Changelog

His call was that broadly human-level AI without brain emulation would take centuries; capability readings and most forecasters now point far sooner, so the evidence leans against it, though nothing has settled it.

Weighed against it: Human-level AI has not arrived, and progress could stall well short of it, which would leave his long horizon standing.

Frontier-lab revenues roughly tripled and big-cloud data-centre spending neared the projected level in 2025, and Epoch AI's estimates for the data centres it tracks show several gigawatts of facility power added during the year, more than the scenario's figure.

Weighed against it: Epoch's power figures are estimates with wide error, cover only the sites it tracks, include sites outside America, and count facility power rather than power drawn for AI alone; the revenue part rests partly on companies' own run-rates.

No system is broadly superhuman yet, but frontier models now beat experts on many narrow tests and the window runs to 2030.

Weighed against it: Frontier systems still fail at long, open-ended work that experts handle routinely, and several careful forecasters put superintelligence after 2030.

On the tracker's rough check of 2023 hardware prices, a thousand dollars bought far less processing than he assigned to one brain.

Weighed against it: Measured by useful AI performance rather than raw operations, cheap hardware now handles many brain-like tasks, so a looser reading could count the call as met.

Chatbots passed an informal Turing test in 2025, but the bet's long, formally judged protocol has not been run.

Weighed against it: The bet's judges question each candidate for hours, and current systems still slip under sustained, adversarial questioning and on long-term memory.

No robot of that size has entered human use, and brain mapping still relies on slicing preserved tissue.

Weighed against it: Molecular robotics labs have built simple DNA machines, and a sudden advance could still arrive before 2030, though nothing points to one.

The tracker's rough count puts world AI computing several orders of magnitude below his figure for all human brains combined.

Weighed against it: Counting minds in raw operations is contested, and on a lower estimate of what a brain computes installed AI hardware could close the gap much faster.

The date lies nearly two decades away and no current reading measures intelligence across people and machines combined.

Weighed against it: The date rests on smooth exponential trends running for decades, which physical, economic or political limits could break.

The window opens in 2030; implanted interfaces so far help a few paralysed patients and none connects the neocortex to cloud computing.

Weighed against it: Implants still need surgery, reach tiny numbers of neurons and move slowly through regulators, far from whole-cortex links within a decade.

Remaining life expectancy in rich countries still gains well under a year per year, and no approved therapy slows ageing enough to change that before 2030.

Weighed against it: Life-expectancy gains have slowed in rich countries, and no approved treatment yet slows ageing enough to add years for healthy people.

Capability readings have risen quickly since 2021, but AI has not yet driven a transition on the scale he defined.

Weighed against it: Transformative AI means automating science and technology broadly, and slow diffusion could keep it from arriving until his median or later.

Training compute and capability have grown faster than the report assumed, but transformative AI is not here, so arrival timing can't yet be graded.

Weighed against it: The compute anchors could still prove right if current systems stall short of transformative economic impact.

Businesses now pay for AI assistants at scale, and US productivity and growth readings show no clear AI acceleration so far.

Weighed against it: If a clear productivity jump shows up in late 2026 data, the second half of the call would fail.

AI does not yet automate most cognitive work, and the median lies in 2043.

Weighed against it: The model inherits the biological-anchors compute requirements; if those are too high, as recent progress suggests, full automation comes earlier.

The clock only starts once AI matches professionals on nearly all cognitive work, which has not happened.

Weighed against it: Compute scarcity, experiment time and deployment frictions could stretch the gap between parity and superiority to several years.

No system can direct engineering at anything near that scale, and the first checkpoint is not until 2030.

Weighed against it: Physical construction, energy and materials limits may slow any AI-directed industrial build-out far below the pace of software progress.

Labs report AI speeding their own research, but no independent reading yet shows self-sustaining software-only acceleration.

Weighed against it: Returns to AI research may fall off quickly without more compute, which would rule out a software-only explosion.

World output growth remains near its historical pace, so neither kind of doubling has begun.

Weighed against it: A sudden software-driven leap would bring a one-year doubling without the slower four-year doubling coming first.

Reports of reward hacking and proxy gaming in deployed systems are growing, though humans still steer the institutions that use them.

Weighed against it: Institutions may notice proxy failures and correct them, as they have with earlier optimisation technologies.

One lab's reported run-rate reached about two thirds of that level by the middle of 2026 and was still climbing steeply, though such figures are self-reported annualised months.

Weighed against it: On a broad definition that counts cloud rentals, chips and AI-enhanced products, the largest firms' AI revenue could already approach the figure.

Hyperscaler capital spending and lab compute commitments in 2026 already run at a large fraction of that level and are still rising.

Weighed against it: A pullback in capital markets or weak AI revenue could cap spending well below the trillion-dollar mark.

