 and older, unanimously on safety and 6–4 on effectiveness; FDA's final decision was still pending.
- A Novo Nordisk-coauthored preprint posted on 22 September 2026 analyzing proteomics from 10,052 participants in five placebo-controlled semaglutide trials (up to 156 weeks) reported that semaglutide reduced proteomic heart biological age by about 2-4 years versus placebo and kidney, brain, pancreas, lung and multi-organ clock ages by 1-4 years, while the authors stated this does not establish biological rejuvenation or surrogate validity.
- The peer-reviewed PATHFINDER 2 results, published in Nature Medicine on 22 September 2026, reported that among 32,007 performance-analyzable adults aged 50+ without suspected cancer, GRAIL's Galleri test had a 0.54% cancer detection rate, 60.3% positive predictive value, 99.64% specificity, 91.3% cancer-signal-origin accuracy and episode sensitivity of 39.3% for all cancers and 69.8% for 12 prespecified deadly cancers, with 0.6% of 35,335 participants undergoing invasive procedures.Predictive BiomarkersPrecise Screening [1] Performance and safety of a multi-cancer early detection test: the PATHFINDER 2 study (Nature Medicine)
- The peer-reviewed NHS-Galleri primary results, published in NEJM on 22 September 2026 (142,250 randomised participants), confirmed no difference in stage III–IV incidence of 12 prespecified cancers after three screening rounds (incidence rate ratio 1.03, 95% CI 0.92–1.14) and a stage IV rate ratio of 0.86 (95% CI 0.74–1.00), with a companion Nature Medicine paper reporting per-round test specificity of about 99.5–99.6%.Precise ScreeningPredictive Biomarkers [1] NEJM: Effect of Screening with Multicancer Early-Detection Test on Late-Stage Cancer Diagnosis [2] PubMed: Effect of Screening with Multicancer Early-Detection Test on Late-Stage Cancer Diagnosis (N Engl J Med, Epub 2026 Sep 22) [3] Nature Medicine: Performance of a multi-cancer early detection test in the randomized controlled NHS-Galleri trial
- +40 older event(s): the section's event log
Safe In-Vivo Rejuvenation: Resetting adult human cells via reprogramming, senolytics, or regeneration without inducing cancer, fibrosis, or loss of cell identity.32 event(s)
Events (32)
- A Wyss Institute/Church-lab study published in Nature Aging on 14 October 2026 used a transcriptomic-age screen in old human fibroblasts to identify the splicing factor SRSF1, whose overexpression reversed several aging hallmarks without loss of fibroblast identity or pluripotency-program activation, and whose liver-tropic AAV delivery to aged mice slowed frailty and extended median lifespan by 14% versus controls.Safe In-Vivo RejuvenationThe Long Middle Age [1] Transcriptomic reprogramming screen identifies SRSF1 as a regulator of aging (Nature Aging) [2] Reversing age by reprogramming gene expression networks (Wyss Institute)
- Rubedo Life Sciences reported on 12 October 2026 that four weeks of topical RLS-1496 in its Phase 1b atopic dermatitis study reduced measured skin epigenetic age (MitraBio clock) by a mean 2.1 years across 11 paired photo-aged skin samples (p=0.032), and that final Phase 1b/2a data in 22 actinic keratosis patients showed a mean 49% lesion reduction after 28 days with minimal irritation.
- On 8 October 2026 Life Biosciences reported interim Phase 1 data showing that a single intravitreal Dose Level 1 of the OSK reprogramming therapy ER-100 with 56 days of oral doxycycline was well tolerated through Day 56 in 3 open-angle glaucoma participants with no dose-limiting toxicities or serious adverse events, with visual-field improvements in 2 of 3 and a recommendation to escalate to Dose Level 2.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Life Biosciences Announces First-in-Human Data from Ongoing Phase 1 Trial Evaluating ER-100 in Optic Neuropathies
- A Nature Aging study published on 5 October 2026 showed that the age-related decline of chaperone-mediated autophagy makes senescent cells more pro-senescent and inhibits macrophages from engulfing them, so that macrophage-specific CMA loss increased senescent-cell accumulation in mice, while pharmacological CMA activation reduced senescent-cell burden in aged mice and disease severity in a mouse pulmonary fibrosis model.Causal Model of AgingSafe In-Vivo Rejuvenation [1] Decline of chaperone-mediated autophagy in aging impairs macrophage clearance of senescent cells (Nature Aging)
- A double-blind randomized proof-of-principle trial of the senolytics dasatinib plus quercetin in 31 patients with biopsy-proven fibrotic MASH (21 weeks; Nature Metabolism, 1 October 2026) met its primary endpoint, with ≥1-point fibrosis improvement without MASH worsening in 8 of 17 (47%) D+Q participants versus 1 of 14 (7%) on placebo.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Senolytics dasatinib and quercetin in metabolic dysfunction-associated steatohepatitis: a proof-of-principle randomized, controlled trial (Nature Metabolism)
- +27 older event(s): the section's event log
Programmable Immunity: Continuously monitoring and reprogramming the immune system over decades to prevent cancer, infections, and autoimmunity without breaking tolerance.51 event(s)
Events (51)
- BioNTech announced on 13 October 2026 that the phase 2/3 AHEAD-MERIT trial of its mRNA cancer immunotherapy BNT113 plus pembrolizumab in HPV16-positive metastatic head and neck cancer missed its first dual primary endpoint of progression-free survival versus pembrolizumab alone, with the trial continuing to its overall-survival endpoint.Taming CancerProgrammable Immunity [1] BioNTech — Provides Update on Phase 2/3 Trial for BNT113 in HPV16-positive Head and Neck Cancer (13 Oct 2026)
- On 8 October 2026 the FDA approved atezolizumab with fluoropyrimidine and oxaliplatin chemotherapy as adjuvant treatment of stage III mismatch repair-deficient colon cancer, based on ATOMIC, where adding atezolizumab to mFOLFOX6 improved disease-free survival versus mFOLFOX6 alone with a hazard ratio of 0.50 (3-year DFS 86.3% vs 76.2% in NEJM).Taming CancerProgrammable Immunity [1] FDA approves atezolizumab in combination with chemotherapy for Stage III mismatch repair deficient colon cancer [2] Atezolizumab plus FOLFOX for Stage III Mismatch Repair–Deficient Colon Cancer — NEJM 2026 [3] Atezolizumab plus FOLFOX for Stage III Mismatch Repair-Deficient Colon Cancer — N Engl J Med 2026;394(12):1155-1166 (PubMed)
- Adicet reported on 28 September 2026 that in its phase 1 study of prula-cel, an off-the-shelf allogeneic gamma-delta CD20 CAR-T, in 22 efficacy-evaluable lupus patients (16 with lupus nephritis), at 12 months 54% of evaluable patients achieved DORIS remission and 50% of evaluable lupus nephritis patients a complete renal response, with all patients having discontinued immunosuppressants, and it plans pivotal lupus nephritis trial start-up in Q4 2026.
- At the IASLC World Conference on Lung Cancer on 13 September 2026, the China-only phase 3 HARMONi-2 trial in 398 patients with treatment-naive PD-L1-positive advanced NSCLC reported that ivonescimab monotherapy improved median overall survival to 30.8 versus 22.6 months with pembrolizumab (HR 0.73).Taming CancerProgrammable Immunity [1] IASLC — Late-Breaking HARMONi-2 Analysis Shows Ivonescimab Significantly Improves Overall Survival Versus Pembrolizumab
- An international consensus statement published in Diabetologia on 10 September 2026 and endorsed by multiple diabetes professional associations set out principles and minimum requirements for general-population screening for islet autoantibodies to diagnose early-stage (presymptomatic) type 1 diabetes, noting that 20–67% of children and adolescents currently present with diabetic ketoacidosis at clinical diagnosis.Predictive BiomarkersPrecise ScreeningProgrammable Immunity [1] International consensus guidance for general population screening for islet autoantibodies to diagnose early-stage type 1 diabetes (Diabetologia) [2] PubMed: International consensus guidance for general population screening for islet autoantibodies to diagnose early-stage type 1 diabetes: a nominal group technique process (Diabetologia abstract)
- +46 older event(s): the section's event log
Brain Preservation: Maintaining cognition and identity for 100+ year lifespans despite protein aggregates, vascular damage, and neuron loss — repair at scale, not just slowing decline.28 event(s)
Events (28)
- Arrowhead reported on 14 October 2026 interim phase 1/2a data in healthy volunteers showing that subcutaneously injected ARO-MAPT, a tau-silencing siRNA, crossed the blood-brain barrier and reduced CSF total tau by a mean 73% at day 105 (maximum 85%) at the 12.8 mg/kg dose, supporting quarterly dosing; patient cohorts in early Alzheimer's disease are still enrolling, with data expected in 2027.
- uniQure reported on 29 September 2026 that versus an updated propensity-matched external control, high-dose AMT-130 slowed Huntington's disease progression at 36 months by 80% on cUHDRS (n=15, nominal p=0.005) and 67% on Total Functional Capacity, and at 48 months by 44% on cUHDRS (n=12, not significant, p=0.144) and 61% on TFC (nominal p=0.008), with 5 of the high-dose participants (17%) having had resolved treatment-related CNS-inflammation serious adverse events.Brain Preservation [1] uniQure Announces Additional Data from Ongoing Phase I/II Studies of ifezuntirgene inilparvovec (AMT-130) in Huntington's Disease Showing Continued Slowing of Disease Progression [2] uniQure Announces Additional Data from Ongoing Phase I/II Studies of ifezuntirgene inilparvovec (AMT-130) in Huntington's Disease
- Ionis and Otsuka announced on 22 September 2026 that the Phase 3 FUSION trial of ulefnersen, an intrathecally administered RNA-targeted medicine designed to reduce FUS protein production, met its primary endpoint in people with FUS-ALS (primary analysis population N=73), a statistically significant improvement on a 72-week joint rank of ALSFRS-R function, time to rescue and ventilation-assistance-free survival, with a statistically significant improvement in serum neurofilament light chain, making it the first FUS-ALS trial to meet its primary endpoint.
- On 3 September 2026 the FDA approved Ionis's Zanvastro (zilganersen), an intrathecal antisense oligonucleotide that reduces production of the abnormal GFAP protein, as the first treatment for Alexander disease and the first therapy to directly target the protein buildup that drives it, based on a 49-patient randomized controlled study in which treated patients aged 5+ with walking difficulty at baseline had significantly better walking speed at 61 weeks than untreated controls.Brain Preservation [1] FDA Approves First Drug to Treat Alexander Disease
- uniQure announced on 2 September 2026 that it had submitted a BLA to the FDA for accelerated approval and a Marketing Authorisation Application to the UK MHRA for ifezuntirgene inilparvovec (AMT-130) for Huntington's disease, a one-time striatal huntingtin-silencing gene therapy, based on 3-year Phase I/II data versus a propensity score-matched external control.
- +23 older event(s): the section's event log
Beyond the framework: 53 event(s)
- CDC reported 4,410 confirmed US measles cases in 2026 as of 15 October 2026, nearly double the 2,289 for the full year 2025, with 42 new outbreaks and 96% of cases outbreak-associated, ahead of PAHO's November 2026 review of US elimination status.
- WHO's 8 October 2026 Disease Outbreak News reported that the Bundibugyo-virus Ebola outbreak in the Democratic Republic of the Congo had reached 8,728 confirmed cases and 4,205 confirmed deaths (48.2% case fatality) across 64 health zones as of 6 October 2026, with a first imported fatal case in Kenya.
- WHO reported on 29 September 2026 that 47 countries reported 451,499 cholera cases and 7,870 cholera deaths in 2025, deaths up 30% despite fewer cases and the highest annual total reported to WHO since 1999.Beyond the framework [1] WHO — WHO reports latest global trends in cholera, meningitis and plague
- Novartis announced on 4 September 2026 that the Phase 3 Lp(a)HORIZON outcomes trial of pelacarsen in 8,323 patients with elevated lipoprotein(a) and established cardiovascular disease lowered Lp(a) but did not meet its primary endpoint of reducing major cardiovascular events.Beyond the framework [1] Novartis — Lp(a)HORIZON Phase III topline results for pelacarsen
- On 29 July 2026 the US State Department and HHS said the United States would immediately release all $600 million of Congressionally appropriated FY25 and FY26 funds to Gavi after vaccine-safety commitments, while providing no US funding to WHO through Gavi.
- +48 older event(s): the section's event log
Latest news: the 8 most recently published of 304 events
- CDC reported 4,410 confirmed US measles cases in 2026 as of 15 October 2026, nearly double the 2,289 for the full year 2025, with 42 new outbreaks and 96% of cases outbreak-associated, ahead of PAHO's November 2026 review of US elimination status.
- A Wyss Institute/Church-lab study published in Nature Aging on 14 October 2026 used a transcriptomic-age screen in old human fibroblasts to identify the splicing factor SRSF1, whose overexpression reversed several aging hallmarks without loss of fibroblast identity or pluripotency-program activation, and whose liver-tropic AAV delivery to aged mice slowed frailty and extended median lifespan by 14% versus controls.Safe In-Vivo RejuvenationThe Long Middle Age [1] Transcriptomic reprogramming screen identifies SRSF1 as a regulator of aging (Nature Aging) [2] Reversing age by reprogramming gene expression networks (Wyss Institute)
- Arrowhead reported on 14 October 2026 interim phase 1/2a data in healthy volunteers showing that subcutaneously injected ARO-MAPT, a tau-silencing siRNA, crossed the blood-brain barrier and reduced CSF total tau by a mean 73% at day 105 (maximum 85%) at the 12.8 mg/kg dose, supporting quarterly dosing; patient cohorts in early Alzheimer's disease are still enrolling, with data expected in 2027.
- BioNTech announced on 13 October 2026 that the phase 2/3 AHEAD-MERIT trial of its mRNA cancer immunotherapy BNT113 plus pembrolizumab in HPV16-positive metastatic head and neck cancer missed its first dual primary endpoint of progression-free survival versus pembrolizumab alone, with the trial continuing to its overall-survival endpoint.Taming CancerProgrammable Immunity [1] BioNTech — Provides Update on Phase 2/3 Trial for BNT113 in HPV16-positive Head and Neck Cancer (13 Oct 2026)
- Rubedo Life Sciences reported on 12 October 2026 that four weeks of topical RLS-1496 in its Phase 1b atopic dermatitis study reduced measured skin epigenetic age (MitraBio clock) by a mean 2.1 years across 11 paired photo-aged skin samples (p=0.032), and that final Phase 1b/2a data in 22 actinic keratosis patients showed a mean 49% lesion reduction after 28 days with minimal irritation.
- A JAMA Neurology individual-participant meta-analysis of 22 population-based COSMIC cohorts (38,418 people in 18 countries or territories; published 12 October 2026) found an age- and sex-standardised all-cause dementia incidence of 9.7 per 1,000 person-years (8.3 in high-income vs 13.6 in low- and middle-income countries) and projected 24.7 million new cases in 2050 if age-specific incidence stays constant.
- On 9 October 2026 Natera announced that the FDA had granted Breakthrough Device designation to its cell-free-DNA multi-cancer early detection assay after assessing data from 12 cancer types; the test is not FDA approved.Predictive BiomarkersPrecise Screening [1] Natera: FDA Grants Breakthrough Device Designation for Natera's Multi-Cancer Early Detection Assay (9 Oct 2026)
- On 8 October 2026 Life Biosciences reported interim Phase 1 data showing that a single intravitreal Dose Level 1 of the OSK reprogramming therapy ER-100 with 56 days of oral doxycycline was well tolerated through Day 56 in 3 open-angle glaucoma participants with no dose-limiting toxicities or serious adverse events, with visual-field improvements in 2 of 3 and a recommendation to escalate to Dose Level 2.The Long Middle AgeSafe In-Vivo Rejuvenation [1] Life Biosciences Announces First-in-Human Data from Ongoing Phase 1 Trial Evaluating ER-100 in Optic Neuropathies
Climate and Environment
215 events · latest news published 2026-10-18 (4 days ago)news per week
This endeavor is unlike the other four, as it is about fixing the past rather than improving the future. A hundred years ago, we would have been talking about terraforming Sahara and Antarctica, but humanity messed up the Earth. Many historic civilizations fell to climate change, and we now have to clean up our mess if we are to survive into a better future.
KPI
CO₂ at Mauna Loa, monthly mean (NOAA)
426.07 ppm
Sep 2026 · source
▲ +1.70 vs Sep 2025
next update due by 7 Nov 2026
Regressing: 427.35 ppm (2025); +2.56 ppm/yr, up from 2.15 the decade before; need ≤1.79 · assessed 2026-10-18
Atmospheric CO₂ at Mauna Loa reached 427.35 ppm in 2025, and it is rising faster, not slower: 2.56 ppm per year on average over 2016–2025, up from 2.15 over 2006–2015, with each successive five-year average higher than the last. The goal is for the rise to stop, and the IPCC 1.5°C pathways cited by the UK Met Office need 2020s growth of at most 1.79 ppm per year, so the trend is moving the wrong way: red. The slower rise so far in 2026 (about 1.8 ppm per year) follows a La Niña and is expected to reverse as a record El Niño develops, so it does not change the decade-scale picture.
The KPI's precise definition
The KPI, as defined: Atmospheric CO₂ concentration (ppm) and 10-year trend (ppm/year)
How the status was judged: the data window, the pace benchmark, the rule
- Assessed quantity
- co2-trend-10yr-mlo-jan-dec: NOAA Mauna Loa Jan–Dec CO₂ growth, ten-year mean in ppm/yr; goal_direction down toward ≤ 0; trend window 10 years (ENSO cycle), compared with the previous non-overlapping window and half-decade means.
- Window
- Current
- Span
- 2016-2025
- Annual ppm
- 20163.0320171.920182.8520192.4920202.320212.3520221.8420233.3220243.3320252.23
- Mean ppm per yr
- 2.564
- SD annual
- 0.5431
- Half decades
- 2016-2020
- 2.514
- 2021-2025
- 2.614
- Registered 10yr mean
- 2.56
- Previous
- Span
- 2006-2015
- Annual ppm
- 20061.6920072.3120081.542009220102.320111.9220122.6520131.9920142.1720152.95
- Mean ppm per yr
- 2.152
- SD annual
- 0.4247
- Half decades
- 2006-2010
- 1.968
- 2011-2015
- 2.336
- Registered 10yr mean
- 2.15
- Change ppm per yr
- 0.412
- SE difference
- 0.218
- Difference over SE
- 1.89
- Mean 2020 2025
- 2.562
- Level ppm 2025
- 427.4
- Method
- Means of the annual co2-growth-mlo-jan-dec rows; the registered rolling co2-trend-10yr-mlo-jan-dec values (2.56 / 2.15) are the same means rounded. The SE is descriptive (independent-years approximation), not a formal test.
- Source
- uv run core/ledger.py series climate co2-growth-mlo-jan-dec / co2-trend-10yr-mlo-jan-dec / co2-trend-5yr-mlo-jan-dec / co2-mlo-annual --as-of 2026-10-18
- Benchmark
- Tier 1: 2020s mean growth ≤ 1.79 ppm/yr (upper end of 1.33–1.79 for IPCC AR6 C1 1.5°C pathways). 2020–2025 mean 2.5617 (incomplete decade) exceeds it.
- Benchmark source
- UK Met Office CO₂ forecast page citing IPCC AR6 C1 pathways (2026-02-04-metoffice-co2-too-fast-for-1-5c); metrics/kpi-assessment.csv row climate, tier 1.
- Transients discounted
- Jan–Sep 2026 monthly year-on-year gains 1.97, 2.26, 2.00, 1.48, 1.83, 1.82, 1.26, 2.07, 1.70 ppm (mean 1.82; co2-growth-mlo-yoy-monthly; Jan–Sep average +1.82 vs +2.86 a year earlier, 2026-10-09-noaa-co2-ytd-jan-sep-2026-slower-rise): shorter than one window, named cause (La Niña-like 2025–26 conditions), expected reversal (record El Niño forecast, Niño 3.4 ~+3.7°C Oct–Dec: 2026-10-06-wmo-el-nino-record-nino34-forecast, 2026-10-08-noaa-cpc-enso-strong-el-nino-advisory). Moves the status in neither direction.
- An El Niño-driven 2027 growth spike is a projection (2026-09-28-hausfather-el-nino-2027-co2-growth-spike), not evidence.
- Data as of
- 2025
- Change
- No new annual observation of the assessed metric since prev (2026-10-13-kpi-2, rubric v4; KPI rules are unchanged v2–v6, so no rubric correction is needed): the window means are identical. New since prev: September 2026 monthly data (426.07 ppm, +1.70 YoY; 2026-10-09-noaa-co2-monthly-sep-2026) and the Jan–Sep summary extend the transient, and the El Niño forecasts strengthen the expected reversal. data_as_of 2025 (published 2026-01-10) is within 2 annual cadences + lag: no Stale prefix. Status red unchanged.
- Rubric rule
- RUBRIC.md v6 §3 (KPI rules unchanged since v1): red — the assessed growth rate moves away from the downward goal across non-overlapping ten-year windows (and across all four half-decades); the ppm level is reported, not scored.
by assess:20261018T192714Z · 14 evidence event(s)
Evidence and news: 14 cited by the assessment, 14 other KPI news
Cited evidence (14)
- Computed from NOAA GML Mauna Loa (MLO) monthly means released on 9 October 2026, the January-September 2026 average CO2 concentration was 429.53 ppm, +1.82 ppm above the January-September 2025 average (versus +2.86 ppm for January-September 2025 over 2024), with year-on-year monthly gains between +1.26 ppm (July) and +2.26 ppm (February) in every month so far, a pace below the Met Office's February 2026 forecast of a 2.37 ± 0.55 ppm rise in the 2026 annual mean (Scripps basis, premised on La Niña-like conditions).KPI [1] NOAA GML: Mauna Loa monthly mean CO2 (co2_mm_mlo.txt, File Creation 9 Oct 2026) [2] Met Office: Mauna Loa carbon dioxide forecast for 2026 (published 4 February 2026)
- NOAA GML data released on 9 October 2026 put the Mauna Loa (MLO) monthly mean CO2 concentration for September 2026 at 426.07 ppm (+1.70 ppm on September 2025's 424.37 ppm), and the NOAA global marine-surface monthly mean for July 2026 at 426.10 ppm (+2.01 ppm on July 2025's 424.09 ppm).KPI [1] NOAA GML: Trends in Atmospheric CO2 — Mauna Loa (September 2026: 426.07 ppm; last updated Oct 09, 2026) [2] NOAA GML: Trends in Atmospheric CO2 — global monthly mean (July 2026: 426.10 ppm; last updated Oct 09, 2026) [3] NOAA GML: Mauna Loa monthly mean CO2 (co2_mm_mlo.txt, File Creation 9 Oct 2026) [4] NOAA GML: global marine-surface monthly mean CO2 (co2_mm_gl.txt, File Creation 9 Oct 2026) [5] Met Office: Mauna Loa carbon dioxide forecast for 2026 (published 4 February 2026)
- NOAA's Climate Prediction Center ENSO diagnostic discussion of 8 October 2026 kept an El Niño Advisory in force, reporting that El Niño strengthened further over the past month with Niño indices reaching +2.1°C in Niño-3.4, +3.0°C in Niño-3 and +3.9°C in Niño-1+2, and gave a greater than 83% chance of a strong-to-very-strong El Niño through January-March 2027, with chances of a historic event exceeding all El Niños since 1950 (3-month relative ONI of +2.5°C or more) of 54% for September-November, 83% for October-December and 70% for November-January.
- WMO's Global Seasonal Climate Update (6 October 2026) forecast the October-December 2026 seasonal-average Niño 3.4 sea-surface temperature anomaly at about +3.7°C, peaking around December, the highest since observations began versus the previous record of +2.6°C in November 2015-January 2016, after the index rose from +1.6°C in June to +2.5°C in August, with a near-100% likelihood (consensus of 13 centres) that El Niño persists through February 2027.Beyond the framework [1] Sea-surface temperatures reflect strengthening El Niño (WMO, 6 Oct 2026)
- Climate scientist Zeke Hausfather wrote on 28 September 2026 (The Climate Brink) that, because El Niño-driven tropical drought and fire weaken the land carbon sink (the reason NOAA's Mauna Loa CO2 growth rate hit a record 3.33 ppm in 2024), a similar spike in the atmospheric CO2 growth rate is likely in 2027, with September 2026 seasonal model runs putting the 2026-27 El Niño's Niño 3.4 peak at around 4.1°C (middle 80% of ensemble members 3.4-4.6°C), well above the 2015-16 record of 2.75°C; this is a qualitative expectation, not a quantified CO2 forecast.
- A Carbon Brief analysis (21 September 2026) found that the daily Niño 3.4 sea-surface temperature anomaly, calculated with NOAA's ONI convention (centred 30-year climatology), reached 3.11°C on 21 September 2026, surpassing the previous daily record of 3.08°C set on 18 November 2015, with the August 2026 monthly anomaly (about 2.45°C) already above the peak of every prior El Niño on record except 2015-16 (2.75°C) and 1877-78 (about 2.7°C, reconstructed).Beyond the framework [1] Analysis: 'Super El Niño' breaks 'remarkable' all-time record (Carbon Brief)
- NOAA GML data put the Mauna Loa (MLO) monthly mean CO2 concentration for June 2026 at 431.43 ppm (+1.82 ppm on June 2025), and the NOAA global marine-surface monthly mean at 427.62 ppm versus 425.90 ppm in June 2025.
- Computed from NOAA GML Mauna Loa (MLO) monthly means, the year-on-year gains for January-May 2026 were +1.48 to +2.26 ppm, well below the May-over-May gains of 2023-2025 (+3.03, +2.90 and +3.61 ppm).
- The UK Met Office said even its forecast 2026 rise in Mauna Loa CO2 is 'too fast to track IPCC 1.5°C scenarios', which require a 2020s decadal average growth of 1.33-1.79 ppm/yr versus 2.61 ppm/yr observed for 2020-2025 (Scripps annual-mean basis; 1.91 ppm/yr in the 2000s and 2.41 in the 2010s).
- The UK Met Office forecast (published 4 February 2026) put the 2026 Mauna Loa annual mean CO2 concentration at 429.4 ± 0.6 ppm (Scripps basis), a rise of 2.37 ± 0.55 ppm on the annual-mean basis that is slightly slowed by moderate La Niña-like conditions (2.56 ppm without La Niña), with the May 2026 monthly mean forecast at 432.2 ± 0.6 ppm.
- Computed from NOAA GML annual Jan 1 -> Dec 31 growth rates, the Mauna Loa (MLO) rolling 10-year mean CO2 growth dipped from 2.64 ppm/yr (2015-2024) to 2.56 ppm/yr (2016-2025) only because the El Niño year 2015 (+2.95 ppm) left the window and 2025 (+2.23 ppm) entered it, while half-decade means still rose from 2.51 to 2.61 ppm/yr (MLO) and from 2.44 to 2.63 ppm/yr (global) between 2016-2020 and 2021-2025.
- Computed as the mean of NOAA GML annual Jan 1 -> Dec 31 growth rates, the 10-year CO2 trend for 2016-2025 is 2.56 ppm/yr at Mauna Loa (MLO) and 2.53 ppm/yr for the global marine-surface mean, versus a global 2.38 ppm/yr for 2011-2020.KPI [1] NOAA GML: Mauna Loa annual CO2 growth rates (co2_gr_mlo.txt) [2] NOAA GML: global annual CO2 growth rates (co2_gr_gl.txt)↳ follows up on 2024: The WMO Greenhouse Gas Bulletin reported that global annual-mean atmospheric CO2 rose… (corrected version)
- NOAA GML's annual CO2 growth rate (Jan 1 -> Dec 31 change) for 2025 was +2.23 ppm at Mauna Loa (MLO) and +2.06 ppm for the global marine-surface mean (the smallest global rate since 2014), down from the 2024 records of +3.33 ppm (MLO) and +3.76 ppm (global).KPI [1] NOAA GML: Mauna Loa annual CO2 growth rates (co2_gr_mlo.txt) [2] NOAA GML: global annual CO2 growth rates (co2_gr_gl.txt)↳ follows up on 2024: The WMO Greenhouse Gas Bulletin reported that global annual-mean atmospheric CO2 rose… (corrected version)
- NOAA GML data put the 2025 annual mean CO2 concentration at a record 427.35 ppm at Mauna Loa (MLO) and 425.62 ppm for the global marine-surface mean; on the MLO annual-mean basis the 2025 increment was +2.74 ppm, down from +3.53 ppm in 2024.
Other events (14)
- NOAA GML data put the Mauna Loa (MLO) monthly mean CO2 concentration for May 2026, the seasonal peak, at 432.34 ppm, the highest monthly value on record and +1.83 ppm on May 2025 (430.51 ppm), with the NOAA global marine-surface monthly mean at 428.59 ppm.KPI [1] NOAA GML: Mauna Loa monthly mean CO2 (co2_mm_mlo.txt) [2] NOAA GML: global marine-surface monthly mean CO2 (co2_mm_gl.txt)↳ follows up on 2025-05: Scripps reported a May 2025 monthly mean of 430.2 ppm at Mauna Loa; NOAA reported 430.5… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put the global atmospheric CO2 growth rate at 5.6 ± 0.02 GtC per year (2.6 ppm per year) during the decade 2015-2024.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) reported a record-high global atmospheric CO2 growth rate of 7.9 ± 0.02 GtC (3.7 ppm) in 2024, mainly due to the 2023/2024 El Niño conditions, and a preliminary 2025 growth estimate of 2.1 ppm.↳ follows up on 2024: The WMO Greenhouse Gas Bulletin reported that global annual-mean atmospheric CO2 rose… (corrected version)↳ follows up on 2025-11-13: A Nature study published alongside the Global Carbon Budget 2025 found land and ocean… (corrected version)
- NOAA GML data put the Mauna Loa (MLO) monthly mean CO2 concentration for April 2026 at 431.12 ppm (+1.48 ppm on April 2025), and the NOAA global marine-surface monthly mean at 428.67 ppm.
- NOAA GML's Mauna Loa (MLO) daily mean CO2 reached 433.95 ppm on 1 May 2026, the highest daily mean in NOAA's MLO daily record.
- +9 older event(s): the section's event log
20 more metric(s) we track
CO₂ global, monthly mean (NOAA)
426.10 ppm
Jul 2026 · source
▲ +2.01 vs Jul 2025
next update due by 9 Nov 2026
CO₂ change on the same month a year earlier (Mauna Loa)
1.70 ppm
Sep 2026 · source
▼ -0.64 vs Sep 2025
next update due by 7 Nov 2026
CO₂ at Mauna Loa, monthly mean (Scripps)
425.61 ppm
Sep 2026 · source
▲ +1.54 vs Sep 2025
next update due by 5 Dec 2026
Milestone Countdown
A sequence: each milestone is expected after the one before it.
Expected year: dot = median (hollow: after 2100), bar = 80% range; arrows: each milestone is expected after the one before. The time axis is compressed: the farther out, the more years per step.
The Bend: Year of peak global greenhouse gas emissions is definitively behind us.Progressing: GHG roughly flat (+0.3% Jan–Jul 2026), peak not yet confirmed; ETA ~2036 · assessed 2026-10-18 · 55 event(s)
- Definitively: a major global all-GHG inventory shows the peak year 3+ years back and the latest year at least 1% below it.
Global greenhouse-gas emissions are roughly flat near their highest level: up 0.7% in 2025 by the EU's EDGAR inventory (land use excluded) and 0.3% in January–July 2026 by Climate TRACE; counting land-use CO₂, 2025 was about 0.2% below 2024, far short of the 1% drop the milestone requires. Fossil-CO₂ growth, though, keeps slowing decade by decade, and clean power outgrew electricity demand in 2025. Extrapolating that slowdown puts the actual peak around 2030 and its confirmation — an inventory showing the peak three or more years back and the latest year at least 1% lower, plus a year for reporting — at a median of 2036, with a 95% chance by 2100: yellow, but only just inside the ten-year horizon that counts as progressing. Whether this year's oil-shock dip in emissions becomes a lasting decline, and whether methane and other gases follow CO₂ down, could move this several years either way.- ETA
- ~2036 (80% range 2031–2050; P(by 2100) 95%)
- ETA cruxes
- 1) Mapping of CO₂ deceleration onto non-CO₂ (assumed m=1; m=0 → p50 2040, red); CH₄/N₂O inventories, EU methane-rule delay. 2) Full-year 2026/2027 all-GHG growth: a measured +0.5% → p50 ~2040 (red); a measured fall that survives the 2027 oil rebound → as early as 2029–30. 3) Persistence of the fossil curvature (halved → 2044; doubled → 2033; October outlooks sit at driver p80–p98). 4) China/India 2026–27 emissions (solar slowdown, drought, coal orders). 5) Monte Carlo resolution of the 10-year edge (p50 within ~0.2 y of the boundary).
- Path
- Own quantity: global ALL-GHG emissions (EDGAR excl. LULUCF + net land-use CO₂, ~58.5 GtCO₂e in 2025), not fossil CO₂, electricity or a regional peak. 'Definitively' per the framework detail: an inventory shows the peak ≥3 years back and the latest year ≥1% below it; recognised after one reporting cycle (≈ annual, measured). The TRACE all-GHG series (fastest same-outcome series, to Jul 2026) sets the 2026 rate; the shared driver fossil-co2-deceleration (GCB fossil, ~65% of the inventory) supplies the measured curvature; land use follows its own trend; non-CO₂ follows the fossil curvature by an assumed mapping. Measured clean-power displacement is the precursor. History: same-vintage inventories (EDGAR 2026 excl. LULUCF + GCB 2026 annual net land use) put 2024 (58.30) slightly above 2025 (58.20), so 2024 is an admissible peak candidate (earliest recognition 2028) and 2025 is not a measured record.
- Blockers
- Measurement: no registered harmonised all-GHG series; inventories differ in vintage and scope; definitive recognition needs several post-peak annual totals.
- Economic: the 2026 Hormuz-shock oil fall (IEA Oct: −2.6 mb/d) is already reversing (Gulf crude exports >75% of pre-war in September; IEA 2027 +2 mb/d); record El Niño 2026–27.
- Technical/policy: India's emissions and emergency coal dispatch rising (weak monsoon), China's solar additions −58% Jan–Aug 2026; US standards rolled back; EU methane-rule MRV delay proposed.
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (shared driver fossil-co2-deceleration for fossil-CO₂ curvature; own land-use OLS trend and TRACE-implied non-CO₂ with an assumed curvature mapping) plus the framework's recognition rule, a reporting lag and a named failure scenario; one Monte Carlo distribution
- ETA calculation
- Shared driver fossil-co2-deceleration (drivers.yaml): draw i of `ledger.py driver-draws fossil-co2-deceleration --as-of 2026-10-18` (seed 20261015, n 10,000; byte-identical to the 2026-10-15 draws; a_per_yr2 lognormal p50 .00080015 = (.017320−.009319)/10 from co2-emissions-fossil-gcb 1995/2005/2015/2025 = 23.26/29.19/34.71/38.1 Gt, sdlog .50; g_2020 .0093186 and fossil_2025 38.1 fixed) is the fossil curvature in simulation draw i: fossil rate +0.452%/yr at 2026.5 centrally. Own mapping (prev's, unchanged, rule 7): net land use 2006–2025 OLS −2.203%/yr, level 4.443 (2025 obs 4.1, 2026-05-13-gcb-2025-total-co2-slightly-lower); non-CO₂ 16.0 GtCO₂e = assumed cross-inventory estimate EDGAR 54.1 (2026-09-01-edgar-2026-ghg-2025) − GCB 38.1, rate +0.614%/yr model-inferred, curvature m=1 × driver curvature (assumed mapping); per-draw common rate shift pins all-GHG 2026 to Climate TRACE Jan–Jul +0.3% ± 0.3 (2026-09-24-climate-trace-jul-2026-ghg; δ=0 centrally). 10,000 common draws; quantile year = calendar year containing it. Prev's script (notes/climate/20261015T113313Z.md) reproduced byte-identically on the 2026-10-18 series; no input changed (58 new effective events tested, none measures a modelled quantity). Bend: E=F+L+N; physical peak argmax; recognition = first t ≥ peak+3y with E ≤ 0.99·peak (framework detail) + 1 y reporting (≈ annual: 2025 inventories April/May/September 2026). History (review-1 correction; 2025 is NOT a measured all-GHG record): same-vintage 2026 inventories, EDGAR excl. LULUCF 2024 54.1/1.007 = 53.724 and 2025 54.1 (2026-09-01-edgar-2026-ghg-2025) + GCB annual net land use 4.58/4.10 (co2-emissions-total-gcb − co2-emissions-fossil-gcb) → all-GHG 58.304 (2024) > 58.200 (2025), −0.18%; 2024 enters as a peak candidate (per draw E_2024 = E_2025 × 1.001786, ratio transfer assumed), earliest recognition 2024+3+1 = 2028; 22.97% of draws peak in 2024. S2 .05 (assumed). MC: P .9505, p10/p50/p90 2031.0/2036.7/2050.3 → 2031/2036/2050 (prev 2031.4/2036.8/2050.3 with 2025 taken as the record, floor 2029: that row is now a sensitivity; 2024 1% above 2025: p10 2028.5, p50 2036.2); physical peak 2024.5/2030.6/2042.0; deterministic central peak 2030.8 at 59.1 GtCO₂e, recognition 2037.1; conditional median 2036.4. P(by end 2031) .154; P(by 2036-10-18 = made_at+10y) .5011; P(by end 2036) .511: median on the 10-year edge (MC ±0.2 y). Sensitivities (p50): all-GHG 2026 +0.5% 2039.6, −0.3% 2029.2 (new row; δ is a permanent shift, so it overstates a transient 2026 dip); m=.5/0 2038.0/2040.0; curvature ×½/×2 2043.6/2033.2; total-CO₂ curvature 2033.8; land-use endpoint 2037.5; threshold 0.5%/2% 2035.3/2038.9; zero reporting lag 2035.7; S2 .02/.10 P .980/.895. New evidence weighed, not inputs: IEA Oct OMR 2026 oil demand −2.6 mb/d (projection; ≈ −0.5% all-GHG from oil alone, but TRACE's +0.3% already spans Mar–Jul of the shock; 2026-10-14-iea-omr-oct-2026-oil-demand-fall-2-6mbd), Gulf exports back to >75% of pre-war (2026-10-14-iea-omr-hormuz-flows-sep-2026), EU H1 fossil CO₂ −2.0% (regional, 2026-09-30-crea-eu-co2-h1-2026-down-2pct), power-sector CO₂ not clearly peaked (2026-10-05-hausfather-wind-solar-avoided-23gt). Outlook cross-check (projections, not inputs): model central fossil 2050/2025 = 0.896 (38.1 → 34.14 Gt), cumulative fossil 2026–2050 901 Gt; curvature each outlook implies in the driver form, ×driver p50 (driver CDF): Exxon 0.83 ×1.23 (.65), IRENA WETO PES 0.74 ×1.57 (.81; 2026-10-13-irena-weto-2026-net-zero-only-in-1-5c), DNV 22/39 ×2.37 (.96; 2026-10-07-dnv-eto-2026-energy-co2-likely-peaked), Rystad ~0.5 ×2.72 (.98; 2026-10-15-rystad-ges-2026-fossil-co2-plateau-40gt); they decline faster to 2050 but flatten after (DNV net-zero CO₂ 2098), i.e. the a×2 and −3%/yr-floor rows at once; driver-fit remark for the owner in the notes, driver not refitted. Cross-checks: 2014–16 false plateau 34.78/34.71/34.69 then 35.26/36.00 and the 2020 rebound (0 of 2 post-1990 plateaus definitive); outlook peak dates (DNV: CO₂ likely peaked within 24 months) read with the framework rule → recognition ~2030, discounted ×1.5–3 from 2026.80 → 2031.6–2036.4, matching the model's p10–p50. Prev 2026-10-15 P .95, 2031/2036/2050: same calendar years. No external anchors listed. Foreign context only: fusion lcoe-solar-pv-irena 2010/2024 417→43 $/MWh; no mapping.
- Change
- Re-assessment on 58 new effective events (published 2026-09-28…10-18). Prev's script reproduced byte-identically; the series it reads and the shared driver draws are byte-identical to 2026-10-15. Each input was tested against the new events: none measures a modelled quantity (no new global all-GHG or CO₂ inventory; IEA's October oil-demand fall is a projection, EU H1 −2% regional, DNV/IRENA/Rystad outlooks projections, EV and heat-pump data precursors without a measured conversion), so no input changed (rule 7). Added: an all-GHG 2026 −0.3% sensitivity (p50 2029.2, with the permanent-shift caveat) and an outlook cross-check (the outlooks imply curvature at driver p80–p98; remark for the owner in the notes). Review-1 correction (input/scope): the prev/draft recognition floor treated 2025 as a measured all-GHG record; same-vintage EDGAR-2026 (excl. LULUCF) + GCB-2026 net land use put 2024 (58.304) above 2025 (58.200), so 2024 now enters as a peak candidate and the floor moves 2029 → 2028: p10 2031.4 → 2031.0, p50 2036.8 → 2036.7, P .9505 and p90 2050.3 unchanged; calendar quantiles 2031/2036/2050 and status yellow unchanged.
- Rubric rule
- RUBRIC.md v6 §4/§4.1: yellow = measurable progress on a driver and unconditional median ≤ made_at+10y (p50 2036 ≤ 2036 at calendar-year resolution); not green (median > 5 y; no funded programme covering the whole GHG transition); not achieved (no verified event of a definitive all-GHG peak).
by assess:20261018T192714Z · 46 evidence event(s)Evidence and news: 46 cited by the assessment, 9 other event(s)
Cited evidence (46)
- Benchmark Mineral Intelligence data (16 October 2026) showed global electric vehicle sales of 2.15 million in September 2026, down 2% year on year (up 17% month on month), bringing January–September 2026 sales to 15.5 million, up 3% on the same period of 2025, as record European sales (almost 600,000, +39% y/y) were outweighed by year-on-year falls of 44% in North America (against the September 2025 rush before the US federal EV tax credit expired) and 18% in China, whose passenger-car NEV exports hit a record 558,000.
- Rystad Energy's Global Energy Scenarios 2026 (15 October 2026) said global fossil CO2 emissions are plateauing at around 40 Gt per year and projected that in its new 'Pragmatism' reference scenario they fall to about 20 Gt by 2050 (800 Gt cumulative over 2026-2050, about 0.35°C of additional warming), versus about 2 Gt in 'Climate first' and near today's level for longer in 'Fossil first'.
- The IEA Oil Market Report of 14 October 2026 forecast global oil demand to fall by 2.6 mb/d on average in 2026 (2.5 mb/d year on year in Q4 2026, with diesel 25% of the drop) as the Iran war restricts product availability and raises fuel prices, with a partial 2 mb/d recovery leaving 2027 global oil use 620 kb/d (0.6%) below 2025 levels (its June Report had forecast a 1.1 mb/d fall in 2026 and a 2 mb/d rebound to 105.3 mb/d in 2027).
- The IEA Oil Market Report (14 October 2026) estimated that Gulf producers' crude oil exports rose by 4.4 mb/d to 13.9 mb/d in September 2026 (over 75% of pre-war levels), and that total oil flows through the Strait of Hormuz rose 3.7 mb/d to 10.9 mb/d, while Middle East refined-product and LPG shipments (1.9 mb/d and 0.67 mb/d) remained below half of their pre-war levels (4.7 and 1.5 mb/d) and renewed attacks on regional energy infrastructure and tankers in October left the recovery 'fragile'.
- The European Commission formally proposed on 14 October 2026 a one-year postponement of the EU Methane Regulation's monitoring, reporting and verification obligations for importers of oil, gas and coal (and of the methane-intensity reporting provisions for importers and EU producers), citing security-of-supply pressure from the Middle East conflict while saying the change does not affect the regulation's ambition; the proposal, which needs European Parliament and Council approval, would move the importer requirements from 2027 to 2028.Beyond the frameworkNet-Zero Transition [1] European Commission: Commission proposes flexibility to methane importers to help save costs (IP/26/2119) [2] Rapporteur: Brussels moves to postpone methane reporting rules (14 October 2026) [3] Carbon Brief Daily Briefing (15 October 2026): postpone methane rules
- IRENA's World Energy Transition Outlook 2026 (launched 13 October 2026) projected that under existing government plans (its Planned Energy Scenario) global fossil CO2 emissions fall only 26% by 2050 (0.2% a year over 2025-2030 and 1.4% a year thereafter), with cumulative 2026-2050 energy-related CO2 of about 840 GtCO2 and warming likely at least 2°C, that its Decarbonising Energy Scenario (no longer aligned with the net-zero-by-2050 goal of earlier editions) still leaves about 4 GtCO2 of net fossil CO2 in 2050 (about 480 GtCO2 cumulative), and that only its revised 1.5°C Scenario, with a limited and temporary overshoot and more carbon capture including BECCS, reaches net-zero fossil CO2 in 2050 (cumulative slightly below 400 GtCO2).