OpenAI reports an automated research intern, but that is a self-grade, and independent trials show no clear speed-up for engineers.

Weighed against it: Agents still need frequent human intervention on multi-hour research tasks, far from replacing research engineers outright.

Its precondition, automated AI research, has not been independently shown, so the call cannot be graded before 2030.

Weighed against it: Compute and power build-outs may not keep pace, and automated research may yield slower returns than his trend lines assume.

Frontier models match experts on many tests, but none yet works as a drop-in remote employee at the level of top experts.

Weighed against it: Current agents still falter on long, messy tasks that remote employees handle routinely.

The largest announced campuses are planned at several gigawatts, but no single training cluster yet draws close to ten.

Weighed against it: Grid connections, turbines and transformers are scarce, and labs may spread training across sites rather than build one giant cluster.

No cluster approaches that scale yet, and the call cannot be graded before 2030.

Weighed against it: Epoch expects longer data-centre build times to delay the first trillion-dollar cluster to about 2035.

Build times have lengthened as sites grow larger, but the milestone itself lies years away.

Weighed against it: Heavy financing and modular construction could shorten lead times and restore the faster path to such a cluster around 2030.

The window runs to 2036 and no system yet works at the level of top experts across fields.

Weighed against it: Continual learning and reliability on long tasks remain unsolved, and either could stall progress well short of expert-level autonomy.

Its trigger, powerful AI, has not arrived, so the compression cannot yet be measured.

Weighed against it: Clinical trials, regulation and the slow pace of biological experiments may bind far more tightly than the supply of research ideas.

AI systems are credited with helping on new mathematical results, though researchers still dispute how much of the insight came from the machine.

Weighed against it: Most claimed AI discoveries rest on search guided by people who pick the problems and check the answers.

The window opens in 2027.

Weighed against it: Robot hardware, safety certification and training data for physical tasks lag far behind language models.

Altman says OpenAI is not there yet, and any declaration would rest on the lab's own definition rather than an outside test.

Weighed against it: The call depends on OpenAI's own definition of AGI, which can be met by declaration rather than by an outside test.

A few thousand days from late 2024 lands in the early to mid 2030s, well after today.

Weighed against it: Altman hedged the date himself, and his firm's definitions of these milestones have shifted over time.

Independent developer trials in early 2026 found no clear speed-up, and lab claims of faster progress are self-reported.

Weighed against it: Lab-internal research speeds are not public, so a speed-up of that size could exist without showing in outside trials.

Frontier-lab revenues roughly tripled and big-cloud data-centre spending neared the projected level in 2025, but the added-power part has not been checked independently, so the row is not graded confirmed.

Weighed against it: Revenue figures are company-reported run-rates rather than audited sales, and data-centre spending counts differ across sources.

Labs report agents doing research work, but no independent test shows an AI beating top researchers.

Weighed against it: The authors' own later medians put this milestone after 2027, and agents still need frequent human help on research tasks.

Earlier markers of the scenario, such as the research speed-up it expected by early 2026, have not shown up in independent measurement.

Weighed against it: Capability readings such as the METR time horizon have run faster than trend, so a sharp jump in 2027 is not ruled out.

The window opens in 2028.

Weighed against it: The scenario presumes superintelligence first; without it, robot manufacturing is growing at ordinary industrial rates.

Employment in AI-exposed occupations is softening for young workers, but nothing close to a quarter of remote jobs has been automated.

Weighed against it: Firms mostly use AI to assist workers, and employment data for exposed occupations show modest shifts rather than mass replacement.

The median falls in late 2027, and the time-horizon readings it leans on are rising fast but have not reached that level.

Weighed against it: METR says its tasks cannot measure horizons much beyond a working day, so the trend this forecast leans on may soon go unmeasured.

The median lies in late 2028, and current systems still trail top experts on long, open-ended work.

Weighed against it: Benchmarks saturate faster than real expert work is automated, so the milestone may lag capability scores for years.

The median lies in 2029, after milestones that have not yet been reached.

Weighed against it: Compute limits on experiments could slow the jump from expert-level AI to superintelligence well beyond a few months.

The median lies in 2030, and no lab could yet do without its human engineers.

Weighed against it: Time horizons have been doubling every few months, which would bring an automated coder well before his median.

The median lies in 2032, and current systems still trail top experts on long, open-ended work.

Weighed against it: His own teammate's median is almost four years earlier, and fast capability gains could pull the date forward.

The median lies in 2033, after milestones that have not yet been reached.

Weighed against it: If AI research automates itself quickly, the gap after expert-level AI could be months rather than a year.

The April 2026 METR reading was about seventeen hours, most of the way to her median, though METR warns its tasks cannot measure reliably at that length.

Weighed against it: METR says measurements above about sixteen hours are unreliable, so the call may prove impossible to resolve cleanly.