- The Forest Declaration Assessment 2026 (13 October 2026) found that global deforestation fell 15% in 2025 to just over 7 million hectares (releasing an estimated 3.9 billion tonnes CO2e), the largest annual decline since the 2018–20 baseline, but that the world remained 70% off track from the pathway to halt deforestation by 2030 (a cumulative 10.7 Mha excess since 2021), with humid tropical primary forest loss down 36% from 2024's fire-driven spike yet still 11% above the 2018–20 baseline, deforestation in tropical Africa still rising (2.4 Mha in 2025), and an estimated 11.8 Mha of forest under active restoration.
- Ember analysis (13 October 2026) of its 56-country monthly EV sales data found that battery-electric cars took 18% of global new car sales in January–August 2026 (up from 16% a year earlier, with BEV sales +11% y/y while plug-in hybrid sales fell 12% and combustion-car sales 6%), that BEV sales outside the US and China grew an estimated 45% and 49 of the 56 tracked countries set record BEV sales shares, and estimated that electric car fleets in oil-importing countries displaced about 200 million barrels of crude oil worth $20 billion in the six months since the Hormuz energy crisis began (about $14 billion of it in China), while the US BEV share of light-duty sales fell to 6%.
- The Climate Action Tracker's US update (13 October 2026) kept its overall US rating at 'Critically insufficient' and projected that under the Trump Administration's policies (including the One Big Beautiful Bill Act and the rescission of the Endangerment Finding and of power-plant and vehicle emission standards) US greenhouse-gas emissions excluding LULUCF will still fall but only to 20-25% below 2005 levels by 2030, versus about 30-40% on the Biden Administration's trajectory (about 600-1,000 MtCO2e, or 11-23%, higher in 2030), with the US exit from the Paris Agreement and move to leave the UNFCCC having nullified its 2030, 2035 and 2050 (net-zero) targets.
- China Passenger Car Association data released on 12 October 2026 showed new energy vehicles (battery-electric, plug-in hybrid and extended-range) took a record 67.0% of China's passenger-car retail sales in September 2026, up 9.1 percentage points year on year, even though NEV retail sales fell 16.7% year on year to 1.084 million (a ninth consecutive monthly decline; January–September 7.758 million, down 12.8%), as gasoline-car retail sales fell 43% and NEV exports rose 163.5% to 558,000.
- DNV's Energy Transition Outlook 2026 (launched 7 October 2026) judged that global energy-related CO2 emissions 'likely peaked in the last 24 months' and that total anthropogenic CO2 emissions (energy, industrial processes and land-use change, net of direct air capture) peaked in 2024 at 43 Gt, forecasting energy-related CO2 falling 44% from 39 Gt today to 22 Gt in 2050 (and 65% to 12 Gt in 2060), with net-zero CO2 only in 2098 and end-of-century warming of about 2.3°C.
- A counterfactual analysis by climate scientist Zeke Hausfather (The Climate Brink, 5 October 2026) estimated that without wind and solar growth since 2005, global power-sector CO2 emissions in 2025 would have been 28% higher (16.9 rather than 13.2 billion tonnes), that wind and solar avoided about 23 billion tonnes of CO2 over 2006-2025 (15-30 billion depending on whether gas or coal fills the gap), and that global electricity-sector emissions 'have yet to clearly peak' but went down in 2025 and may be on track to decline in 2026.The BendNet-Zero Transition [1] The Climate Brink (Zeke Hausfather): What if the world hadn't built wind and solar?
- India Meteorological Department data reported on 30 September 2026 put India's June-September 2026 southwest monsoon rainfall at 759.4 mm, 87% of the long-period average (a 13% deficit), the fourth-lowest since 2001 and 13th-lowest since 1901, with June 35.4% and August 16.3% below normal, as IMD said strong El Niño conditions would keep strengthening through year-end and forecast below-normal (<85% of LPA) October rainfall.Beyond the framework [1] Monsoon rainfall in 2026 fourth-lowest since 2001, says IMD (Mint, 30 Sep 2026)
- CREA's EU CO2 emissions tracker (30 September 2026) put EU fossil-fuel CO2 emissions (combustion in power, industry, buildings and domestic transport; excluding international aviation and shipping, industrial processes and land use) at 1,040.8 Mt in H1 2026, down 2.0% (20.9 Mt) on H1 2025, with oil emissions down 4.3% (oil used in industry and buildings down 10.8%, mostly in March-June after the Hormuz crisis), coal down 1.3% and gas up 0.9%, just over half the 3.6% annual pace needed for the EU's 2030 target.
- The EDGAR 2026 report (September 2026) put total greenhouse-gas emissions excluding LULUCF at 54.1 GtCO2e in 2025, up 0.7% on 2024, with China up 0.1% and the US up 2.2%.↳ follows up on 2024: The EDGAR 2025 report put 2024 global greenhouse-gas emissions at 53.2 GtCO2e,… (corrected version)
- Preliminary Kpler ship-tracking data reported by Reuters on 28 September 2026 showed crude oil exports from key Middle East producers rebounding in September 2026 to 12.8 million barrels per day, the highest since the US-Israeli war with Iran began on 28 February 2026 but still about 6 million b/d below February's 18.8 million b/d, with exports through the Strait of Hormuz expected to reach about 7.4 million b/d in September (the strait carried about a fifth of world crude and LNG supply before the war), while QatarEnergy-linked LNG carriers resumed visible Hormuz transits in September after none were seen in August.Beyond the frameworkThe Bend [1] Reuters: Mideast oil exports rebound in September as Saudi Arabia boosts shipments (28 Sep 2026) [2] Reuters: Qatar-linked LNG traffic through Hormuz strait picks up despite conflict (28 Sep 2026) [3] Arab Times: Middle East oil exports rebound to highest level since Iran war began: Reuters (28 Sep 2026) [4] TradingView (Reuters syndication): Qatar-linked LNG traffic through Hormuz strait picks up despite conflict (28 Sep 2026)
- The European Heat Pump Association reported (28 September 2026) that residential heat pump sales in 12 European countries rose 11% in the first half of 2026 to about 1,163,000 units (from 1,048,000 in H1 2025), which it attributed to the oil and gas price spike after the Strait of Hormuz was closed in February 2026 and to Commission-encouraged cuts in electricity taxes.Net-Zero Transition [1] EHPA: Post-Hormuz heat pump sales climb 11% (28 September 2026)
- India's power ministry ordered on 25 September 2026, under the emergency provisions (Section 11) of the Electricity Act, 112 captive coal-fired power plants of at least 50 MW (serving aluminium, steel, cement and refinery sites) to run at maximum capacity from 1 October to 31 December 2026 and sell surplus power through exchanges to meet an expected rise in electricity demand, as nearly 40% of India's coal-fired plants were running on critically low fuel stocks amid El Niño-driven demand, according to Reuters.
- China's National Energy Administration data, as reported on 25 September 2026, showed China added 97.19 GW of solar PV capacity in January-August 2026, down about 58% from 230.61 GW in the same period of 2025 (August alone: 11.04 GW), alongside 53.26 GW of wind additions.
- Climate TRACE (24 September 2026) estimated global greenhouse-gas emissions of 4.9 GtCO2e in July 2026, up 0.1% on July 2025, and 33.7 GtCO2e year-to-date for January-July 2026, 0.3% above the same period of 2025, with July emissions down 0.3% in China and 2.6% in the US but up 3.4% in India.
- IRENA's third UAE Consensus tracking report, released with the COP31 Presidency and the Global Renewables Alliance on 21 September 2026, found that a record 693 GW of renewable power capacity was added worldwide in 2025 (513 GW solar PV, 158 GW wind), bringing total installed renewable capacity to about 5.15 TW, but that tripling to about 11.2 TW by 2030 now requires average net additions of about 1.2 TW per year in 2026-2030, and that sustaining the 2025 growth rate would leave a shortfall of about 0.6 TW in 2030 (down from 0.9 TW in the previous edition).Net-Zero Transition [1] IRENA et al. (2026), Delivering on the UAE Consensus: Tracking progress toward tripling renewable energy capacity and doubling energy efficiency by 2030 [2] Global renewable deployment must double to hit 2030 climate target, report says (Reuters, 21 Sep 2026) [3] Report launched at the Global Renewables Summit highlights renewable energy progress but warns of challenges ahead to 2030 (ALER)
- CREA's China Energy and Emissions Trends August 2026 snapshot (17 September 2026) said indicators point to a broad-based decline in China's emissions in August 2026, with thermal power generation down 4% (coal power -5.2%, a second consecutive monthly fall), cement output -12%, crude steel -4% and refinery throughput -7% year on year, while thermal power capacity additions reached 50.3 GW in January-July 2026, up 20% and the highest for that period in 15 years.
- A CREA analysis for Carbon Brief (17 September 2026) found that India's CO2 emissions grew 3.7% year on year in the first half of 2026, driven by an 8% rise in steel and cement emissions (now 23% of India's CO2), even though India's power-sector CO2 was flat versus H1 2024 as clean energy met all of the 7% (63 TWh) electricity-demand growth over two years and oil and gas CO2 fell 7% year on year.
- A Carbon Brief analysis (16 September 2026) projected that global fossil CO2 emissions are set to fall by about 0.5% in 2026, driven by the Strait of Hormuz crisis.
- A Carbon Brief/CREA analysis (3 September 2026) found that China's CO2 emissions fell 1% in Q2 2026 as oil use dropped 9% amid the Strait of Hormuz crisis, leaving first-half 2026 emissions 'up marginally' but still below their 2023-24 peak.
- Climate TRACE (27 August 2026) estimated total greenhouse-gas emissions (CO2e) of 29.7 GtCO2e in H1 2026, up 0.2% on H1 2025, with China down 0.3%, the US down 0.4% and India up 1.5%.
- The IEA Oil Market Report (17 June 2026) forecast global oil demand to fall by 1.1 mb/d in 2026 after Q2 2026 deliveries plunged 5 mb/d year on year in the Hormuz shock, then to rise by 2 mb/d to 105.3 mb/d in 2027, above the implied 2025 level of about 104.4 mb/d.The Bend [1] IEA Oil Market Report, June 2026
- A Carbon Brief/CREA analysis (4 June 2026) found that China's CO2 emissions rose 2% year on year in Q1 2026, with power-sector emissions up 4% because of 'wasted' (curtailed) wind and solar, though emissions remained below the March 2024 peak.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put total CO2 emissions (fossil + land-use change) at 42.4 GtCO2 in 2024 and a slightly lower preliminary 42.2 GtCO2 in 2025, due to lower net land-use emissions (to about 4.1 GtCO2) mainly attributable to the end of El Niño conditions.↳ confirms 2025-11-13: The Global Carbon Budget 2025 projected CO2 emissions from land-use change at about 4.1… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) found that total CO2 emissions (fossil + land-use change) grew 0.3% per year over 2015-2024, compared with 1.9% per year over 2005-2014.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put global fossil CO2 emissions (fossil fuels + cement, net of cement carbonation) at 38.1 GtCO2 in 2025, up 1.0% (range 0.2% to 1.7%) on 2024 and 'an historical record high', with coal, oil and gas emissions each above 2024 levels (+1.0%, +1.1% and +1.3%).↳ updates 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, projected global fossil… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) estimated China's 2025 fossil CO2 emissions at +0.4% (range -0.1% to +0.9%) on 2024, a plateau rather than a confirmed decline.↳ follows up on 2025-11-13: According to the Global Carbon Budget 2025, China produces 32% of global fossil CO2…↳ follows up on 2025-11-11: Carbon Brief's 11 November 2025 analysis found China's CO2 emissions in 2025 were… (corrected version)
- The IEA Global Energy Review 2026 (April 2026) estimated that global energy-related CO2 emissions (fuel combustion + industrial processes) rose by around 0.4% in 2025 to a record 38,082 MtCO2, 'the slowest rate since 2021', or nearly 38.4 Gt including flaring.
- Ember's Global Electricity Review 2026 (21 April 2026) found that renewables (33.8% share, 10,730 TWh) overtook coal (33.0%, 10,476 TWh) in global electricity generation in 2025.
- Ember's Global Electricity Review 2026 (21 April 2026) found that global fossil power generation fell 0.2% (-38 TWh) in 2025, the first time since 2020 and only the fifth time this century that it did not rise, as low-carbon generation (+887 TWh) outpaced electricity demand growth (+849 TWh).
- Ember's Global Electricity Review 2026 (21 April 2026) found that fossil power generation fell in 2025 in China (-0.9%) and India (-3.3%).
- Carbon Monitor (Nature Reviews Earth & Environment, 14 April 2026) estimated global fossil + industry CO2 emissions at a record 37.2 GtCO2 in 2025, with annual growth flattened to 0.7%, while global power-sector emissions fell 0.9%.
- A Carbon Brief/CREA analysis (12 February 2026) found that China's CO2 emissions fell 0.3% in 2025 (Q4 2025 down 1%) and had been flat or falling for 21 months since March 2024, leaving open whether China has passed its peak.↳ updates 2025-11-11: A Carbon Brief analysis published on 11 November 2025 found that China's CO2 emissions… (corrected version)↳ confirms 2025-11-11: Carbon Brief's 11 November 2025 analysis found China's CO2 emissions in 2025 were… (corrected version)
- Climate scientist Zeke Hausfather wrote on 5 January 2026 that 'global emissions have yet to decline (even if they have plateaued)'.
- The IEA Net Zero Roadmap pathway requires cutting emissions 35% (2°C) or 55% (1.5°C) by 2030 versus 2019, tripling renewables to 11,000 GW by 2030, 4%/yr energy efficiency improvement (versus roughly 2% achieved), a 75% methane cut by 2030 and coal peaking before 2025.
- A Carbon Brief analysis published on 11 November 2025 found that China's CO2 emissions had been flat or falling for 18 months, with China's transport emissions down 5%.The Bend [1] Carbon Brief: Analysis: China’s CO2 emissions have now been flat or falling for 18 months↳ corrects (or adds sources to) an earlier record:
2025-11-11: A Carbon Brief analysis published on 11 November 2025 found that China's CO2 emissions… - China submitted its 2035 climate target at the UN; Climate Action Tracker said the target is unlikely to drive down emissions.↳ corrects (or adds sources to) an earlier record:
2025-09-25: China submitted its 2035 NDC on 25 September 2025, targeting a 7-10% emissions… - A Carbon Brief analysis published 15 May 2025 found China’s CO2 emissions fell 1.6% year-on-year in the first quarter of 2025 as clean power growth drove the decline despite rapid power demand growth.The Bend [1] Carbon Brief: Analysis: Clean energy just put China’s CO2 emissions into reverse for first time [2] Wayback capture 20250514234347↳ corrects (or adds sources to) an earlier record:
2025-05-15: A Carbon Brief analysis published 15 May 2025 found China’s CO2 emissions fell 1.6%… - The EDGAR 2025 report put 2024 global greenhouse-gas emissions at 53.2 GtCO2e, excluding land use, land-use change and forestry.The Balance [1] EDGAR 2025 Report↳ corrects (or adds sources to) an earlier record:
2024: The EDGAR 2025 report put global greenhouse gas emissions (all gases, CO2-equivalent)… - According to IEA-PVPS, about 600 GW of solar PV was added worldwide in 2024, bringing global cumulative solar capacity above 2.2 TW.Net-Zero Transition [1] IEA-PVPS: Snapshot of Global PV Markets 2025↳ corrects (or adds sources to) an earlier record:
2024: According to IEA-PVPS, about 600 GW of solar PV was added worldwide in 2024, bringing… - According to the IEA-PVPS Snapshot of Global PV Markets 2025, China installed 357.3 GW of solar PV in 2024, about 60% of the global total.↳ corrects (or adds sources to) an earlier record:
2024: According to the IEA-PVPS Snapshot of Global PV Markets 2025, China installed 357.3 GW…
Other events (9)
- A Carbon Brief analysis (26 May 2026) found that China appears to have retrospectively changed the scope of its official carbon-intensity metric, so official figures now imply CO2 rose about 7% over 2020-2025 instead of 14% under the previous statistics, a gap of about 700-730 MtCO2 per year.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put the remaining carbon budget for a 50% chance of limiting warming to 1.5°C at 170 GtCO2 (50 GtC) from the start of 2026, about 4 years at 2025 emission levels.↳ updates 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, concluded that the… (corrected version)
- The IEA Global Energy Review 2026 (April 2026) estimated 2025 energy-related CO2 at 12,718 MtCO2 for China (-0.5%) and 3,114 MtCO2 for India (-0.1%, India's first fall on record under normal economic conditions, 'largely due to cyclical factors resulting from the strong monsoon').↳ follows up on 2025-11-11: Carbon Brief's 11 November 2025 analysis found China's CO2 emissions in 2025 were… (corrected version)
- The IEA Global Energy Review 2026 (April 2026) found that energy-related CO2 in advanced economies rose 0.5% in 2025, the first annual increase since 2018 excluding the post-Covid rebound and faster than emerging and developing economies (+0.3%) for the first time since the 1990s, with the US up 2.2% to 4,606 MtCO2 and the EU down 0.8% to 2,372 MtCO2.
- Ember analysis reported by Carbon Brief (28 April 2026) found that even in a worst case global coal power generation would rise by no more than 1.8% in 2026 despite the Iran crisis, with CREA data showing no return to coal as of March 2026.
- +4 older event(s): the section's event log
The Balance: Global net anthropogenic greenhouse gas emissions reach approximately zero.Distant: net GHG ~58 GtCO₂e/yr, +0.3% Jan–Jul 2026 vs ≤2.9 needed; P(by 2100) 28% · assessed 2026-10-18 · 65 event(s)
- Net emissions in GWP100 (the UNFCCC convention), durable removals counted; approximately zero = within ±5% of 2025 gross emissions (about ±3 GtCO₂e).
Net human greenhouse-gas emissions are about 58 billion tonnes of CO₂-equivalent a year and still growing slightly, while durable carbon removal runs at roughly a million tonnes a year. Getting net emissions down to about 2.9 billion tonnes (5% of today's gross emissions) needs a 20-fold cut plus several billion tonnes a year of removals; extrapolating today's slowing emissions growth and a removal industry growing as fast as renewables did gives only a 28% chance by 2100 — less likely than not even then, so red. The answer hinges most on how big and how fast carbon removal can grow, and on whether methane and other gases fall along with CO₂.- ETA
- unlikely by 2100 (P 28%; 10% chance by 2087)
- ETA cruxes
- 1) Removal mapping and scale: renewables-rate transfer (r ×½ → 7%, ×1.5 → 42%), saturation K (4 → 13%, 16 → 42%) and seed (×.1, delivery below capacity as CDR.fyi's ~3%-delivered suggests → 18%); observe audited durable-removal throughput and its growth. 2) Non-CO₂ curvature mapping (m=.5 → 19%, m=0 → 0.5%); CH₄/N₂O inventories. 3) Persistence and form of the fossil-CO₂ deceleration (×2 → 50%, p10 2077; −3%/yr floor → 13%); full-year 2026–27 GCB/EDGAR; the October outlooks argue faster to 2050, slower after. 4) A harmonised net all-GHG inventory with durable removals.
- Path
- Own quantity: global net anthropogenic all-GHG emissions in GWP100 (UNFCCC convention), net land-use CO₂ included, durable removals subtracted (not natural sinks or fossil CCS) ≈ 58.5 GtCO₂e in 2025. Owner's wording (d964534): approximately zero = within ±5% of 2025 gross emissions; 2025 gross = 54.1 (EDGAR excl. LULUCF) + 4.1 (GCB net land use) = 58.2, so the target is net ≤ 2.91 GtCO₂e (reading 'gross excl. LULUCF' gives 2.71: P .27, under 2 years on p10). Drivers: the shared driver fossil-co2-deceleration (GCB fossil curvature, measured median) for CO₂, land use on its OLS trend, non-CO₂ (level 16.0 an assumed EDGAR−GCB cross-inventory estimate) following the fossil curvature by an assumed mapping, its rate model-inferred from TRACE (the fastest same-outcome series); gross emissions floored at hard residuals (assumed); removals from operating storage capacity growing at the renewables rate (assumed transfer) and saturating at an assumed AR6-anchored scale.
- Blockers
- Measurement: no harmonised net all-GHG inventory with durable removals; removal capacity is not audited net delivery (CDR.fyi: ~3% of ~50 Mt contracted CDR delivered).
- Technical/economic: durable removal ~0.0012 Gt/yr of operating capacity; 2026 CDR offtake value on track to halve; Kawasaki dropped DAC expansion; Bison paused; Mammoth underperforming; removal scale bounded by biomass/land, storage build-out and energy.
- Policy: US withdrawal from Paris and power-plant standards repealed, opposition to the IMO net-zero framework, banks leaving NZBA, EU methane-rule MRV delay proposed (no gate is assumed away; no policy lag added).
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (shared driver fossil-co2-deceleration for fossil-CO₂ curvature; own land-use OLS trend and TRACE-implied non-CO₂) with consistent net flows, plus a removal S-curve (renewables-rate reference class, AR6-anchored saturation) and named failure scenarios; one Monte Carlo distribution
- ETA calculation
- Shared driver fossil-co2-deceleration (drivers.yaml): draw i of `ledger.py driver-draws fossil-co2-deceleration --as-of 2026-10-18` (seed 20261015, n 10,000; byte-identical to the 2026-10-15 draws; a_per_yr2 lognormal p50 .00080015 = (.017320−.009319)/10 from co2-emissions-fossil-gcb 1995/2005/2015/2025 = 23.26/29.19/34.71/38.1 Gt, sdlog .50; g_2020 .0093186 and fossil_2025 38.1 fixed) is the fossil curvature in simulation draw i: fossil rate +0.452%/yr at 2026.5 centrally. Own mapping (prev's, unchanged, rule 7): net land use 2006–2025 OLS −2.203%/yr, level 4.443 (2025 obs 4.1, 2026-05-13-gcb-2025-total-co2-slightly-lower); non-CO₂ 16.0 GtCO₂e = assumed cross-inventory estimate EDGAR 54.1 (2026-09-01-edgar-2026-ghg-2025) − GCB 38.1, rate +0.614%/yr model-inferred, curvature m=1 × driver curvature (assumed mapping); per-draw common rate shift pins all-GHG 2026 to Climate TRACE Jan–Jul +0.3% ± 0.3 (2026-09-24-climate-trace-jul-2026-ghg; δ=0 centrally). 10,000 common draws; quantile year = calendar year containing it. Prev's script (notes/climate/20261015T113313Z.md) reproduced byte-identically on the 2026-10-18 series; no input changed (58 new effective events tested, none measures a modelled quantity). Removals: S-curve, seed .0012 Gt/yr (2026-09-18-greensand-eu-first-offshore-co2-storage, 2026-09-21-adm-columbus-beccs-removal-credits), r=ln(5.15/4.457)=.14452±.05 (2026-09-21-irena-tripling-tracker-693gw-2025; transfer assumed), K~LN(8,.54) Gt/yr (assumed; AR6 WG3 C.3.5 ≈3–16 Gt/yr); floors Rc U(1,4), RN U(3,4) (assumed). Failure scenarios: S1 removal stall .10 — reference class policy-revenue-only abatement services (first unit by 1996; stall = deployment ≤10× first unit after 30 y): FGD, SCR, catalytic converters, CFC substitution, landfill gas: 0 of 5 stalls (assumed classification), Jeffreys .083 / Laplace .143; climate-only CCS as a stall 1/6 → .21–.25 (the .25 row); new market signals both ways (2026-10-02-qcintel-cdr-offtake-value-halving-2026, 2026-10-06-kawasaki-abandons-dac-plans vs 2026-10-06-cdrfyi-sept-2026-over-1mt-contracted, 2026-10-15-frontier-4401-in-situ-mineralization-offtake), no member outcome: .10 kept (sens .05/.25: Balance P .296/.233, Ceiling .184/.178); S2 decarbonisation stall .05 (assumed, <10%). Seed evidence new both ways: Gevo ND BECCS >700 kt stored since capture began (≈.00016 Gt/yr, start date assumed) and CDR.fyi ~3% of ~50 Mt contracted delivered (2026-10-08-gevo-north-dakota-beccs-10000-delivered): +Gevo row Balance .287 / Ceiling .184, ×.1 row .182 / .170; kept (log-SD 1). Balance: first t with max(Rc,G)+max(RN,N)−C ≤ 0.05×2025 gross (54.1 EDGAR + 4.1 land use = 58.2) = 2.91 GtCO₂e. Gap: net 58.5 → ≤2.91 = 20×; removals .0012 → 3.87/6.87/13.39 Gt/yr at balance (ln(6.87/.0012)/.14452 = 60 y unslowed). Deterministic central: 2100 gross CO₂ 7.0 + non-CO₂ floor 3.5 = 10.5 needs C ≥ 7.6 vs S-curve 6.9 → after 2100. MC: P .2820 → .28, p10 2087.5 → 2087, p50/p90 null (identical to prev); conditional-success median 2090.6 (not p50); P(CO₂ net zero by 2100) .316. Sensitivities (P): curvature ×½/×2 .066/.499 (p10 2077.5); total-CO₂ curvature .463; rate floor −3%/−5% .129/.255; m=.5/0 .189/.005; r ×½/×1.5 .070/.423; K 4/16 .130/.417; unbounded exponential .585 (p50 2094, needs 17/36 Gt/yr removals p50/p90, above AR6; held out); seed ×10/+Gevo/×.1 .365/.287/.182; TRACE +0.5%/−0.3% .249/.376; tolerance 5% of 54.1 .274; ≤0 .162. Outlook cross-check (projections, not inputs): model central fossil 2050/2025 = 0.896 (38.1 → 34.14 Gt), cumulative fossil 2026–2050 901 Gt; curvature each outlook implies in the driver form, ×driver p50 (driver CDF): Exxon 0.83 ×1.23 (.65), IRENA WETO PES 0.74 ×1.57 (.81; 2026-10-13-irena-weto-2026-net-zero-only-in-1-5c), DNV 22/39 ×2.37 (.96; 2026-10-07-dnv-eto-2026-energy-co2-likely-peaked), Rystad ~0.5 ×2.72 (.98; 2026-10-15-rystad-ges-2026-fossil-co2-plateau-40gt); they decline faster to 2050 but flatten after (DNV net-zero CO₂ 2098), i.e. the a×2 and −3%/yr-floor rows at once; driver-fit remark for the owner in the notes, driver not refitted. Cross-checks: AR6 WG3 C.2.4 ~30% of C3 reach net-zero GHG by 2100; DNV net-zero CO₂ 2098 and IRENA PES (net zero only in its 1.5°C scenario) have no balance by 2100 centrally; RMI announced-project growth 51%/yr (2026-10-08-rmi-new-frontiers-cdr-51pct-growth, broader scope) discounted ×1.5–3 in time = .137–.275/yr vs r .1445 (low end); discounted ×1.5–3 from 2026.80: Climate Pledge 2040 → 2046.6–2066.4 (2026-09-15-climate-pledge-report-2026-700-signatories), Germany 2045 → 2054.1–2081.4 (2026-09-23-germany-fossil-fuel-transition-roadmap), IMO 2050 → 2061.6–2096.4 (partial populations). Prev P .28, p10 2087: unchanged. No external anchors listed. Foreign context only; no mapping.
- Change
- Re-assessment on 58 new effective events (published 2026-09-28…10-18). Prev's script reproduced byte-identically on byte-identical series and driver draws. Inputs tested against the new evidence: the removal seed has new evidence both ways (Gevo's ≈.00016 Gt/yr storage would add ~13%; CDR.fyi's ~3%-delivered points to throughput below capacity), both inside its log-SD 1, so the measured proxy .0012 stays (rows +Gevo .287, ×.1 .182); RMI's proponent growth puts r at the low end of its discounted range, not outside it; removal-market and policy signals give S1/S2 no member outcome; outlooks are projections (cross-check). No input changed (rule 7). P .2820, p10 2087.5: no move; status red unchanged.
- Rubric rule
- RUBRIC.md v6 §4/§4.1: red, p_achieved < 0.5 so the unconditional median is null; no verified achievement of global net all-GHG within ±5% of 2025 gross.
by assess:20261018T192714Z · 54 evidence event(s)Evidence and news: 54 cited by the assessment, 11 other event(s)
Cited evidence (54)
- As seen on 18 October 2026 (tracker last updated 8 October), the Climate Action Tracker's 2035 Climate Target Update Tracker counted 153 countries that had submitted a 2035 NDC target since November 2024 (a count that lists the USA among submitters), covering 89.8% of global emissions and 86.9% of global population, and 41 countries that had not, including CAT-analysed Argentina, Egypt, Iran and Viet Nam, with only 3 submitted NDCs (Nigeria, Norway and the United Kingdom) rated 1.5°C-compatible against modelled domestic pathways; the UNFCCC NDC registry lists Chad (5 October), Grenada (7 October) and Equatorial Guinea (15 October 2026) as the latest submissions.
- France and Germany announced on 16 October 2026, on the sidelines of the EU summit in Brussels, a joint position on the EU automotive package under which the 2035 CO2 fleet target for new cars would be lowered from a 100% to a 90% reduction with no requirement for carmakers to offset the remaining 10% (going beyond the Commission's December 2025 proposal, which also set 90% but with offsetting), in exchange for stronger 'Made in Europe' requirements, with Chancellor Friedrich Merz declaring the combustion-engine phase-out 'history'.Net-Zero Transition [1] Bundesregierung: Joint Franco-German Press Statement on the Future of the European Automotive Industry [2] Deutschlandfunk: Merz: 'Verbrenner-Aus ist Geschichte' (16 October 2026) [3] electrive: France and Germany propose 90 per cent CO2 cut for new cars by 2035 (16 October 2026)
- At the 15–16 October 2026 European Council, EU leaders did not agree to the postponement of the ETS2 carbon market for road-transport and building fuels (due to start in 2028 after an earlier one-year delay) until 2030 sought by the Visegrad countries and Italy, instead calling on the Commission to update its assessment of all impacts of ETS2, in particular its design and safeguards, and urging co-legislators to speed up reform of the EU ETS market stability reserve to curb carbon-price volatility.Net-Zero Transition [1] Il Sole 24 Ore: EU summit: leaders say 'The Commission should assess the impact of the ETS 2 system' (16 October 2026) [2] EU Today: EU resists further delay to carbon pricing as energy costs divide European leaders (17 October 2026) [3] Politico Europe: EU leaders' summit descends into 3-hour argument over taxing car emissions (15 October 2026)
- Computed from the SBTi's by-company target data file (records updated to 15 October 2026), 2,848 organisations have validated net-zero targets (2,126 corporates, 718 SMEs and 4 financial institutions) and a further 994 have net-zero commitments, while 12,162 have validated near-term targets, among 15,816 listed organisations.
- Rystad Energy's Global Energy Scenarios 2026 (15 October 2026) said global fossil CO2 emissions are plateauing at around 40 Gt per year and projected that in its new 'Pragmatism' reference scenario they fall to about 20 Gt by 2050 (800 Gt cumulative over 2026-2050, about 0.35°C of additional warming), versus about 2 Gt in 'Climate first' and near today's level for longer in 'Fossil first'.
- Frontier buyers (including Stripe, Google, Shopify, Autodesk, H&M Group and Workday) signed offtake agreements, announced on 15 October 2026, to pay 44.01 $11.7 million to permanently remove 20,932 tonnes of CO2 between 2028 and 2031 by injecting dissolved biogenic CO2 from wastewater-treatment biogas upgrading into peridotite rock, where it mineralises within 1-3 years; it is Frontier's first purchase of in-situ mineralization removal and its first in the Middle East, and 44.01 calls it the first commercial-scale project of its kind, moving from short-duration injection pilots towards continuous operation over multiple years.Permanent Removal [1] Frontier: 44.01 and Frontier advance permanent CO₂ storage through in-situ mineralization (15 October 2026) [2] 44.01: Frontier backs 44.01 to take in-situ CO₂ mineralisation to commercial scale (15 October 2026) [3] ESG Dive: Frontier Climate buyers ink deal to store CO2 as solid rock (16 October 2026)
- The European Commission formally proposed on 14 October 2026 a one-year postponement of the EU Methane Regulation's monitoring, reporting and verification obligations for importers of oil, gas and coal (and of the methane-intensity reporting provisions for importers and EU producers), citing security-of-supply pressure from the Middle East conflict while saying the change does not affect the regulation's ambition; the proposal, which needs European Parliament and Council approval, would move the importer requirements from 2027 to 2028.Beyond the frameworkNet-Zero Transition [1] European Commission: Commission proposes flexibility to methane importers to help save costs (IP/26/2119) [2] Rapporteur: Brussels moves to postpone methane reporting rules (14 October 2026) [3] Carbon Brief Daily Briefing (15 October 2026): postpone methane rules
- IRENA's World Energy Transition Outlook 2026 (launched 13 October 2026) projected that under existing government plans (its Planned Energy Scenario) global fossil CO2 emissions fall only 26% by 2050 (0.2% a year over 2025-2030 and 1.4% a year thereafter), with cumulative 2026-2050 energy-related CO2 of about 840 GtCO2 and warming likely at least 2°C, that its Decarbonising Energy Scenario (no longer aligned with the net-zero-by-2050 goal of earlier editions) still leaves about 4 GtCO2 of net fossil CO2 in 2050 (about 480 GtCO2 cumulative), and that only its revised 1.5°C Scenario, with a limited and temporary overshoot and more carbon capture including BECCS, reaches net-zero fossil CO2 in 2050 (cumulative slightly below 400 GtCO2).
- The Forest Declaration Assessment 2026 (13 October 2026) found that global deforestation fell 15% in 2025 to just over 7 million hectares (releasing an estimated 3.9 billion tonnes CO2e), the largest annual decline since the 2018–20 baseline, but that the world remained 70% off track from the pathway to halt deforestation by 2030 (a cumulative 10.7 Mha excess since 2021), with humid tropical primary forest loss down 36% from 2024's fire-driven spike yet still 11% above the 2018–20 baseline, deforestation in tropical Africa still rising (2.4 Mha in 2025), and an estimated 11.8 Mha of forest under active restoration.
- The Climate Action Tracker's US update (13 October 2026) kept its overall US rating at 'Critically insufficient' and projected that under the Trump Administration's policies (including the One Big Beautiful Bill Act and the rescission of the Endangerment Finding and of power-plant and vehicle emission standards) US greenhouse-gas emissions excluding LULUCF will still fall but only to 20-25% below 2005 levels by 2030, versus about 30-40% on the Biden Administration's trajectory (about 600-1,000 MtCO2e, or 11-23%, higher in 2030), with the US exit from the Paris Agreement and move to leave the UNFCCC having nullified its 2030, 2035 and 2050 (net-zero) targets.
- Parties to the London Convention and London Protocol meeting at the IMO (LC 48/LP 21, 5-9 October 2026) adopted a resolution establishing assessment frameworks for ocean fertilization, ocean alkalinity enhancement and ocean sinking of biomass, under which proposals are assessed case by case to show legitimate scientific research purpose and acceptable environmental risk, with the Parties stating that further research is needed before any activities beyond the research stage can be considered.
- RMI's New Frontiers Report for Carbon Dioxide Removal (8 October 2026), covering 14 CDR approaches including nature-based living-biomass removal, estimated end-of-2026 deployed CDR capacity at about 42 MtCO2 per year with $13.6 billion of funding raised across 514 tracked companies and 694 active or announced projects, and found that the 217 MtCO2 per year of 2030 supply implied by announced projects and company targets would require roughly 51% annual capacity growth.
- Gevo announced on 8 October 2026 the sale and delivery of 10,000 carbon dioxide removal credits through ClimeFi to a corporate buyer from its Gevo North Dakota ethanol BECCS facility, where fermentation CO2 is stored geologically more than a mile beneath the plant and which Gevo said has generated more than 700,000 tonnes of CO2 removal since capture began, while citing CDR.fyi data that only about 3% of roughly 50 million tonnes of contracted CDR has been delivered market-wide.
- DNV's Energy Transition Outlook 2026 (launched 7 October 2026) judged that global energy-related CO2 emissions 'likely peaked in the last 24 months' and that total anthropogenic CO2 emissions (energy, industrial processes and land-use change, net of direct air capture) peaked in 2024 at 43 Gt, forecasting energy-related CO2 falling 44% from 39 Gt today to 22 Gt in 2050 (and 65% to 12 Gt in 2060), with net-zero CO2 only in 2098 and end-of-century warming of about 2.3°C.
- ClimeFi's CDR Market Review for Q3 2026 (7 October 2026) found that enhanced rock weathering made up a majority of a quarter's durable-CDR purchase volume for the first time, driven by the largest ERW deal on record, while biochar purchases (above 1 MtCO2 in each of the two previous quarters) fell to their lowest since Q1 2023, Microsoft's buying slowed to about 2% of Q3 volume with buyers other than Microsoft accounting for 64% of 2026 volume to date, and debt made up 94% of all disclosed Q3 private funding.
- Japanese industrial group Kawasaki Heavy Industries abandoned its direct air capture expansion plans and shifted its focus to CO2 capture, utilisation and storage, Quantum Commodity Intelligence reported on 6 October 2026.
- CDR.fyi's September 2026 market recap (6 October 2026) reported that more than 1 million tonnes of durable carbon removal were contracted in September 2026, led by Google's 1 Mt enhanced-weathering purchase from Terradot, up from about 128,000 tonnes contracted in August 2026.
- Quantum Commodity Intelligence's monthly carbon-removal offtake tracking (2 October 2026) found the value of CDR offtake agreements on track to more than halve in 2026 compared with 2025, although September 2026 offered a partial reprieve with 21 offtake deals worth an estimated $148 million.
- A coalition of 26 US states, counties and cities led by New York Attorney General Letitia James petitioned the US Court of Appeals for the D.C. Circuit on 1 October 2026 to overturn the EPA's September 2026 repeal of the 2024 Carbon Pollution Standards (greenhouse gas limits for coal- and gas-fired power plants, a sector producing about a quarter of US CO2 emissions), arguing the repeal was arbitrary and capricious, and also filed a notice of intent to sue the EPA for failing to regulate greenhouse gases from existing gas-fired plants.Beyond the framework [1] Attorney General Tong Challenges EPA Rollback of Power Plant Pollution Limits (Connecticut AG, 1 Oct 2026) [2] Attorney General James Challenges EPA Rollback of Power Plant Pollution Limits (New York AG) [3] Daily Briefing: Diesel 'crisis' talks | US states sue EPA (Carbon Brief, 2 Oct 2026)
- Swedish utility Öresundskraft took a final investment decision on carbon capture at its Filbornaverket waste-to-energy plant in Helsingborg, described as the first EU waste-to-energy plant to reach FID for the full capture, transport and storage chain, and signed an agreement (announced 1 October 2026) with Northern Lights to ship and permanently store up to 200,000 tCO2 per year (both biogenic and fossil CO2) beneath the Norwegian North Sea from a targeted start-up in Q4 2029.
- The EDGAR 2026 report (September 2026) put total greenhouse-gas emissions excluding LULUCF at 54.1 GtCO2e in 2025, up 0.7% on 2024, with China up 0.1% and the US up 2.2%.↳ follows up on 2024: The EDGAR 2025 report put 2024 global greenhouse-gas emissions at 53.2 GtCO2e,… (corrected version)
- The Integrity Council for the Voluntary Carbon Market approved (announced by Verra on 28 September 2026) Verra's VM0049 Carbon Capture and Storage v1.0 methodology and its accompanying modules, including direct air capture (VMD0056) and bioenergy with carbon capture and storage (VMD0059) plus CO2 transport and geological storage, under its Core Carbon Principles, so that DACCS and BECCS removal credits issued under VM0049 can qualify for the CCP high-integrity label.
- A European Commission spokesperson said on 28 September 2026 that the Commission plans to propose postponing the entry into force of the EU Methane Regulation (length not specified), after pressure from governments including France and the United States and after Energy Commissioner Dan Jørgensen said on 25 September that Brussels was exploring a one-year delay of the rules for imported oil, gas and coal, whose monitoring and reporting obligations for importers are due to take effect on 1 January 2027.Beyond the frameworkNet-Zero Transition [1] Reuters: EU to delay methane emissions rules after pushback from governments (28 Sep 2026) [2] Argus: EU plans one-year delay to methane law for oil, gas (28 Sep 2026) [3] eKathimerini (Reuters syndication, Kate Abnett): EU to delay methane emissions rules after pushback from governments (28 Sep 2026)
- Climate TRACE (24 September 2026) estimated global greenhouse-gas emissions of 4.9 GtCO2e in July 2026, up 0.1% on July 2025, and 33.7 GtCO2e year-to-date for January-July 2026, 0.3% above the same period of 2025, with July emissions down 0.3% in China and 2.6% in the US but up 3.4% in India.
- The Government of Canada said on 24 September 2026 that it will develop a policy framework under Paris Agreement Article 6 to let Canadian companies, including carbon removal suppliers, trade internationally transferred mitigation outcomes (ITMOs), which it said could contribute billions of dollars to the Canadian economy and catalyse domestic climate mitigation projects and technologies.Permanent Removal [1] Canada to develop carbon removal market framework under Paris agreement (Reuters) [2] Canada Takes First Steps to Enter International Carbon Market (ESG Today)↳ follows up on 2024-11: COP29 in November 2024 finalised the Paris Agreement Article 6 rules for international… (corrected version)
- Germany's federal cabinet approved, and Environment Minister Carsten Schneider presented at the UN in New York on 23 September 2026, a national roadmap for transitioning away from coal, oil and gas by 2045 (Germany's climate-neutrality target year), making Germany the third country after France and the Netherlands to publish such a roadmap; it reaffirms existing goals such as a coal exit no later than 2038 and at least 80% renewable electricity by 2030 (versus about 55% of gross electricity consumption now), with fossil fuels still 65% of German energy consumption in 2024.
- In his 22 September 2026 address to the UN General Assembly, US President Donald Trump attacked the IMO Net-Zero Framework (the proposed global marine-fuel greenhouse-gas-intensity standard and emissions-pricing mechanism meant to deliver the IMO goal of net-zero international shipping emissions by or around 2050) as a 'global carbon tax' and vowed 'there will be no global taxes' while he is president, signalling continued US opposition ahead of the framework's expected adoption vote at an extraordinary MEPC session on 4 December 2026, after US pressure had already postponed adoption by a year in October 2025.The BalanceNet-Zero Transition [1] Remarks to the United Nations General Assembly in New York City, September 22, 2026 (American Presidency Project transcript) [2] Trump Takes Aim at Shipping's Net-Zero Plan at UN (gCaptain, 23 Sep 2026) [3] Competing Climate Visions Clash at UN General Assembly (Inside Climate News, 22 Sep 2026)
- The IEA's Electrification special report (22 September 2026), requested by COP31 hosts Türkiye and Australia, found that electricity supplied about 23% of global final energy consumption in 2025 (up from 17% in 2000), that fully exploiting today's cost-competitive electrification potential would raise this to about 33%, and that reaching the COP31 presidency's goal of 35% by 2035 (achieved in its High Electrification Scenario) requires going beyond today's policy settings, under which the share rises to slightly less than 30% by 2035.
- IRENA's third UAE Consensus tracking report, released with the COP31 Presidency and the Global Renewables Alliance on 21 September 2026, found that a record 693 GW of renewable power capacity was added worldwide in 2025 (513 GW solar PV, 158 GW wind), bringing total installed renewable capacity to about 5.15 TW, but that tripling to about 11.2 TW by 2030 now requires average net additions of about 1.2 TW per year in 2026-2030, and that sustaining the 2025 growth rate would leave a shortfall of about 0.6 TW in 2030 (down from 0.9 TW in the previous edition).Net-Zero Transition [1] IRENA et al. (2026), Delivering on the UAE Consensus: Tracking progress toward tripling renewable energy capacity and doubling energy efficiency by 2030 [2] Global renewable deployment must double to hit 2030 climate target, report says (Reuters, 21 Sep 2026) [3] Report launched at the Global Renewables Summit highlights renewable energy progress but warns of challenges ahead to 2030 (ALER)
- IRENA's UAE Consensus tracking report (21 September 2026) found that global energy intensity improved by only 1.7% in 2025 (after 1.1% in both 2023 and 2024), well below the 4% per year doubling-efficiency goal, so that meeting the 2030 goal now requires 5.6% improvement per year, while grid and flexibility investment of up to about $1 trillion per year is needed in 2026-2030 versus $525 billion in 2025.