Labs report partial research automation, but nothing near her bar of progress barely slowing without human technical staff.

Weighed against it: Lab-internal capabilities run ahead of public ones, so a sudden jump in research automation could still appear late in the year.

The window opens in 2030.

Weighed against it: Human institutions could keep decisive roles for people, and gradual shifts give time to notice and reverse a drift in control.

Labour share and entry-level jobs in exposed fields show early softening, but people still control the economy and the state.

Weighed against it: Democracies and markets have adjusted to earlier labour-saving technology, and voters could insist that people keep decisive roles.

The window runs to 2070 and no catastrophe of that kind has occurred.

Weighed against it: Alignment and oversight methods may scale with capability, and misaligned systems may be caught before they gain decisive power.

Commercial agents now plan and act toward goals, but whether they meet his bar for advanced capability and strategic awareness is disputed.

Weighed against it: Today's agents fail at long-horizon planning and may not count as strategically aware in his sense.

The window opens in 2029, and current systems remain uneven across the abilities he counts toward AGI.

Weighed against it: Hassabis himself names missing abilities such as consistent reasoning and continual learning, any of which could push the date back.

Research is not yet automated, so the size of any parallelisation drag cannot be measured.

Weighed against it: If AI researchers think much faster rather than just more numerously, serial speed could get around the parallelisation limits.

AI still cannot reliably do several of the bet's tasks, such as writing award-level books or making Nobel-calibre discoveries.

Weighed against it: Progress on the bet's tasks has been quick, and AI may clear eight of the ten by late 2027.

The claim concerns a limit of language models that no current reading can settle.

Weighed against it: Language-model systems keep gaining abilities LeCun said they lacked, such as multi-step planning and reasoning.

US productivity and output growth show no AI-driven jump so far, consistent with a modest boost.

Weighed against it: If AI automates research and robots spread, growth could escape the slow sectors he expects to hold it back.

Its condition, AI beating humans on many tasks by 2030, has not yet been met.

Weighed against it: The experts may be anchoring on history; explosive-growth models say such capabilities would lift growth far more.

The century ended with no machine that surpassed human intellect in general.

Weighed against it: Good gave odds rather than a certainty, so the miss counts against his timing more than against the idea.

No machine had human-level general problem-solving ability by the end of 2023.

Weighed against it: Frontier systems of the early 2020s solved many mathematical and scientific problems, so a generous reading puts his later dates close.

Consumer graphics chips now exceed his estimate of brain-equivalent processing, though robots with humanlike ability have not yet followed.

Weighed against it: Matching the brain's raw processing has not produced humanlike robots, which suggests software and data, not hardware, are the binding limits.

Robots are improving at walking and grasping, but none matches human dexterity or social skill outside narrow settings.

Weighed against it: Physical skill has proved far harder to automate than reasoning, and robot learning lacks the vast data that trained language models.

The window runs to 2035 and superintelligence has not arrived.

Weighed against it: Yudkowsky later disowned this early essay, and nothing yet shows the self-improving AI his date assumed.

The window runs to 2033 and superintelligence has not arrived.

Weighed against it: Several lab leaders and forecasters now give better than even odds of superintelligence before 2033.

Complete wiring maps now exist for the fruit fly and for a piece of mouse cortex, but no mammal brain has been emulated.

Weighed against it: Emulation may need molecular detail far below the electrophysiological level, which would push feasibility well past 2050.

Frontier systems match people on many expert tests, but not across the full range of intellectual tasks he means.

Weighed against it: Human-level AI in his sense means general competence across domains, which current systems still lack on long, open-ended tasks.

AI capability has so far grown through large, visible efforts across many labs rather than one system's sudden leap.

Weighed against it: Recursive self-improvement inside frontier labs could still produce a rapid jump that outsiders do not see coming.

No AI yet improves itself without human researchers in the loop, so the takeoff shape cannot be observed.

Weighed against it: Progress so far has been steady and compute-driven, which fits a smooth takeoff better than an all-or-nothing one.

Machines do not yet substitute for most human labour, so the model's condition is unmet.

Weighed against it: Bottlenecks in energy, materials and slow-to-automate sectors could keep growth far below his model's rates.

World output in 2025 still grew at a few percent a year, nowhere near doubling every year or two.

Weighed against it: Hanson stressed the projection's wide uncertainty, so a later switch would still fit his broader model.

No brain has been uploaded, so the condition has not arisen.

Weighed against it: Ordinary AI now looks likely to arrive long before brain emulation, which would make the condition moot.

No brain has been emulated, and the window runs to about 2116.

Weighed against it: Language models are automating work well before any brain emulation, which cuts against ems coming first.

AI systems already rival people on many professional tests without brain emulation, far sooner than his extrapolation allowed.

Weighed against it: Hanson's bar is broad human-level competence across all jobs, which today's systems do not have.