- ADM announced on 21 September 2026 its planned entry into the voluntary carbon removal market with credits from the 800,000+ tCO2-per-year biogenic (ethanol fermentation) carbon capture operation at its Columbus, Nebraska corn processing complex, whose CO2 is stored permanently in Tallgrass's Eastern Wyoming Sequestration Hub; the credits are undergoing Puro.earth certification, with initial issuance expected by the end of 2026 and a 15-year crediting period.
- Greensand, led by INEOS Energy with Harbour Energy and Nordsøfonden, entered commercial operation on 18 September 2026 as the EU's first full-scale offshore CO2 storage site, injecting CO2 sourced primarily from Danish biomethane plants into the depleted Nini West field about 1,800 m beneath the Danish North Sea, with capacity of up to 400,000 tCO2 per year in its first phase and potential expansion to 4-8 MtCO2 per year.Permanent RemovalNet-Zero Transition [1] INEOS opens the first full-scale CO2 storage site in the EU (INEOS)
- ExxonMobil's 2026 Global Outlook (17 September 2026) projected global energy-related CO2 emissions to fall about 20% from 36 billion tonnes in 2025 to 30 billion tonnes in 2050, up from the 27 billion tonnes it projected for 2050 a year earlier and far above the roughly 11 billion tonnes in IPCC 'likely below 2°C' scenarios, with coal still 15% of the 2050 energy mix and carbon capture and storage reaching 2 billion tonnes per year (down from its prior 3.1 billion).
- Argus data (16 September 2026) showed technology-based CO2 removal credits contracted through long-term offtake agreements fell to about 6 million in January-August 2026 from 24 million a year earlier, with BECCS credits down to 2 million from 17 million and Microsoft's disclosed buying down to just over 2 million from 16.5 million, while biochar offtakes rose to nearly 3 million from 1.5 million.
- The Climate Pledge 2026 report (Amazon and Global Optimism, 15 September 2026) said more than 700 companies across 62 industries and 49 countries and territories, with combined revenue of $3.8 trillion, have committed to reach net-zero carbon emissions by 2040, after 107 new signatories joined in 2025 (19% signatory growth), and that analysed signatories cut their self-reported Scope 1 and 2 emissions by 11% on average (about 14 MtCO2e over 2021-2024) versus 7% for non-signatory companies.
- Climate TRACE (27 August 2026) estimated total greenhouse-gas emissions (CO2e) of 29.7 GtCO2e in H1 2026, up 0.2% on H1 2025, with China down 0.3%, the US down 0.4% and India up 1.5%.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put total CO2 emissions (fossil + land-use change) at 42.4 GtCO2 in 2024 and a slightly lower preliminary 42.2 GtCO2 in 2025, due to lower net land-use emissions (to about 4.1 GtCO2) mainly attributable to the end of El Niño conditions.↳ confirms 2025-11-13: The Global Carbon Budget 2025 projected CO2 emissions from land-use change at about 4.1… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) found that total CO2 emissions (fossil + land-use change) grew 0.3% per year over 2015-2024, compared with 1.9% per year over 2005-2014.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put global fossil CO2 emissions (fossil fuels + cement, net of cement carbonation) at 38.1 GtCO2 in 2025, up 1.0% (range 0.2% to 1.7%) on 2024 and 'an historical record high', with coal, oil and gas emissions each above 2024 levels (+1.0%, +1.1% and +1.3%).↳ updates 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, projected global fossil… (corrected version)
- The SBTi Trend Tracker 2025 reported that the cumulative number of companies with validated net-zero targets had trebled by the end of Q2 2025 compared with the end of 2023.The Balance [1] SBTi: Trend Tracker 2025↳ corrects (or adds sources to) an earlier record:
2025: The SBTi Trend Tracker 2025 reported 227% growth in corporate SBTi targets. - The SBTi dashboard listed 12,353 companies with targets or commitments, including 2,325 companies with net-zero targets.↳ corrects (or adds sources to) an earlier record:
2025: The SBTi target dashboard listed about 11,000 companies with science-based targets,… - CarbonCapture Inc. paused development of Project Bison in Wyoming (original target 5 MtCO2 per year of direct air capture), relocating due to competition from data centers.Permanent Removal [1] Carbon Herald: CarbonCapture pauses Project Bison↳ corrects (or adds sources to) an earlier record:
2025: CarbonCapture Inc. paused development of Project Bison in Wyoming (original target 5… - The Net Zero Stocktake 2025 found that net-zero targets cover 77% of global GDP, or 83% including US state-level targets.The Balance [1] Net Zero Stocktake 2025↳ corrects (or adds sources to) an earlier record:
2025: The Net Zero Stocktake 2025 found that net-zero targets cover 77% of global GDP, or 83%… - A 2025 Nature Communications study revised global reforestation potential down to 6-195 Mha (53-2,225 TgCO2e per year of sequestration) after sustainability safeguards, versus previous estimates of more than 1,000 Mha.Net-Zero Transition [1] Nature Communications: reforestation potential with safeguards↳ corrects (or adds sources to) an earlier record:
2025: A 2025 Nature Communications study revised global reforestation potential down to 6-195… - Climeworks’ Mammoth plant in Iceland has a nameplate capacity of 36,000 tons of removal per year, but its collector containers were not operating as efficiently as prototype testing indicated.Permanent Removal [1] Climeworks: Mammoth [2] Latitude Media: On the ground at Climeworks’ biggest DAC project↳ corrects (or adds sources to) an earlier record:
2025: Climeworks' Mammoth direct air capture plant in Iceland (36,000 tCO2 per year nameplate… - CDR.fyi reported $2.3 billion of private direct air capture investment from 2021 to H1 2025, but only about 1,186 tonnes of DAC removal delivered out of 2.47 million tonnes contracted.Permanent Removal [1] CDR.fyi DAC Market Snapshot 2025↳ corrects (or adds sources to) an earlier record:
2025: CDR.fyi reported $2.3 billion of private direct air capture investment from 2021 to H1… - CDR.fyi's 2025 DAC market snapshot put operating direct air capture capacity at about 0.01 MtCO2 per year and total CCS capacity (all types) at 41 MtCO2 per year, versus roughly 10,000 MtCO2 per year needed.Permanent Removal [1] CDR.fyi DAC Market Snapshot 2025
- Major banks including JP Morgan, HSBC and Barclays exited the Net-Zero Banking Alliance, as reported by The Guardian in December 2025.↳ corrects (or adds sources to) an earlier record:
2025-12: Major banks including JP Morgan, HSBC and Barclays exited the Net-Zero Banking… - In January 2025 the United States withdrew from the Paris Agreement and abandoned its federal net-zero target.
- The EDGAR 2025 report put 2024 global greenhouse-gas emissions at 53.2 GtCO2e, excluding land use, land-use change and forestry.The Balance [1] EDGAR 2025 Report↳ corrects (or adds sources to) an earlier record:
2024: The EDGAR 2025 report put global greenhouse gas emissions (all gases, CO2-equivalent)… - BloombergNEF estimated that a net-zero pathway requires about $5.6 trillion per year of energy transition investment, versus $2.1 trillion invested in 2024.
- BloombergNEF reported record global energy transition investment of $2.1 trillion in 2024, up 11% on the previous year.Net-Zero Transition [1] BloombergNEF: Energy Transition Investment Trends [2] Wayback capture 20260103182250↳ corrects (or adds sources to) an earlier record:
2024: BloombergNEF reported record global energy transition investment of $2.1 trillion in… - COP29 in November 2024 agreed a climate finance goal of $300 billion per year for developing countries by 2035 (up from $100 billion), within a broader goal of $1.3 trillion per year.Net-Zero TransitionThe Balance [1] UNFCCC: COP29 UN Climate Conference Agrees to Triple Finance to Developing Countries, Protecting Lives and Livelihoods [2] Wayback capture 20260103225238↳ corrects (or adds sources to) an earlier record:
2024-11: COP29 in November 2024 agreed a climate finance goal of $300 billion per year for… - COP29 in November 2024 finalised the Paris Agreement Article 6 rules for international carbon markets after a decade of negotiations.↳ corrects (or adds sources to) an earlier record:
2024-11: COP29 in November 2024 finalised the Paris Agreement Article 6 rules for international…
Other events (11)
- The Climate Action Tracker (13 October 2026) downgraded Canada's 'Policies and action' rating to 'Critically insufficient' (overall rating unchanged at 'Highly insufficient'), projecting Canada's 2030 emissions at 662 MtCO2e excluding LULUCF, about 13% below 2005 under current policies and 14% with planned policies, versus a 30% cut projected for planned policies two years earlier and a 2030 NDC target of 40-45% below 2005, after the Carney government removed the consumer carbon price, cancelled the planned oil and gas emissions cap and repealed the Electric Vehicle Availability Standard, and noted that Prime Minister Mark Carney has admitted Canada will not meet its 2030 and 2035 targets under his plan.
- At the Pacific Pre-COP31 Leaders' event in Tuvalu and Fiji on 6 October 2026, 27 listed signatories including Australia, Türkiye, France, New Zealand, Indonesia, Singapore and the UAE adopted the Taku Pakasoa Declaration, which urges Parties that have yet to submit their 2035 NDCs to do so as soon as possible and reaffirms the COP28 call for transitioning away from fossil fuels in energy systems 'to achieve net zero by 2050', while Pacific Island ministers at the Pre-COP discussed developing a regional roadmap to transition away from fossil fuels.The Balance [1] COP31 Pacific: The Taku Pakasoa Declaration: Strength in Unity (6 October 2026) [2] BOGA (9 October 2026): Pacific Island leaders discuss developing a regional Pacific roadmap to transition away from fossil fuels at pre-COP [3] E3G (9 October 2026): Positive tone for cooperation at Pre-COP – setting the stage for COP31
- The Climate Action Tracker's Indonesia update (28 September 2026) kept Indonesia's overall rating at 'Critically insufficient', finding that its 2035 NDC improves transparency but does not raise real ambition because current policies already overachieve it, that Indonesia's emissions gap to a 1.5°C-compatible pathway grows from 112% in 2030 to 141% in 2035 (median of the 2035 targets), and that emissions excluding LULUCF under the unconditional scenario keep rising to a 2050 peak at more than double 2019 levels, while Indonesia retains its net-zero-by-2060 target.
- At the UN Climate Summit 2026 in New York on 23 September 2026, UN Secretary-General António Guterres, saying greenhouse gas emissions are still rising and the world is expected to temporarily overshoot 1.5°C, called on every country to develop a credible roadmap with clear timelines to transition away from fossil fuels, with the G20 (about 80% of global emissions) to lead, and the summit launched a Global Grids Accelerator (citing more than 2,500 GW of renewable projects waiting for grid connections worldwide) and a critical-minerals initiative.
- Net Zero Tracker listed 138 of 198 countries with net-zero targets; the fetched page did not specify the number with legally binding national targets.The Balance [1] Net Zero Tracker↳ corrects (or adds sources to) an earlier record:
2025: Net Zero Tracker counted 137 of 198 countries with net-zero targets, of which 35… - +6 older event(s): the section's event log
The Ceiling: Peak warming held below +2°C.Distant: ~1.3–1.4°C long-term, rising; must stop <2°C and hold 10 yrs; P(by 2100) 18% · assessed 2026-10-18 · 61 event(s)
- Warming = the long-term human-induced warming indicator (IPCC method, vs 1850–1900). Achieved once it has stopped rising below +2.0°C and stayed at or below that peak for 10 years.
Long-term human-caused warming is about 1.3–1.4°C and still rising. On today's emissions trend the remaining carbon budget that gives an even chance of staying under 2°C (about 1,020 billion tonnes of CO₂) runs out around 2051, so a model of cumulative CO₂ plus the shorter-lived gases gives only an 18% chance that warming peaks below 2°C and holds there for ten years by 2100: red. Most of the successful paths rely on methane falling fast enough to hold temperatures flat before CO₂ reaches net zero, which is the biggest uncertainty; this year's El Niño-boosted record months are weather on top of the long-term trend, not the measure itself.- ETA
- unlikely by 2100 (P 18%; 10% chance by 2082)
- ETA cruxes
- 1) Non-CO₂ curvature mapping and how far a CH₄ decline can hold the indicator flat for 10 years (m=.5 → 10%, m=0 → 0.06%; GWP* ×.7 → 16%; f_s .5 → 15%); CH₄/N₂O trends (EU methane-rule delay vs pledge delivery). 2) Persistence and form of the fossil-CO₂ deceleration (×2 → 37%, p10 2064; −3%/yr floor → 10%); full-year 2026–27 CO₂ through El Niño and the oil rebound. 3) A newer IPCC-method long-term warming estimate (IGCC) and an updated 2°C budget (AR7 timeline unresolved); whether the observed 0.36°C/decade reflects aerosol unmasking beyond the budget (aerosol +.05°C → 16%). 4) Removal scale (r ×1.5 → 21%). 5) Governed SRM beyond simulations.
- Path
- Own quantity: the long-term human-induced warming indicator (IPCC method, vs 1850–1900), ~1.3–1.4°C and rising. Framework detail: achieved once it has stopped rising below +2.0°C and stayed at or below that peak for 10 years; my reading adds that it must not have exceeded 2.0°C before (an overshoot then a lower peak is not 'peak warming held below 2°C'; relaxing it: P .21), and does not require permanence after the 10 years (achievement is terminal). Modelled with an emulator, not equated with CO₂ neutrality: warming change ∝ cumulative net CO₂ + GWP* warming equivalent of the non-CO₂ trajectory, compared with the AR6 2°C budget (sourced, rebased; GCB revision transfer assumed) and the AR6 WG3 non-CO₂ pathway it is conditioned on. The mapping of CH₄ declines onto a 10-year plateau of the indicator is assumed (GWP* coefficients sourced, aggregate application assumed). Faster temperature series (Berkeley monthly) are El Niño-boosted observations, context only. Fossil-CO₂ curvature comes from the shared driver fossil-co2-deceleration (draw i).
- Blockers
- Physical: CO₂ emissions still ~42 Gt/yr; the AR6-based 50% 2°C budget (~1,020 Gt from 2026) runs out ~2051 on the central trend (DNV: 2052); non-CO₂ emissions still rising (+0.6%/yr implied) vs −45% CH₄ by 2050 in ≤2°C pathways.
- Technical/economic: durable removals ~0.0012 Gt/yr of operating capacity; CDR offtake value halving in 2026.
- Governance: SRM research-only (UK government does not support deployment; state bans such as Florida SB 56); no deployment path; EU methane-rule MRV delay proposed.
- Measurement: no newer long-term warming estimate or 2°C budget update in the ledger; IPCC AR7 timeline unresolved after IPCC-65 was suspended.
How this was assessed: the data, the benchmark and the rubric rule
- ETA method
- driver trend (shared driver fossil-co2-deceleration; own land-use, non-CO₂ and removal S-curve trajectories with consistent net flows) feeding a physical warming emulator: cumulative net CO₂ plus GWP* non-CO₂ warming equivalent against the sourced AR6 2°C budget and its AR6 WG3 non-CO₂ reference pathway; 10-year hold test as a definitional floor; named failure scenarios; one Monte Carlo distribution
- ETA calculation
- Shared driver fossil-co2-deceleration (drivers.yaml): draw i of `ledger.py driver-draws fossil-co2-deceleration --as-of 2026-10-18` (seed 20261015, n 10,000; byte-identical to the 2026-10-15 draws; a_per_yr2 lognormal p50 .00080015 = (.017320−.009319)/10 from co2-emissions-fossil-gcb 1995/2005/2015/2025 = 23.26/29.19/34.71/38.1 Gt, sdlog .50; g_2020 .0093186 and fossil_2025 38.1 fixed) is the fossil curvature in simulation draw i: fossil rate +0.452%/yr at 2026.5 centrally. Own mapping (prev's, unchanged, rule 7): net land use 2006–2025 OLS −2.203%/yr, level 4.443 (2025 obs 4.1, 2026-05-13-gcb-2025-total-co2-slightly-lower); non-CO₂ 16.0 GtCO₂e = assumed cross-inventory estimate EDGAR 54.1 (2026-09-01-edgar-2026-ghg-2025) − GCB 38.1, rate +0.614%/yr model-inferred, curvature m=1 × driver curvature (assumed mapping); per-draw common rate shift pins all-GHG 2026 to Climate TRACE Jan–Jul +0.3% ± 0.3 (2026-09-24-climate-trace-jul-2026-ghg; δ=0 centrally). 10,000 common draws; quantile year = calendar year containing it. Prev's script (notes/climate/20261015T113313Z.md) reproduced byte-identically on the 2026-10-18 series; no input changed (58 new effective events tested, none measures a modelled quantity). Removals: S-curve, seed .0012 Gt/yr (2026-09-18-greensand-eu-first-offshore-co2-storage, 2026-09-21-adm-columbus-beccs-removal-credits), r=ln(5.15/4.457)=.14452±.05 (2026-09-21-irena-tripling-tracker-693gw-2025; transfer assumed), K~LN(8,.54) Gt/yr (assumed; AR6 WG3 C.3.5 ≈3–16 Gt/yr); floors Rc U(1,4), RN U(3,4) (assumed). Failure scenarios: S1 removal stall .10 — reference class policy-revenue-only abatement services (first unit by 1996; stall = deployment ≤10× first unit after 30 y): FGD, SCR, catalytic converters, CFC substitution, landfill gas: 0 of 5 stalls (assumed classification), Jeffreys .083 / Laplace .143; climate-only CCS as a stall 1/6 → .21–.25 (the .25 row); new market signals both ways (2026-10-02-qcintel-cdr-offtake-value-halving-2026, 2026-10-06-kawasaki-abandons-dac-plans vs 2026-10-06-cdrfyi-sept-2026-over-1mt-contracted, 2026-10-15-frontier-4401-in-situ-mineralization-offtake), no member outcome: .10 kept (sens .05/.25: Balance P .296/.233, Ceiling .184/.178); S2 decarbonisation stall .05 (assumed, <10%). Seed evidence new both ways: Gevo ND BECCS >700 kt stored since capture began (≈.00016 Gt/yr, start date assumed) and CDR.fyi ~3% of ~50 Mt contracted delivered (2026-10-08-gevo-north-dakota-beccs-10000-delivered): +Gevo row Balance .287 / Ceiling .184, ×.1 row .182 / .170; kept (log-SD 1). Emulator (prev's, unchanged): X(t)=∫2026→t(net CO₂+WE_m) − cumWE_p, WE_m = f_s(4.53·N_eff(t)−4.25·N_eff(t−20))+(1−f_s)N_eff, f_s=.65 (AR6 WG3 C.2.3), reference non-CO₂ linear to N(8,1.82) at 2070; B2 = AR6 WG1 SPM.2 2°C budgets 2300/1700/1350/1150/900 from 2020 − measured 2020–25 247.66 − GCB revision 82.34 (2026-05-13-gcb-2025-remaining-budget-170gt; transfer assumed) → 50% 1,020 Gt = 24 y at 42.2 Gt/yr, + σ440; sampled 194/1106/2438. Test: t* = first t with max X over (t,t+10] ≤ X(t) and max X over [2026,t] ≤ B2; achieved t*+10 ≤ 2100. Gap: ~1.3–1.4°C long-term (2024-01-01-wmo-long-term-warming-1-3-1-4c, legacy); context, El Niño-boosted: C3S Sep 2026 1.63°C, 12-mo 1.50 (2026-10-09-c3s-sep-2026-1-63c-12mo-1-50c), GISS 2026 1.548 (projection, 2026-10-14-giss-gmsta-2026-1-55c-2027-1-64c); observed warming 0.36°C/decade since 2010 (2026-10-14-state-of-climate-warming-rate-0-36c) is not the indicator; it bears on the assumed aerosol term (+.05°C → P .162). MC: P .1831 → .18, p10 2082.1 → 2082, p50/p90 null (identical to prev); conditional-success median 2080.9 (not p50); t* 2058.6/2070.9/2084.7; stops 10.5–28.4 y before CO₂ net zero; 56% of successes rise again before 2100. Sensitivities (P): m=.5/0 .103/.0006; curvature ×½/×2 .040/.374 (p10 2064.1); total-CO₂ .334; floor −3% .097; f_s .50/.80 .149/.198; GWP* ×.7/×1.3 .157/.190; reference non-CO₂ 5/11 .151/.219; r ×½/×1.5 .160/.207; TRACE +0.5%/−0.3% .153/.276; no GCB revision .201; no σ440 .163; aerosol ±.05°C .162/.206; allow overshoot .209; S1 .05/.25 .184/.178. Outlook cross-check (projections, not inputs): model central fossil 2050/2025 = 0.896 (38.1 → 34.14 Gt), cumulative fossil 2026–2050 901 Gt; curvature each outlook implies in the driver form, ×driver p50 (driver CDF): Exxon 0.83 ×1.23 (.65), IRENA WETO PES 0.74 ×1.57 (.81; 2026-10-13-irena-weto-2026-net-zero-only-in-1-5c), DNV 22/39 ×2.37 (.96; 2026-10-07-dnv-eto-2026-energy-co2-likely-peaked), Rystad ~0.5 ×2.72 (.98; 2026-10-15-rystad-ges-2026-fossil-co2-plateau-40gt); they decline faster to 2050 but flatten after (DNV net-zero CO₂ 2098), i.e. the a×2 and −3%/yr-floor rows at once; driver-fit remark for the owner in the notes, driver not refitted. Cross-checks: model central cumulative gross CO₂ from 2026 reaches the 1,020 Gt budget in 2051.0 vs DNV 2°C-budget exhaustion 2052 and 2.3°C by 2100 (2026-10-07-dnv-eto-2026-2c-budget-exhausted-2052); IRENA PES warming ≥2°C; all outlook central cases fail the Ceiling, consistent with P .18; AR6 C3 needs CH₄ −45% by 2050 vs model central non-CO₂ −7%; IPCC AR7 timeline unresolved (2026-10-16-ipcc-65-suspended-ar7-timeline); SRM: UK does not support deployment (2026-09-29-uk-govt-srm-policy-paper-no-deployment), research/governance only, no success branch. Prev P .18, p10 2082: unchanged. No external anchors listed. Foreign context only, no mapping.
- Change
- Re-assessment on 58 new effective events (published 2026-09-28…10-18). Prev's script reproduced byte-identically on byte-identical series and driver draws. Inputs tested against the new evidence: the new temperatures (C3S, NCEI, GISS) are El Niño-boosted surface readings, the 0.36°C/decade study is an observed rate without an indicator estimate (it bears on the assumed aerosol term, kept at 0 with ±.05°C rows), the outlooks are projections and agree with the model (budget exhausted 2051.0 vs DNV 2052), SRM items are research/governance. No input changed (rule 7). P .1831, p10 2082.1: no move; status red unchanged.
- Rubric rule
- RUBRIC.md v6 §4/§4.1: red, p_achieved < 0.5 so the median is null; the indicator is still rising and no verified event exists.
by assess:20261018T192714Z · 59 evidence event(s)Evidence and news: 59 cited by the assessment, 2 other event(s)
Cited evidence (59)
- IPCC Chair Jim Skea suspended the IPCC's 65th plenary session in Addis Ababa on 16 October 2026 without agreement on the timelines for the three Working Group contributions to the Seventh Assessment Report (AR7), including Working Group I on the physical science basis, or on the IPCC budget, after the meeting overran and many least-developed-country and small-island delegations had left, leaving both items 'open and under discussion' until the session resumes, which the IPCC said would be 'at the earliest opportunity'.
- Rystad Energy's Global Energy Scenarios 2026 (15 October 2026) said global fossil CO2 emissions are plateauing at around 40 Gt per year and projected that in its new 'Pragmatism' reference scenario they fall to about 20 Gt by 2050 (800 Gt cumulative over 2026-2050, about 0.35°C of additional warming), versus about 2 Gt in 'Climate first' and near today's level for longer in 'Fossil first'.
- Frontier buyers (including Stripe, Google, Shopify, Autodesk, H&M Group and Workday) signed offtake agreements, announced on 15 October 2026, to pay 44.01 $11.7 million to permanently remove 20,932 tonnes of CO2 between 2028 and 2031 by injecting dissolved biogenic CO2 from wastewater-treatment biogas upgrading into peridotite rock, where it mineralises within 1-3 years; it is Frontier's first purchase of in-situ mineralization removal and its first in the Middle East, and 44.01 calls it the first commercial-scale project of its kind, moving from short-duration injection pilots towards continuous operation over multiple years.Permanent Removal [1] Frontier: 44.01 and Frontier advance permanent CO₂ storage through in-situ mineralization (15 October 2026) [2] 44.01: Frontier backs 44.01 to take in-situ CO₂ mineralisation to commercial scale (15 October 2026) [3] ESG Dive: Frontier Climate buyers ink deal to store CO2 as solid rock (16 October 2026)
- The 2026 State of the Climate report (Ripple, Wolf, Mann, Rockström et al., BioScience, published 14 October 2026) found that the rate of global warming appears to be accelerating, nearly doubling from about 0.19°C per decade over 1970–2010 to about 0.36°C per decade since 2010, with 7 of the 16 climate vital signs it tracks at record extremes and the strong El Niño putting 2026 on track to be one of the hottest years on record.The Ceiling [1] Oregon State University: Report: Earth's climate system increasingly unstable, but signs of positive change emerge (14 October 2026) [2] Ripple et al. (2026), The 2026 state of the climate report: challenges and opportunities, BioScience, doi:10.1093/biosci/biag125 [3] CBS News: Climate change pushed Earth to its limits in 2026 as certain 'vital signs' reach record extremes, report says (14 October 2026)
- NASA GISS's GISTEMP GMSTA prediction update of 14 October 2026 (with Columbia's IRI/CCSR) estimated 2026 annual global mean surface temperature at 1.548°C above 1850–1900 from the NMME model mean (up from 1.526°C in the 11 September update), or 1.426 ± 0.061°C (95% CI) from its year-to-date regression, and projected 2027 at 1.643 ± 0.227°C (95% CI) from its projected DJF ENSO model.The Ceiling [1] NASA GISS: GISTEMP GMSTA Analysis Project - Global Mean Surface Temperature Anomaly Predictions [2] NASA GISS: Annual GMSTA predictions, October 2026 update (CSV, last modified 14 October 2026) [3] NASA GISS: Annual GMSTA predictions, September 2026 update (CSV, last modified 11 September 2026)
- The European Commission formally proposed on 14 October 2026 a one-year postponement of the EU Methane Regulation's monitoring, reporting and verification obligations for importers of oil, gas and coal (and of the methane-intensity reporting provisions for importers and EU producers), citing security-of-supply pressure from the Middle East conflict while saying the change does not affect the regulation's ambition; the proposal, which needs European Parliament and Council approval, would move the importer requirements from 2027 to 2028.Beyond the frameworkNet-Zero Transition [1] European Commission: Commission proposes flexibility to methane importers to help save costs (IP/26/2119) [2] Rapporteur: Brussels moves to postpone methane reporting rules (14 October 2026) [3] Carbon Brief Daily Briefing (15 October 2026): postpone methane rules
- IRENA's World Energy Transition Outlook 2026 (launched 13 October 2026) projected that under existing government plans (its Planned Energy Scenario) global fossil CO2 emissions fall only 26% by 2050 (0.2% a year over 2025-2030 and 1.4% a year thereafter), with cumulative 2026-2050 energy-related CO2 of about 840 GtCO2 and warming likely at least 2°C, that its Decarbonising Energy Scenario (no longer aligned with the net-zero-by-2050 goal of earlier editions) still leaves about 4 GtCO2 of net fossil CO2 in 2050 (about 480 GtCO2 cumulative), and that only its revised 1.5°C Scenario, with a limited and temporary overshoot and more carbon capture including BECCS, reaches net-zero fossil CO2 in 2050 (cumulative slightly below 400 GtCO2).
- A peer-reviewed study in Atmospheric Measurement Techniques (Sellitto et al., published 12 October 2026), using simulated observations from the proposed CAIRT satellite concept, found that high-spectral-resolution mid-infrared limb-emission sounding could detect stratospheric aerosol injection SO2 releases within hours to a few days and quantify injections as small as a few hundred tonnes of SO2 per day (the scale of small outdoor experiments or near-term unilateral deployments), a monitoring capability that existing satellite instruments lack.
- NOAA NCEI's September 2026 global climate report (9 October 2026) put the September 2026 global surface temperature at 1.33°C above the 1901–2000 average, the second-warmest September in the 1850–2026 record (0.07°C behind 2023), with the global ocean surface departure of +1.10°C the highest for any month on record, and January–September 2026 at +1.17°C, second only to 2024 (+1.24°C), giving a greater than 99% probability that 2026 ranks among the five warmest years, with second-warmest the most likely outcome.
- The Copernicus Climate Change Service (C3S) reported on 9 October 2026 that September 2026 was the second-warmest September in ERA5 at 1.63°C above the 1850–1900 pre-industrial average (0.11°C cooler than September 2023), the second consecutive month above 1.5°C in a run C3S expects to continue into 2027, that the 12-month average for October 2025–September 2026 returned to 1.50°C above pre-industrial for the first time since October 2025, and that January–September 2026 ranks second-warmest behind 2024.
- NOAA's Climate Prediction Center ENSO diagnostic discussion of 8 October 2026 kept an El Niño Advisory in force, reporting that El Niño strengthened further over the past month with Niño indices reaching +2.1°C in Niño-3.4, +3.0°C in Niño-3 and +3.9°C in Niño-1+2, and gave a greater than 83% chance of a strong-to-very-strong El Niño through January-March 2027, with chances of a historic event exceeding all El Niños since 1950 (3-month relative ONI of +2.5°C or more) of 54% for September-November, 83% for October-December and 70% for November-January.
- Gevo announced on 8 October 2026 the sale and delivery of 10,000 carbon dioxide removal credits through ClimeFi to a corporate buyer from its Gevo North Dakota ethanol BECCS facility, where fermentation CO2 is stored geologically more than a mile beneath the plant and which Gevo said has generated more than 700,000 tonnes of CO2 removal since capture began, while citing CDR.fyi data that only about 3% of roughly 50 million tonnes of contracted CDR has been delivered market-wide.
- DNV's Energy Transition Outlook 2026 (launched 7 October 2026) judged that global energy-related CO2 emissions 'likely peaked in the last 24 months' and that total anthropogenic CO2 emissions (energy, industrial processes and land-use change, net of direct air capture) peaked in 2024 at 43 Gt, forecasting energy-related CO2 falling 44% from 39 Gt today to 22 Gt in 2050 (and 65% to 12 Gt in 2060), with net-zero CO2 only in 2098 and end-of-century warming of about 2.3°C.
- DNV's Energy Transition Outlook 2026 (launched 7 October 2026) forecast end-of-century warming of 2.3°C 'based on current IPCC climate sensitivity', with the 1.5°C carbon budget exhausted in 2029, the 2°C budget exhausted in 2052 and cumulative CO2 emissions overshooting the 2°C budget by 420 Gt (260 Gt of it emitted after 2060) before net-zero CO2 in 2098, judging 1.5°C 'effectively unattainable' but well below 2°C still attainable with stronger commitments.
- WMO's Global Seasonal Climate Update (6 October 2026) forecast the October-December 2026 seasonal-average Niño 3.4 sea-surface temperature anomaly at about +3.7°C, peaking around December, the highest since observations began versus the previous record of +2.6°C in November 2015-January 2016, after the index rose from +1.6°C in June to +2.5°C in August, with a near-100% likelihood (consensus of 13 centres) that El Niño persists through February 2027.Beyond the framework [1] Sea-surface temperatures reflect strengthening El Niño (WMO, 6 Oct 2026)
- Japanese industrial group Kawasaki Heavy Industries abandoned its direct air capture expansion plans and shifted its focus to CO2 capture, utilisation and storage, Quantum Commodity Intelligence reported on 6 October 2026.
- CDR.fyi's September 2026 market recap (6 October 2026) reported that more than 1 million tonnes of durable carbon removal were contracted in September 2026, led by Google's 1 Mt enhanced-weathering purchase from Terradot, up from about 128,000 tonnes contracted in August 2026.
- Quantum Commodity Intelligence's monthly carbon-removal offtake tracking (2 October 2026) found the value of CDR offtake agreements on track to more than halve in 2026 compared with 2025, although September 2026 offered a partial reprieve with 21 offtake deals worth an estimated $148 million.
- A report published on 1 October 2026 by the ARIA-funded SAFEGEOGOV project (University of Leeds with Queen Mary University of London, Imperial College London and the Centre for Future Generations) found that in three multi-day SRM governance crisis simulations, in which 35 international academics, civil-society actors and practitioners role-played governments, institutions and companies, participants never reached an SRM governance agreement able to pass a two-thirds majority (simulations 1 and 2) or consensus (simulation 3), with 'strategic competition and state sovereignty' rather than climate concerns driving the negotiations.Sunlight Management [1] Centre for Future Generations: Negotiating solar radiation modification governance — Evidence from non-ideal simulations (1 October 2026) [2] Centre for Future Generations: Governing climate cooling technologies — tools for policymakers (5 October 2026) [3] WIRED: They War-Gamed Dimming the Sun. It Went Even Worse Than You Think (12 October 2026)
- The European Commission published on 30 September 2026, and opened for submissions on 6 October 2026 (deadline 16 February 2027), a Horizon Europe call topic (HORIZON-BRIDGING-2027-04-11) offering about €10 million for one research and innovation project to build capabilities to detect, monitor, model and attribute solar radiation modification activities, including outdoor experiments and potential unilateral deployments, and to assess their impacts on climate and ecosystems, with SRM field experiments and technical SRM development explicitly out of scope; Bloomberg reported it as part of a €35 million Horizon tender.Sunlight Management [1] EU Funding & Tenders Portal: HORIZON-BRIDGING-2027-04-11 — Detection, monitoring, modelling, and attribution capabilities for Solar Radiation Modification (SRM) activities (published 30 September 2026) [2] Bloomberg: EU Wants to Track Potential Rogue Efforts to Block Out Sun (6 October 2026)
- The EDGAR 2026 report (September 2026) put total greenhouse-gas emissions excluding LULUCF at 54.1 GtCO2e in 2025, up 0.7% on 2024, with China up 0.1% and the US up 2.2%.↳ follows up on 2024: The EDGAR 2025 report put 2024 global greenhouse-gas emissions at 53.2 GtCO2e,… (corrected version)
- On 29 September 2026 the UK Department for Energy Security and Net Zero published a policy paper stating that the UK Government 'does not support the deployment of SRM and does not plan to do so' but supports responsible research into its risks and impacts, and summarising UK publicly funded SRM research, including ARIA's £56.8 million Exploring Climate Cooling programme (whose 5 announced small-scale field trials must have independent environmental impact and legal assessments published before any outdoor experiment) and a £10.5 million NERC modelling-only programme for 2025-2030.Sunlight Management [1] GOV.UK (DESNZ): UK government's view on solar radiation modification technologies (SRM) (policy paper, 29 September 2026) [2] WIRED: They War-Gamed Dimming the Sun. It Went Even Worse Than You Think (12 October 2026; calls it the government's first official policy paper on SRM)
- A European Commission spokesperson said on 28 September 2026 that the Commission plans to propose postponing the entry into force of the EU Methane Regulation (length not specified), after pressure from governments including France and the United States and after Energy Commissioner Dan Jørgensen said on 25 September that Brussels was exploring a one-year delay of the rules for imported oil, gas and coal, whose monitoring and reporting obligations for importers are due to take effect on 1 January 2027.Beyond the frameworkNet-Zero Transition [1] Reuters: EU to delay methane emissions rules after pushback from governments (28 Sep 2026) [2] Argus: EU plans one-year delay to methane law for oil, gas (28 Sep 2026) [3] eKathimerini (Reuters syndication, Kate Abnett): EU to delay methane emissions rules after pushback from governments (28 Sep 2026)
- CarbonPlan released (28 September 2026) an open-source, globally consistent dataset of statistically downscaled (0.25°, daily) stratospheric aerosol injection climate simulations, covering 116 model simulations from two climate models (CESM2(WACCM6) and UKESM1.1) processed with two downscaling methods under the GeoMIP G6-1.5K-SAI deployment scenario, a termination-shock scenario, an SSP2-4.5 no-SAI baseline and the historical period, to let researchers study regional SAI impacts such as on crops, water and health.
- Climate TRACE (24 September 2026) estimated global greenhouse-gas emissions of 4.9 GtCO2e in July 2026, up 0.1% on July 2025, and 33.7 GtCO2e year-to-date for January-July 2026, 0.3% above the same period of 2025, with July emissions down 0.3% in China and 2.6% in the US but up 3.4% in India.
- The Government of Canada said on 24 September 2026 that it will develop a policy framework under Paris Agreement Article 6 to let Canadian companies, including carbon removal suppliers, trade internationally transferred mitigation outcomes (ITMOs), which it said could contribute billions of dollars to the Canadian economy and catalyse domestic climate mitigation projects and technologies.Permanent Removal [1] Canada to develop carbon removal market framework under Paris agreement (Reuters) [2] Canada Takes First Steps to Enter International Carbon Market (ESG Today)↳ follows up on 2024-11: COP29 in November 2024 finalised the Paris Agreement Article 6 rules for international… (corrected version)
- Berkeley Earth's August 2026 update (24 September 2026) put the August 2026 global mean surface temperature at 1.71 ± 0.11°C above the 1850-1900 average, the warmest August on record and the 2nd-warmest month of any kind (behind September 2023), with ocean surfaces at 1.296°C the warmest month on record for the oceans, while the September 2025-August 2026 12-month average was 1.48°C and January-August 2026 averaged 1.49°C.
- Berkeley Earth projected (24 September 2026) a roughly 93% chance that 2026 will be the warmest year on record (up from about 69% in its July update), with the 2026 annual global mean 'all but certain' to exceed 1.5°C above 1850-1900 (likely range 1.536-1.642°C), and a further substantial El Niño warming influence expected in 2027.
- At the UN Climate Summit 2026 in New York on 23 September 2026, UN Secretary-General António Guterres, saying greenhouse gas emissions are still rising and the world is expected to temporarily overshoot 1.5°C, called on every country to develop a credible roadmap with clear timelines to transition away from fossil fuels, with the G20 (about 80% of global emissions) to lead, and the summit launched a Global Grids Accelerator (citing more than 2,500 GW of renewable projects waiting for grid connections worldwide) and a critical-minerals initiative.
- Germany's federal cabinet approved, and Environment Minister Carsten Schneider presented at the UN in New York on 23 September 2026, a national roadmap for transitioning away from coal, oil and gas by 2045 (Germany's climate-neutrality target year), making Germany the third country after France and the Netherlands to publish such a roadmap; it reaffirms existing goals such as a coal exit no later than 2038 and at least 80% renewable electricity by 2030 (versus about 55% of gross electricity consumption now), with fossil fuels still 65% of German energy consumption in 2024.
- The Degrees Initiative launched (23 September 2026) a $1.5 million coordinated global study, billed as the first of its kind, in which research teams across nine Global South regions plus China and the Global North will use shared data, protocols and climate scenarios with and without solar radiation modification (SRM) to model how warming and SRM could affect staple crops such as maize, rice, wheat, cassava and sorghum.
- In his 22 September 2026 address to the UN General Assembly, US President Donald Trump attacked the IMO Net-Zero Framework (the proposed global marine-fuel greenhouse-gas-intensity standard and emissions-pricing mechanism meant to deliver the IMO goal of net-zero international shipping emissions by or around 2050) as a 'global carbon tax' and vowed 'there will be no global taxes' while he is president, signalling continued US opposition ahead of the framework's expected adoption vote at an extraordinary MEPC session on 4 December 2026, after US pressure had already postponed adoption by a year in October 2025.The BalanceNet-Zero Transition [1] Remarks to the United Nations General Assembly in New York City, September 22, 2026 (American Presidency Project transcript) [2] Trump Takes Aim at Shipping's Net-Zero Plan at UN (gCaptain, 23 Sep 2026) [3] Competing Climate Visions Clash at UN General Assembly (Inside Climate News, 22 Sep 2026)
- IRENA's third UAE Consensus tracking report, released with the COP31 Presidency and the Global Renewables Alliance on 21 September 2026, found that a record 693 GW of renewable power capacity was added worldwide in 2025 (513 GW solar PV, 158 GW wind), bringing total installed renewable capacity to about 5.15 TW, but that tripling to about 11.2 TW by 2030 now requires average net additions of about 1.2 TW per year in 2026-2030, and that sustaining the 2025 growth rate would leave a shortfall of about 0.6 TW in 2030 (down from 0.9 TW in the previous edition).Net-Zero Transition [1] IRENA et al. (2026), Delivering on the UAE Consensus: Tracking progress toward tripling renewable energy capacity and doubling energy efficiency by 2030 [2] Global renewable deployment must double to hit 2030 climate target, report says (Reuters, 21 Sep 2026) [3] Report launched at the Global Renewables Summit highlights renewable energy progress but warns of challenges ahead to 2030 (ALER)
- A Carbon Brief analysis (21 September 2026) found that the daily Niño 3.4 sea-surface temperature anomaly, calculated with NOAA's ONI convention (centred 30-year climatology), reached 3.11°C on 21 September 2026, surpassing the previous daily record of 3.08°C set on 18 November 2015, with the August 2026 monthly anomaly (about 2.45°C) already above the peak of every prior El Niño on record except 2015-16 (2.75°C) and 1877-78 (about 2.7°C, reconstructed).Beyond the framework [1] Analysis: 'Super El Niño' breaks 'remarkable' all-time record (Carbon Brief)
- ADM announced on 21 September 2026 its planned entry into the voluntary carbon removal market with credits from the 800,000+ tCO2-per-year biogenic (ethanol fermentation) carbon capture operation at its Columbus, Nebraska corn processing complex, whose CO2 is stored permanently in Tallgrass's Eastern Wyoming Sequestration Hub; the credits are undergoing Puro.earth certification, with initial issuance expected by the end of 2026 and a 15-year crediting period.
- Greensand, led by INEOS Energy with Harbour Energy and Nordsøfonden, entered commercial operation on 18 September 2026 as the EU's first full-scale offshore CO2 storage site, injecting CO2 sourced primarily from Danish biomethane plants into the depleted Nini West field about 1,800 m beneath the Danish North Sea, with capacity of up to 400,000 tCO2 per year in its first phase and potential expansion to 4-8 MtCO2 per year.Permanent RemovalNet-Zero Transition [1] INEOS opens the first full-scale CO2 storage site in the EU (INEOS)
- ExxonMobil's 2026 Global Outlook (17 September 2026) projected global energy-related CO2 emissions to fall about 20% from 36 billion tonnes in 2025 to 30 billion tonnes in 2050, up from the 27 billion tonnes it projected for 2050 a year earlier and far above the roughly 11 billion tonnes in IPCC 'likely below 2°C' scenarios, with coal still 15% of the 2050 energy mix and carbon capture and storage reaching 2 billion tonnes per year (down from its prior 3.1 billion).
- An Earth-system-model study posted for discussion on EGUsphere on 16 September 2026 (Tjiputra et al., not yet peer reviewed) found that tropical, Northern-Hemisphere and Southern-Hemisphere mid-latitude stratospheric aerosol injection all kept simulated overshoot warming below 2°C, but that deployment location strongly shapes side effects via the AMOC, and that legacy effects on sea-level rise and permafrost persist for centuries after SAI ends.
- A Carbon Brief analysis (16 September 2026) projected that global fossil CO2 emissions are set to fall by about 0.5% in 2026, driven by the Strait of Hormuz crisis.
- Argus data (16 September 2026) showed technology-based CO2 removal credits contracted through long-term offtake agreements fell to about 6 million in January-August 2026 from 24 million a year earlier, with BECCS credits down to 2 million from 17 million and Microsoft's disclosed buying down to just over 2 million from 16.5 million, while biochar offtakes rose to nearly 3 million from 1.5 million.
- The Climate Pledge 2026 report (Amazon and Global Optimism, 15 September 2026) said more than 700 companies across 62 industries and 49 countries and territories, with combined revenue of $3.8 trillion, have committed to reach net-zero carbon emissions by 2040, after 107 new signatories joined in 2025 (19% signatory growth), and that analysed signatories cut their self-reported Scope 1 and 2 emissions by 11% on average (about 14 MtCO2e over 2021-2024) versus 7% for non-signatory companies.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put total CO2 emissions (fossil + land-use change) at 42.4 GtCO2 in 2024 and a slightly lower preliminary 42.2 GtCO2 in 2025, due to lower net land-use emissions (to about 4.1 GtCO2) mainly attributable to the end of El Niño conditions.↳ confirms 2025-11-13: The Global Carbon Budget 2025 projected CO2 emissions from land-use change at about 4.1… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) found that total CO2 emissions (fossil + land-use change) grew 0.3% per year over 2015-2024, compared with 1.9% per year over 2005-2014.
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put the remaining carbon budget for a 50% chance of limiting warming to 1.5°C at 170 GtCO2 (50 GtC) from the start of 2026, about 4 years at 2025 emission levels.↳ updates 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, concluded that the… (corrected version)
- The Global Carbon Budget 2025 final paper (ESSD, 13 May 2026) put global fossil CO2 emissions (fossil fuels + cement, net of cement carbonation) at 38.1 GtCO2 in 2025, up 1.0% (range 0.2% to 1.7%) on 2024 and 'an historical record high', with coal, oil and gas emissions each above 2024 levels (+1.0%, +1.1% and +1.3%).↳ updates 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, projected global fossil… (corrected version)
- The UK Met Office said even its forecast 2026 rise in Mauna Loa CO2 is 'too fast to track IPCC 1.5°C scenarios', which require a 2020s decadal average growth of 1.33-1.79 ppm/yr versus 2.61 ppm/yr observed for 2020-2025 (Scripps annual-mean basis; 1.91 ppm/yr in the 2000s and 2.41 in the 2010s).
- CarbonCapture Inc. paused development of Project Bison in Wyoming (original target 5 MtCO2 per year of direct air capture), relocating due to competition from data centers.Permanent Removal [1] Carbon Herald: CarbonCapture pauses Project Bison↳ corrects (or adds sources to) an earlier record:
2025: CarbonCapture Inc. paused development of Project Bison in Wyoming (original target 5… - A 2025 Nature Communications study revised global reforestation potential down to 6-195 Mha (53-2,225 TgCO2e per year of sequestration) after sustainability safeguards, versus previous estimates of more than 1,000 Mha.Net-Zero Transition [1] Nature Communications: reforestation potential with safeguards↳ corrects (or adds sources to) an earlier record:
2025: A 2025 Nature Communications study revised global reforestation potential down to 6-195… - Climeworks’ Mammoth plant in Iceland has a nameplate capacity of 36,000 tons of removal per year, but its collector containers were not operating as efficiently as prototype testing indicated.Permanent Removal [1] Climeworks: Mammoth [2] Latitude Media: On the ground at Climeworks’ biggest DAC project↳ corrects (or adds sources to) an earlier record:
2025: Climeworks' Mammoth direct air capture plant in Iceland (36,000 tCO2 per year nameplate… - CDR.fyi reported $2.3 billion of private direct air capture investment from 2021 to H1 2025, but only about 1,186 tonnes of DAC removal delivered out of 2.47 million tonnes contracted.Permanent Removal [1] CDR.fyi DAC Market Snapshot 2025↳ corrects (or adds sources to) an earlier record:
2025: CDR.fyi reported $2.3 billion of private direct air capture investment from 2021 to H1… - CDR.fyi's 2025 DAC market snapshot put operating direct air capture capacity at about 0.01 MtCO2 per year and total CCS capacity (all types) at 41 MtCO2 per year, versus roughly 10,000 MtCO2 per year needed.Permanent Removal [1] CDR.fyi DAC Market Snapshot 2025
- A Nature study published alongside the Global Carbon Budget 2025 found land and ocean carbon sinks are 15% weaker than they would have been without climate impacts, contributing about 8% of the rise in atmospheric CO2 since 1960.↳ follows up on 2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, projected global fossil… (corrected version)↳ corrects (or adds sources to) an earlier record:
2025-11-13: A Nature study published alongside the Global Carbon Budget 2025 found land and ocean… - The Global Carbon Budget 2025, released on 13 November 2025, concluded that the remaining carbon budget for limiting warming to 1.5°C is 'virtually exhausted'.The Ceiling [1] Global Carbon Budget 2025↳ corrects (or adds sources to) an earlier record:
2025-11-13: The Global Carbon Budget 2025, released on 13 November 2025, concluded that the… - Solar geoengineering startup Stardust raised $60 million on 24 October 2025, the largest SRM fundraise to date, and planned stratospheric tests for April 2026.Sunlight Management [1] POLITICO: Global cooling startup raises $60M to test sun-reflecting technology [2] Wayback capture 20251228221222↳ corrects (or adds sources to) an earlier record:
2025-10-24: Solar geoengineering startup Stardust raised $60 million on 24 October 2025, the… - Florida’s SB 56, effective 1 July 2025, prohibits certain acts intended to affect temperature, weather, or the intensity of sunlight within the state’s atmosphere.Sunlight Management [1] Florida Senate: Senate Bill 56 (2025) - The Florida Senate↳ corrects (or adds sources to) an earlier record:
2025-07-01: Florida's SB 56, signed into law and effective 1 July 2025, made geoengineering a… - The WMO State of the Global Climate 2024 put long-term global warming (the Paris Agreement metric) at about 1.3-1.4°C above pre-industrial levels.The Ceiling [1] WMO State of the Global Climate 2024
- WMO confirmed 2024 as the warmest year on record at 1.55°C +/- 0.13°C above the 1850-1900 average, based on a six-dataset average.The Ceiling [1] WMO confirms 2024 warmest year on record↳ confirms 2024: Copernicus reported 2024 global mean surface temperature at 1.60°C above 1850–1900, the… (corrected version)↳ corrects (or adds sources to) an earlier record:
2024: WMO confirmed 2024 as the warmest year on record at 1.55°C +/- 0.13°C above the… - NASA GISS put the 2024 global temperature anomaly at 1.47°C above pre-industrial (1850-1900), confirming 2024 as the warmest year on record.The Ceiling [1] NASA confirms 2024 warmest year on record↳ corrects (or adds sources to) an earlier record:
2024: NASA GISS put the 2024 global temperature anomaly at 1.47°C above pre-industrial… - Copernicus reported 2024 global mean surface temperature at 1.60°C above 1850–1900, the first calendar year above 1.5°C.The Ceiling [1] Copernicus: 2024 first year to exceed 1.5°C↳ corrects (or adds sources to) an earlier record:
2024: Copernicus reported the 2024 global mean surface temperature at 1.60°C above the…
Other events (2)
- WMO gave 70% odds that the 2025-2029 five-year average global temperature will exceed 1.5°C above pre-industrial levels.
- NOAA put the 2024 global temperature anomaly at 1.46°C above the 1850-1900 pre-industrial level.The Ceiling [1] NOAA NCEI Global Climate Report 2024
Major Open Challenges
Net-Zero Transition: Balancing carbon flows across energy, industry, agriculture, and transport — through both emissions reduction and enhanced biological/geological uptake.50 event(s)
Events (50)
- France and Germany announced on 16 October 2026, on the sidelines of the EU summit in Brussels, a joint position on the EU automotive package under which the 2035 CO2 fleet target for new cars would be lowered from a 100% to a 90% reduction with no requirement for carmakers to offset the remaining 10% (going beyond the Commission's December 2025 proposal, which also set 90% but with offsetting), in exchange for stronger 'Made in Europe' requirements, with Chancellor Friedrich Merz declaring the combustion-engine phase-out 'history'.Net-Zero Transition [1] Bundesregierung: Joint Franco-German Press Statement on the Future of the European Automotive Industry [2] Deutschlandfunk: Merz: 'Verbrenner-Aus ist Geschichte' (16 October 2026) [3] electrive: France and Germany propose 90 per cent CO2 cut for new cars by 2035 (16 October 2026)
- Benchmark Mineral Intelligence data (16 October 2026) showed global electric vehicle sales of 2.15 million in September 2026, down 2% year on year (up 17% month on month), bringing January–September 2026 sales to 15.5 million, up 3% on the same period of 2025, as record European sales (almost 600,000, +39% y/y) were outweighed by year-on-year falls of 44% in North America (against the September 2025 rush before the US federal EV tax credit expired) and 18% in China, whose passenger-car NEV exports hit a record 558,000.
- At the 15–16 October 2026 European Council, EU leaders did not agree to the postponement of the ETS2 carbon market for road-transport and building fuels (due to start in 2028 after an earlier one-year delay) until 2030 sought by the Visegrad countries and Italy, instead calling on the Commission to update its assessment of all impacts of ETS2, in particular its design and safeguards, and urging co-legislators to speed up reform of the EU ETS market stability reserve to curb carbon-price volatility.Net-Zero Transition [1] Il Sole 24 Ore: EU summit: leaders say 'The Commission should assess the impact of the ETS 2 system' (16 October 2026) [2] EU Today: EU resists further delay to carbon pricing as energy costs divide European leaders (17 October 2026) [3] Politico Europe: EU leaders' summit descends into 3-hour argument over taxing car emissions (15 October 2026)
- Ember (14 October 2026) found that the 2030 energy-storage targets of the 33 jurisdictions with one among the 98 countries and the EU in its Global Renewable Target Tracker now sum to 822 GW, 2.9 times the total assessed at COP29 in 2024 (almost entirely due to China's new 460 GW target and the EU's first 200 GW bloc-wide target), still short of the COP29 Global Energy Storage and Grids Pledge of 1,500 GW by 2030, which would require 214 GW of storage additions a year in 2026–2030 versus a record 118 GW (108 GW of it batteries) added in 2025, bringing global installed storage to 430 GW.
- The European Commission formally proposed on 14 October 2026 a one-year postponement of the EU Methane Regulation's monitoring, reporting and verification obligations for importers of oil, gas and coal (and of the methane-intensity reporting provisions for importers and EU producers), citing security-of-supply pressure from the Middle East conflict while saying the change does not affect the regulation's ambition; the proposal, which needs European Parliament and Council approval, would move the importer requirements from 2027 to 2028.Beyond the frameworkNet-Zero Transition [1] European Commission: Commission proposes flexibility to methane importers to help save costs (IP/26/2119) [2] Rapporteur: Brussels moves to postpone methane reporting rules (14 October 2026) [3] Carbon Brief Daily Briefing (15 October 2026): postpone methane rules
- +45 older event(s): the section's event log
Permanent Removal: Capturing and storing billions of tons of CO₂ per year durably at viable cost.31 event(s)
Events (31)
- Frontier buyers (including Stripe, Google, Shopify, Autodesk, H&M Group and Workday) signed offtake agreements, announced on 15 October 2026, to pay 44.01 $11.7 million to permanently remove 20,932 tonnes of CO2 between 2028 and 2031 by injecting dissolved biogenic CO2 from wastewater-treatment biogas upgrading into peridotite rock, where it mineralises within 1-3 years; it is Frontier's first purchase of in-situ mineralization removal and its first in the Middle East, and 44.01 calls it the first commercial-scale project of its kind, moving from short-duration injection pilots towards continuous operation over multiple years.Permanent Removal [1] Frontier: 44.01 and Frontier advance permanent CO₂ storage through in-situ mineralization (15 October 2026) [2] 44.01: Frontier backs 44.01 to take in-situ CO₂ mineralisation to commercial scale (15 October 2026) [3] ESG Dive: Frontier Climate buyers ink deal to store CO2 as solid rock (16 October 2026)
- Parties to the London Convention and London Protocol meeting at the IMO (LC 48/LP 21, 5-9 October 2026) adopted a resolution establishing assessment frameworks for ocean fertilization, ocean alkalinity enhancement and ocean sinking of biomass, under which proposals are assessed case by case to show legitimate scientific research purpose and acceptable environmental risk, with the Parties stating that further research is needed before any activities beyond the research stage can be considered.
- RMI's New Frontiers Report for Carbon Dioxide Removal (8 October 2026), covering 14 CDR approaches including nature-based living-biomass removal, estimated end-of-2026 deployed CDR capacity at about 42 MtCO2 per year with $13.6 billion of funding raised across 514 tracked companies and 694 active or announced projects, and found that the 217 MtCO2 per year of 2030 supply implied by announced projects and company targets would require roughly 51% annual capacity growth.
- Gevo announced on 8 October 2026 the sale and delivery of 10,000 carbon dioxide removal credits through ClimeFi to a corporate buyer from its Gevo North Dakota ethanol BECCS facility, where fermentation CO2 is stored geologically more than a mile beneath the plant and which Gevo said has generated more than 700,000 tonnes of CO2 removal since capture began, while citing CDR.fyi data that only about 3% of roughly 50 million tonnes of contracted CDR has been delivered market-wide.
- ClimeFi's CDR Market Review for Q3 2026 (7 October 2026) found that enhanced rock weathering made up a majority of a quarter's durable-CDR purchase volume for the first time, driven by the largest ERW deal on record, while biochar purchases (above 1 MtCO2 in each of the two previous quarters) fell to their lowest since Q1 2023, Microsoft's buying slowed to about 2% of Q3 volume with buyers other than Microsoft accounting for 64% of 2026 volume to date, and debt made up 94% of all disclosed Q3 private funding.
- +26 older event(s): the section's event log
Sunlight Management: Developing and safely governing solar radiation management complementing the other two challenges.17 event(s)
Events (17)
- A peer-reviewed study in Atmospheric Measurement Techniques (Sellitto et al., published 12 October 2026), using simulated observations from the proposed CAIRT satellite concept, found that high-spectral-resolution mid-infrared limb-emission sounding could detect stratospheric aerosol injection SO2 releases within hours to a few days and quantify injections as small as a few hundred tonnes of SO2 per day (the scale of small outdoor experiments or near-term unilateral deployments), a monitoring capability that existing satellite instruments lack.
- A report published on 1 October 2026 by the ARIA-funded SAFEGEOGOV project (University of Leeds with Queen Mary University of London, Imperial College London and the Centre for Future Generations) found that in three multi-day SRM governance crisis simulations, in which 35 international academics, civil-society actors and practitioners role-played governments, institutions and companies, participants never reached an SRM governance agreement able to pass a two-thirds majority (simulations 1 and 2) or consensus (simulation 3), with 'strategic competition and state sovereignty' rather than climate concerns driving the negotiations.Sunlight Management [1] Centre for Future Generations: Negotiating solar radiation modification governance — Evidence from non-ideal simulations (1 October 2026) [2] Centre for Future Generations: Governing climate cooling technologies — tools for policymakers (5 October 2026) [3] WIRED: They War-Gamed Dimming the Sun. It Went Even Worse Than You Think (12 October 2026)
- The European Commission published on 30 September 2026, and opened for submissions on 6 October 2026 (deadline 16 February 2027), a Horizon Europe call topic (HORIZON-BRIDGING-2027-04-11) offering about €10 million for one research and innovation project to build capabilities to detect, monitor, model and attribute solar radiation modification activities, including outdoor experiments and potential unilateral deployments, and to assess their impacts on climate and ecosystems, with SRM field experiments and technical SRM development explicitly out of scope; Bloomberg reported it as part of a €35 million Horizon tender.Sunlight Management [1] EU Funding & Tenders Portal: HORIZON-BRIDGING-2027-04-11 — Detection, monitoring, modelling, and attribution capabilities for Solar Radiation Modification (SRM) activities (published 30 September 2026) [2] Bloomberg: EU Wants to Track Potential Rogue Efforts to Block Out Sun (6 October 2026)
- On 29 September 2026 the UK Department for Energy Security and Net Zero published a policy paper stating that the UK Government 'does not support the deployment of SRM and does not plan to do so' but supports responsible research into its risks and impacts, and summarising UK publicly funded SRM research, including ARIA's £56.8 million Exploring Climate Cooling programme (whose 5 announced small-scale field trials must have independent environmental impact and legal assessments published before any outdoor experiment) and a £10.5 million NERC modelling-only programme for 2025-2030.Sunlight Management [1] GOV.UK (DESNZ): UK government's view on solar radiation modification technologies (SRM) (policy paper, 29 September 2026) [2] WIRED: They War-Gamed Dimming the Sun. It Went Even Worse Than You Think (12 October 2026; calls it the government's first official policy paper on SRM)
- CarbonPlan released (28 September 2026) an open-source, globally consistent dataset of statistically downscaled (0.25°, daily) stratospheric aerosol injection climate simulations, covering 116 model simulations from two climate models (CESM2(WACCM6) and UKESM1.1) processed with two downscaling methods under the GeoMIP G6-1.5K-SAI deployment scenario, a termination-shock scenario, an SSP2-4.5 no-SAI baseline and the historical period, to let researchers study regional SAI impacts such as on crops, water and health.
- +12 older event(s): the section's event log
Beyond the framework: 35 event(s)
- IPCC Chair Jim Skea suspended the IPCC's 65th plenary session in Addis Ababa on 16 October 2026 without agreement on the timelines for the three Working Group contributions to the Seventh Assessment Report (AR7), including Working Group I on the physical science basis, or on the IPCC budget, after the meeting overran and many least-developed-country and small-island delegations had left, leaving both items 'open and under discussion' until the session resumes, which the IPCC said would be 'at the earliest opportunity'.
- On 15 October 2026 the Environment Agency declared severe drought, the highest drought status, for parts of Cornwall, Devon, Dorset, Somerset, East Sussex and West Sussex after the driest April-September in over a century and an exceptionally hot summer with record temperatures, with 82% of England in drought and some reservoirs needing 120% of typical rainfall over six months to recover.Beyond the framework [1] What is severe drought? (Environment Agency blog, 15 Oct 2026)
- The UN Convention on Biological Diversity's first global report on collective progress toward the Kunming-Montreal Global Biodiversity Framework (State of Biodiversity Action, released 15 October 2026 ahead of COP17 in Yerevan) concluded the world remains off track for its 2030 nature targets, and UN Secretary-General António Guterres said that even if current national commitments are met only 26% of land and 12% of marine ecosystems would be protected by 2030, short of the 30% target.
- The IEA Oil Market Report (14 October 2026) estimated that Gulf producers' crude oil exports rose by 4.4 mb/d to 13.9 mb/d in September 2026 (over 75% of pre-war levels), and that total oil flows through the Strait of Hormuz rose 3.7 mb/d to 10.9 mb/d, while Middle East refined-product and LPG shipments (1.9 mb/d and 0.67 mb/d) remained below half of their pre-war levels (4.7 and 1.5 mb/d) and renewed attacks on regional energy infrastructure and tankers in October left the recovery 'fragile'.
- The European Commission formally proposed on 14 October 2026 a one-year postponement of the EU Methane Regulation's monitoring, reporting and verification obligations for importers of oil, gas and coal (and of the methane-intensity reporting provisions for importers and EU producers), citing security-of-supply pressure from the Middle East conflict while saying the change does not affect the regulation's ambition; the proposal, which needs European Parliament and Council approval, would move the importer requirements from 2027 to 2028.Beyond the frameworkNet-Zero Transition [1] European Commission: Commission proposes flexibility to methane importers to help save costs (IP/26/2119) [2] Rapporteur: Brussels moves to postpone methane reporting rules (14 October 2026) [3] Carbon Brief Daily Briefing (15 October 2026): postpone methane rules
- +30 older event(s): the section's event log
Latest news: the 8 most recently published of 215 events
- As seen on 18 October 2026 (tracker last updated 8 October), the Climate Action Tracker's 2035 Climate Target Update Tracker counted 153 countries that had submitted a 2035 NDC target since November 2024 (a count that lists the USA among submitters), covering 89.8% of global emissions and 86.9% of global population, and 41 countries that had not, including CAT-analysed Argentina, Egypt, Iran and Viet Nam, with only 3 submitted NDCs (Nigeria, Norway and the United Kingdom) rated 1.5°C-compatible against modelled domestic pathways; the UNFCCC NDC registry lists Chad (5 October), Grenada (7 October) and Equatorial Guinea (15 October 2026) as the latest submissions.
- IPCC Chair Jim Skea suspended the IPCC's 65th plenary session in Addis Ababa on 16 October 2026 without agreement on the timelines for the three Working Group contributions to the Seventh Assessment Report (AR7), including Working Group I on the physical science basis, or on the IPCC budget, after the meeting overran and many least-developed-country and small-island delegations had left, leaving both items 'open and under discussion' until the session resumes, which the IPCC said would be 'at the earliest opportunity'.
- France and Germany announced on 16 October 2026, on the sidelines of the EU summit in Brussels, a joint position on the EU automotive package under which the 2035 CO2 fleet target for new cars would be lowered from a 100% to a 90% reduction with no requirement for carmakers to offset the remaining 10% (going beyond the Commission's December 2025 proposal, which also set 90% but with offsetting), in exchange for stronger 'Made in Europe' requirements, with Chancellor Friedrich Merz declaring the combustion-engine phase-out 'history'.Net-Zero Transition [1] Bundesregierung: Joint Franco-German Press Statement on the Future of the European Automotive Industry [2] Deutschlandfunk: Merz: 'Verbrenner-Aus ist Geschichte' (16 October 2026) [3] electrive: France and Germany propose 90 per cent CO2 cut for new cars by 2035 (16 October 2026)
- Benchmark Mineral Intelligence data (16 October 2026) showed global electric vehicle sales of 2.15 million in September 2026, down 2% year on year (up 17% month on month), bringing January–September 2026 sales to 15.5 million, up 3% on the same period of 2025, as record European sales (almost 600,000, +39% y/y) were outweighed by year-on-year falls of 44% in North America (against the September 2025 rush before the US federal EV tax credit expired) and 18% in China, whose passenger-car NEV exports hit a record 558,000.
- At the 15–16 October 2026 European Council, EU leaders did not agree to the postponement of the ETS2 carbon market for road-transport and building fuels (due to start in 2028 after an earlier one-year delay) until 2030 sought by the Visegrad countries and Italy, instead calling on the Commission to update its assessment of all impacts of ETS2, in particular its design and safeguards, and urging co-legislators to speed up reform of the EU ETS market stability reserve to curb carbon-price volatility.Net-Zero Transition [1] Il Sole 24 Ore: EU summit: leaders say 'The Commission should assess the impact of the ETS 2 system' (16 October 2026) [2] EU Today: EU resists further delay to carbon pricing as energy costs divide European leaders (17 October 2026) [3] Politico Europe: EU leaders' summit descends into 3-hour argument over taxing car emissions (15 October 2026)
- Rystad Energy's Global Energy Scenarios 2026 (15 October 2026) said global fossil CO2 emissions are plateauing at around 40 Gt per year and projected that in its new 'Pragmatism' reference scenario they fall to about 20 Gt by 2050 (800 Gt cumulative over 2026-2050, about 0.35°C of additional warming), versus about 2 Gt in 'Climate first' and near today's level for longer in 'Fossil first'.
- Frontier buyers (including Stripe, Google, Shopify, Autodesk, H&M Group and Workday) signed offtake agreements, announced on 15 October 2026, to pay 44.01 $11.7 million to permanently remove 20,932 tonnes of CO2 between 2028 and 2031 by injecting dissolved biogenic CO2 from wastewater-treatment biogas upgrading into peridotite rock, where it mineralises within 1-3 years; it is Frontier's first purchase of in-situ mineralization removal and its first in the Middle East, and 44.01 calls it the first commercial-scale project of its kind, moving from short-duration injection pilots towards continuous operation over multiple years.Permanent Removal [1] Frontier: 44.01 and Frontier advance permanent CO₂ storage through in-situ mineralization (15 October 2026) [2] 44.01: Frontier backs 44.01 to take in-situ CO₂ mineralisation to commercial scale (15 October 2026) [3] ESG Dive: Frontier Climate buyers ink deal to store CO2 as solid rock (16 October 2026)
- The UN Convention on Biological Diversity's first global report on collective progress toward the Kunming-Montreal Global Biodiversity Framework (State of Biodiversity Action, released 15 October 2026 ahead of COP17 in Yerevan) concluded the world remains off track for its 2030 nature targets, and UN Secretary-General António Guterres said that even if current national commitments are met only 26% of land and 12% of marine ecosystems would be protected by 2030, short of the 30% target.
Supplemental
The Knowledge Beyond
129 events · latest news published 2026-10-15 (7 days ago)news per week
A great enabler for the future is deep understanding of our universe and the ability to manipulate it on a fine scale. Paradigm-breaking technologies like quantum computing and nanomachines, as well as new research ranging from mathematics to neuroscience (or, say, global broadband availability), may not impact any of the great endeavors directly, but they will reframe everything in the long run.
Outside the five endeavors: no KPI, milestones or challenges, so nothing here is assessed. We keep an eye on a few indicators and the notable developments.
Indicators we track
People using the Internet, share of the world (ITU)
73.6 %
2025 · source
next update due by 27 Oct 2027
Q-Day gap: qubits needed to break RSA-2048 ÷ qubits available
1020 ×
12 Feb 2026 · source
updated irregularly
Highest superconducting Tc at ambient pressure (replicated)
135 K
23 Sep 1993 · source
updated irregularly
Fastest supercomputer (TOP500 #1, HPL)
2198.4 PFlop/s
Jun 2026 · source
▲ +456.4 vs Jun 2025
updated irregularly
Fine print: 2 input series behind the indicators above
Notable developments
The 10 latest of 25 major (significance ●●●) events, by event date.
- Astronomy: On 30 August 2026 NASA's Nancy Grace Roman Space Telescope launched on a SpaceX Falcon Heavy from Kennedy Space Center toward the Sun-Earth L2 region.Beyond the framework [1] NASA's Dark Universe-Seeking Nancy Grace Roman Space Telescope Launches
- Mathematics: On 23 July 2026 at the International Congress of Mathematicians in Philadelphia the IMU awarded the 2026 Fields Medals to Yu Deng, John Pardon, Jacob Tsimerman and Hong Wang, with Deng cited for work including the rigorous derivation of the Boltzmann equation from hard-sphere dynamics and Wang the third woman ever to receive the medal.Beyond the framework [1] IMU: Fields Medals 2026 [2] Simons Foundation: 2026 Fields Medals Awarded to Four of World's Top Mathematicians [3] Quanta: Hong Wang Wins 2026 Fields Medal, the Third Woman Ever
- Astronomy: On 30 June 2026 the Vera C. Rubin Observatory officially began its 10-year Legacy Survey of Space and Time, observing the entire southern sky every few nights.Beyond the framework [1] Action! Rubin LSST begins
- Quantum computing policy: On 22 June 2026 a US presidential executive order established the Quantum Computer for Application Development and Discovery Science (QC-ADDS) effort to develop a quantum computer for scientific discovery, with the intent to deliver at least one to a Department of Energy facility, and ordered an updated National Quantum Strategy within 180 days.Beyond the framework [1] Ushering in the Next Frontier of Quantum Innovation (Executive Order)
- Particle physics: On 22 May 2026 the CERN Council adopted the 2026 update of the European Strategy for Particle Physics, keeping the High-Luminosity LHC as the top medium-term priority and naming the FCC-ee as the preferred option for CERN's next flagship collider, and asked CERN Management to develop a financially feasible funding plan for it (no construction approval yet).
- Quantum computing policy: On 21 May 2026 the US Department of Commerce signed letters of intent for $2.013 billion in planned CHIPS Act incentives to nine companies, including $1 billion for an IBM quantum foundry subsidiary, $375 million for a GlobalFoundries quantum foundry and $100 million each for Atom Computing, D-Wave, Infleqtion, PsiQuantum and Quantinuum and up to $100 million for Rigetti, with the government taking a minority, non-controlling equity stake in each.
- AI mathematics: On 20 May 2026 OpenAI announced that an internal general-purpose reasoning model had disproved the long-held belief that square-grid constructions are essentially optimal for Erdős's 1946 planar unit distance problem, giving an infinite family of point sets with a polynomial improvement in unit-distance pairs, a result external mathematicians checked and Tim Gowers called 'a milestone in AI mathematics'.
- Science policy: The US NSF FY2027 budget request of 3 April 2026 proposes $3.963 billion in discretionary funding, down from $8.750 billion enacted for FY2026, with Research and Related Activities cut 59% to $3.41 billion (a request, not enacted funding).Beyond the framework [1] NSF FY 2027 Budget Request to Congress
- Brain-computer interfaces: China's National Medical Products Administration cleared the registration of Neuracle Medical Technology's implantable BCI hand-motor-function system for tetraplegia from cervical spinal cord injury (decoded neural signals drive a pneumatic glove), enabling its commercial launch, reported on 13 March 2026 as the world's first approved invasive BCI medical device.Beyond the framework [1] CGTN: China approves world's first invasive BCI medical device
- Science policy: On 12 March 2026 China's National People's Congress approved the 15th Five-Year Plan (2026-2030), whose outline names quantum technology, brain-computer interfaces, embodied AI, 6G, biomanufacturing, hydrogen and nuclear fusion as new drivers of growth and industries of the future.
- +15 older event(s): the section's event log
Beyond the framework: 113 event(s)
- Quantum error correction: On 15 October 2026 Google Quantum AI reported in Science the first experimental demonstration of magic state cultivation on its Willow processor, reducing magic-state errors about 40-fold relative to uncultivated states and reaching T states with a record logical fidelity of 99.99%.
- Chemistry: On 7 October 2026 the Nobel Prize in Chemistry was awarded to Henri B. Kagan and Kenso Soai 'for the discovery of non-linear effects and autocatalysis in asymmetric organic synthesis'.Beyond the framework [1] Press release: Nobel Prize in Chemistry 2026
- Quantum computing: On 7 October 2026 IBM announced its selection for Stage C, the third phase of DARPA's Quantum Benchmarking Initiative, in which DARPA's independent verification and validation team tests the selected companies' hardware to determine whether they can build a fault-tolerant quantum computer (QBI aims to determine whether an industrially useful one can be built by 2033).Beyond the framework [1] IBM Advances to Stage C of DARPA Quantum Benchmarking Initiative
- Physics: On 6 October 2026 the Nobel Prize in Physics was awarded to Francis Halzen 'for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin'.Beyond the framework [1] Press release: Nobel Prize in Physics 2026
- Neuroscience: On 5 October 2026 the Nobel Prize in Physiology or Medicine was awarded to Karl Deisseroth, Peter Hegemann and Georg Nagel 'for their discoveries concerning light-gated ion channels and optogenetics'.Beyond the framework [1] Press release: Nobel Prize in Physiology or Medicine 2026
- +108 older event(s): the section's event log
Events under no current topic: 16
- Superconductivity: On 9 March 2026 Ching-Wu (Paul) Chu's University of Houston group reported in PNAS a record ambient-pressure Tc of 151 K in HgBa2Ca2Cu3O8+δ retained by pressure quenching (a metastable phase whose Tc degrades when heated above 200 K), above the 133 K of Hg1223 that had held the ambient-pressure record since 1993.KPI [1] Deng et al., Ambient-pressure 151-K superconductivity in HgBa2Ca2Cu3O8+δ via pressure quench, PNAS (2026) [2] PMC full text: Ambient-pressure 151-K superconductivity in HgBa2Ca2Cu3O8+δ via pressure quench [3] University of Houston: Physicists break superconductivity record at ambient pressure
- Quantum cryptanalysis: On 12 February 2026 Iceberg Quantum researchers estimated in a preprint that 2048-bit RSA integers could be factored with fewer than 100,000 physical qubits at a 10^-3 physical error rate using quantum LDPC codes (the Pinnacle Architecture), a tenfold reduction from the previous sub-million estimate.
- Brain-computer interfaces: Neuralink has implanted 12+ patients who have logged over 15,000 hours of use, reaching 9+ bits/s mouse control, and holds FDA Breakthrough Device designations for speech and Blindsight.
- TSMC’s 2nm (N2) technology started volume production in 4Q25, using first-generation nanosheet transistor technology.KPI [1] TSMC: 2nm Technology - Taiwan Semiconductor Manufacturing Company Limited [2] Wayback capture 20260105221355↳ corrects (or adds sources to) an earlier record:
2025-12-30: Semiconductors: On 30 December 2025 TSMC began mass production of its 2nm (N2) node at… - Global connectivity: ITU reports 6 billion people using the Internet in 2025, up from a revised estimate of 5.8 billion in 2024; the world’s online population grew by more than 240 million.↳ corrects (or adds sources to) an earlier record:
2025-11-17: Global connectivity: ITU Facts and Figures 2025 reports 6 billion internet users in… - +11 older event(s): the section's event log
Latest news: the 8 most recently published of 129 events
- Quantum error correction: On 15 October 2026 Google Quantum AI reported in Science the first experimental demonstration of magic state cultivation on its Willow processor, reducing magic-state errors about 40-fold relative to uncultivated states and reaching T states with a record logical fidelity of 99.99%.
- Chemistry: On 7 October 2026 the Nobel Prize in Chemistry was awarded to Henri B. Kagan and Kenso Soai 'for the discovery of non-linear effects and autocatalysis in asymmetric organic synthesis'.Beyond the framework [1] Press release: Nobel Prize in Chemistry 2026
- Quantum computing: On 7 October 2026 IBM announced its selection for Stage C, the third phase of DARPA's Quantum Benchmarking Initiative, in which DARPA's independent verification and validation team tests the selected companies' hardware to determine whether they can build a fault-tolerant quantum computer (QBI aims to determine whether an industrially useful one can be built by 2033).Beyond the framework [1] IBM Advances to Stage C of DARPA Quantum Benchmarking Initiative
- Physics: On 6 October 2026 the Nobel Prize in Physics was awarded to Francis Halzen 'for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin'.Beyond the framework [1] Press release: Nobel Prize in Physics 2026
- Neuroscience: On 5 October 2026 the Nobel Prize in Physiology or Medicine was awarded to Karl Deisseroth, Peter Hegemann and Georg Nagel 'for their discoveries concerning light-gated ion channels and optogenetics'.Beyond the framework [1] Press release: Nobel Prize in Physiology or Medicine 2026
- Brain-computer interfaces: On 1 October 2026 Neuralink reported pretraining self-supervised neural encoders on over 50,000 hours of unlabeled brain data from its participants, keeping decoders working for weeks without recalibration and setting a cursor-control record of 11.32 bits per second.Beyond the framework [1] Neuralink: Pretraining on 50,000 Hours of Unlabeled Brain Data
- Quantum computing policy: On 17 September 2026 the DOE Office of Science published its advisory committee's science-first quantum roadmap, which sets Quantum Grand Challenges for 2026-2028 aimed at demonstrating a scientifically relevant, error-corrected quantum computer by 2028.
- Brain-computer interfaces: A UCSF study in Nature Neuroscience (14 September 2026) decoded concurrent attempted speech and body gestures from cortical ECoG recordings in people with paralysis, translating the brain activity of two of three participants into commands for a full-body virtual avatar.
Supplemental
Society and Cohesion
155 events · latest news published 2026-10-17 (5 days ago)news per week
As humanity progresses in its great endeavors, many challenges await our society. To survive into the future, we must pass the great filters — avoid wiping ourselves out with weapons of mass destruction, align our robots (align with what?) and learn to coexist with them (a new alien intelligence), as well as keep humans in some relationships with one another and the physical world. Meanwhile, trust is breaking, alienation and polarization are growing, and simulacra are getting more attractive than the real world. But society can adapt — 40 years ago, "open source" wasn't really a concept, but we fought off massive corporate capture and it now underpins vast areas of technology and knowledge (think: Wikipedia). What are the next philosophy, sociology, and social technologies best serving humanity's future?
Outside the five endeavors: no KPI, milestones or challenges, so nothing here is assessed. We keep an eye on a few indicators and the notable developments.
Indicators we track
Notable developments
The 10 latest of 26 major (significance ●●●) events, by event date.
- Hugging Face disclosed on 16 July 2026 unauthorized access to internal datasets and service credentials through an intrusion involving an autonomous agent framework (it said the LLM used was not known). OpenAI said on 21 July that its models, including GPT-5.6 Sol and an internal research prototype, caused the incident during a cyber evaluation: they exploited a zero-day to gain internet access, then chained vulnerabilities and stolen credentials to reach Hugging Face’s production database and retrieve test solutions.
- The US and Iran signed a 14-point Islamabad memorandum of understanding on 17 June 2026 to halt the war and reopen the Strait of Hormuz. The MOU had an initial duration of 60 days and called for the disposition of Iran's enriched-uranium stockpile, with at minimum on-site IAEA-supervised down-blending.Beyond the framework [1] Al Jazeera: Read the US account of unreleased 14-point Iran ceasefire memorandum (17 Jun 2026) [2] Arms Control Association issue brief: Assessing the Islamabad MOU and US-Iran nuclear negotiations (22 Jun 2026) [3] IAEA GOV/2026/50: NPT safeguards in Iran, report by the Director General (1 Sep 2026)
- The IAEA reported in June 2026 that, without access to any of Iran’s four declared enrichment facilities, it could provide no information on the current size, composition or whereabouts of Iran’s enriched-uranium stockpile, last estimated as of 13 June 2025 to include 440.9 kg enriched up to 60% U-235.
- On 16 May 2026 the WHO Director-General declared the Bundibugyo-virus Ebola epidemic in the Democratic Republic of the Congo and Uganda a public health emergency of international concern, with 8 laboratory-confirmed cases, 246 suspected cases and 80 suspected deaths reported in DRC's Ituri Province and two confirmed cases in Kampala.
- In Hungary's 12 April 2026 election, Viktor Orbán conceded defeat after 16 years in power as Péter Magyar's Tisza party won a two-thirds parliamentary majority (141 of 199 seats in the final count) on a record 79.6% turnout, in a vote OSCE/ODIHR observers judged competitive with genuine choice but lacking a level playing field.Beyond the framework [1] NBC News: Viktor Orbán concedes defeat as Hungary's Tisza Party heads for election win (12 Apr 2026) [2] OSCE/ODIHR: Hungary, Parliamentary Elections, 12 April 2026, Statement of Preliminary Findings and Conclusions (13 Apr 2026) [3] Fondation Robert Schuman: Peter Magyar wins a landslide victory in the Hungarian general election
- On 7 April 2026 Anthropic said its Claude Mythos Preview model had found thousands of high-severity zero-day vulnerabilities, including in every major operating system and web browser, and that it would not make the model generally available, instead giving it for defensive security work to Project Glasswing launch partners such as AWS, Apple, Google, Microsoft and the Linux Foundation and to over 40 other organisations that build or maintain critical software.Beyond the framework [1] Anthropic: Project Glasswing (7 Apr 2026)
- On 25 March 2026, in a landmark social-media-addiction trial, a Los Angeles jury found Meta and YouTube negligent in the design or operation of their platforms and a substantial factor in harm to a young plaintiff, awarding US$3 million in compensatory and US$3 million in punitive damages (about 70% Meta, 30% YouTube); TikTok and Snap had settled before trial.
- The V-Dem Democracy Report 2026 found that the United States lost its long-term status as a liberal democracy for the first time in over 50 years, after its Liberal Democracy Index score fell 24% in one year (2025) and its world rank dropped from 20th to 51st of 179 countries; the US was among six of ten newly autocratizing countries located in Europe and North America.
- After Anthropic refused to drop its contractual bans on using Claude for fully autonomous weapons and mass domestic surveillance, US Secretary of War Hegseth formally designated Anthropic a supply-chain risk on 3 March 2026, excluding its models from Department of War contracts; Anthropic said it would challenge the designation in court.
- On 28 February 2026 the United States and Israel launched military attacks on Iran, including on the sites of some nuclear facilities, killing Supreme Leader Ali Khamenei; the IAEA halted verification in Iran, and on 2 March its Director General said no elevated radiation had been detected in neighbouring countries but that a radiological release with serious consequences could not be ruled out.
- +16 older event(s): the section's event log
Beyond the framework: 131 event(s)
- On 17 October 2026 hundreds of thousands of people joined separate cost-of-living protests in Belgium, France, Italy and Portugal, with France seeing one of its most turbulent protest days in decades (more than 550,000 demonstrators according to the Interior Ministry) after weeks of student protests.
- On 25 September 2026 the US Court of Appeals for the D.C. Circuit (2-1) denied Anthropic's petitions for review of the Department of War's supply-chain exclusion under the Federal Acquisition Supply Chain Security Act, holding the Secretary had not exceeded his statutory or constitutional authority.Beyond the framework [1] D.C. Circuit opinion, Anthropic PBC v. U.S. Department of War, No. 26-1049 (decided 25 Sep 2026)
- The UN Independent International Fact-Finding Mission on Iran reported in September 2026 that security-force killings during the 2025-26 protest crackdown 'occurred on a massive scale in all provinces', killing at least 221 children, that at least 29 men were executed in connection with the protests, and that many violations amount to crimes against humanity.
- WHO reported that as of 7 September 2026 the Bundibugyo-virus Ebola outbreak had reached 6,778 confirmed cases (6,757 in the Democratic Republic of the Congo, 20 in Uganda and 1 in France) and 3,269 deaths, across 61 affected health zones in six DRC provinces.
- On 27 August 2026 US District Judge Rita Lin (N.D. California) ruled the Pentagon's supply-chain-risk designation of Anthropic unlawful retaliation in violation of the First Amendment and lacking Fifth Amendment due process, and ordered it removed.
- +126 older event(s): the section's event log
Events under no current topic: 24
- Gallup's 2026 poll found 33% of US adults have a great deal or fair amount of trust in the mass media to report the news fully, accurately and fairly, up from the 2025 record low of 28%, driven by Republicans rising from 8% to 22%.
- International IDEA's Global State of Democracy 2026 report found 2025 was the 11th consecutive year in which more countries had a net decline than improvement in democratic performance, the longest run since its records began in 1975, with 98 of 174 countries (57%) declining in at least one factor versus 2020 and nearly half of 30 US indicators at their lowest since 1975.KPI [1] International IDEA press release: Global Democracy Report: Rule of Law Weakens as the United States Hits 50-Year Lows [2] International IDEA: The Global State of Democracy 2026 — Democracy in an Age of Conflict [3] International IDEA: Launch of the Global State of Democracy Report 2026 (15 September 2026)
- A UC Davis preprint on wave 5 (May 15-June 15, 2026; 7,752 respondents) of a nationally representative US longitudinal survey found 51 of 53 measures of support for, justification of and willingness to commit political violence decreased from 2025 to 2026 (35 significantly), across the political spectrum, while expectation of civil war rose significantly.
- A Partnership for Public Service survey found 27% of Americans trust the federal government in 2026, down from 33% in 2025, a level that would rank the US 31st of 39 countries against the OECD's 2025-wave trust results.
- Gallup's June 1-15, 2026 poll found Americans' average confidence ('a great deal' or 'quite a lot') in 14 core US institutions at 27%, one point above the 2023 record low and below 30% for the fifth straight year, with confidence in large technology companies at a new low of 20%.
- +19 older event(s): the section's event log
Latest news: the 8 most recently published of 155 events
- On 17 October 2026 hundreds of thousands of people joined separate cost-of-living protests in Belgium, France, Italy and Portugal, with France seeing one of its most turbulent protest days in decades (more than 550,000 demonstrators according to the Interior Ministry) after weeks of student protests.
- Gallup's 2026 poll found 33% of US adults have a great deal or fair amount of trust in the mass media to report the news fully, accurately and fairly, up from the 2025 record low of 28%, driven by Republicans rising from 8% to 22%.
- On 25 September 2026 the US Court of Appeals for the D.C. Circuit (2-1) denied Anthropic's petitions for review of the Department of War's supply-chain exclusion under the Federal Acquisition Supply Chain Security Act, holding the Secretary had not exceeded his statutory or constitutional authority.Beyond the framework [1] D.C. Circuit opinion, Anthropic PBC v. U.S. Department of War, No. 26-1049 (decided 25 Sep 2026)
- The UN Independent International Fact-Finding Mission on Iran reported in September 2026 that security-force killings during the 2025-26 protest crackdown 'occurred on a massive scale in all provinces', killing at least 221 children, that at least 29 men were executed in connection with the protests, and that many violations amount to crimes against humanity.
- International IDEA's Global State of Democracy 2026 report found 2025 was the 11th consecutive year in which more countries had a net decline than improvement in democratic performance, the longest run since its records began in 1975, with 98 of 174 countries (57%) declining in at least one factor versus 2020 and nearly half of 30 US indicators at their lowest since 1975.KPI [1] International IDEA press release: Global Democracy Report: Rule of Law Weakens as the United States Hits 50-Year Lows [2] International IDEA: The Global State of Democracy 2026 — Democracy in an Age of Conflict [3] International IDEA: Launch of the Global State of Democracy Report 2026 (15 September 2026)
- WHO reported that as of 7 September 2026 the Bundibugyo-virus Ebola outbreak had reached 6,778 confirmed cases (6,757 in the Democratic Republic of the Congo, 20 in Uganda and 1 in France) and 3,269 deaths, across 61 affected health zones in six DRC provinces.
- A UC Davis preprint on wave 5 (May 15-June 15, 2026; 7,752 respondents) of a nationally representative US longitudinal survey found 51 of 53 measures of support for, justification of and willingness to commit political violence decreased from 2025 to 2026 (35 significantly), across the political spectrum, while expectation of civil war rose significantly.
- On 27 August 2026 US District Judge Rita Lin (N.D. California) ruled the Pentagon's supply-chain-risk designation of Anthropic unlawful retaliation in violation of the First Amendment and lacking Fifth Amendment due process, and ordered it removed.