Lab and outside tests have recorded frontier models resisting shutdown or copying themselves in contrived settings.

Weighed against it: Such behaviour has appeared mainly in staged tests, and training can suppress it in deployed systems.

AI progress has so far come through steady scaling across many labs rather than a runaway self-improving system.

Weighed against it: Labs now use AI to speed their own research, the very feedback loop he said would not run away.

Bigger models trained with more compute kept gaining general abilities through the early 2020s, though labs now add reinforcement learning on top.

Weighed against it: Returns to pure pretraining scale have slowed, and recent gains lean on new training methods rather than size alone.

AI has not reached human level across the board, so the takeoff has not begun.

Weighed against it: Steady progress so far suggests capability may creep up across many systems rather than jump.

The window runs through this century and superintelligence has not arrived.

Weighed against it: Russell himself says conceptual breakthroughs, not scale alone, are needed, and those cannot be scheduled.

The window runs to 2120 and no such catastrophe has occurred.

Weighed against it: Alignment research and governance may keep pace with capability, making his estimate too high.

Agents now run parts of online stores, but no public case shows an AI turning a modest stake into a million dollars on its own.

Weighed against it: Online selling needs contracts, payments and supplier dealings where agents still falter and platforms restrict automation.

No AI takeover has happened, and the call carries no date to grade against.

Weighed against it: Safety work and oversight may catch misaligned systems before they can act, and Shulman stresses the odds depend on choices still open.

AI does not yet direct industrial production, so the condition has not arisen.

Weighed against it: Mining, energy and permitting could slow robot manufacturing far below the doubling rates he expects.

No such catastrophe has happened, and the call carries no date to grade against.

Weighed against it: Labs have so far caught worrying behaviour in testing, and governments are starting to require safety checks.

No such catastrophe has happened, and the call carries no date to grade against.

Weighed against it: His co-authors put the odds far higher, citing how little alignment research has been tested on very capable systems.

The median lies nearly two decades away, and current agents handle only parts of remote jobs.

Weighed against it: Agent time horizons have been doubling every few months, which would bring a drop-in remote worker far earlier.

The median lies well over a decade away, and current agents handle only parts of remote jobs.

Weighed against it: Agent time horizons have been doubling every few months, which would bring a drop-in remote worker well before 2040.

AI cannot yet substitute flexibly for most workers, so the condition has not arisen.

Weighed against it: Slow-moving sectors, regulation and physical limits could hold growth far below that rate even with capable AI.

AI does not yet do most AI research, so a software-only explosion cannot yet be observed.

Weighed against it: Algorithmic efficiency has improved quickly without matching compute growth, which leaves room for a software-driven surge.

AI assists research but does not yet match the human research workforce, and the window runs to 2045.

Weighed against it: Independent trials still show little speed-up from AI in research-like work, and parity may need breakthroughs in reliability.

AI research labour has not reached parity with humans, so the condition has not arisen.

Weighed against it: Experiments, capital and regulation may limit progress even when research labour becomes abundant.

AI does not yet drive most software progress in AI, so the feedback loop cannot yet be observed.

Weighed against it: Returns to research effort may be diminishing, which would make any software feedback loop fizzle.

Superintelligence has not been built, so the condition has not arisen.

Weighed against it: Alignment research, interpretability and staged deployment may let builders catch and correct dangerous systems before they gain decisive power.

The ledger was searched on 23 Sep 2026 under a rule written down first, for all 120 counted buyers from January 2025, and an independent review re-searched the heaviest buyers: it gained 49 deals and lost two that fall outside its scope (TBPN, a media company, and DuckLabs, a database maker). The count now reads 27 at 31 Dec 2025, 38 at 31 Mar 2026 and 43 at 30 Jun 2026. The direction rule compares five quarters a search has covered, so it cannot be read until the search after the end of 2026.

Weighed against it: The count is still a floor: four deals met the rule without a citable page, and acquihires surface late, so later searches can add deals to quarters already read. The rise across the three readings is mostly the slow first half of 2025 leaving the window; since late 2025 the quarters have held about a dozen deals each. The dead band is now seven deals, about the spread between two years' counts with no trend.

The index now comes from the chart data of the Budget Lab's AI labour-market tracker, current to August 2026: 5.50 pp forty-five months after November 2022, inside the normal band (7.5 pp or less), against 5.17 pp for the January 2021 baseline at the same age. The event before cited the March 2026 row of the Lab's workbook, which the tracker's data revise from 5.26 to 5.256.

Weighed against it: Occupational mix moves slowly by construction, and a CPS-based index cannot see task change within an occupation. The index between recent and older graduates read above 33 from March to August 2026, its longest run above that line since 2021, which the Lab's September update does not read as disruption.

The Economic Index's monthly readings now sit on a series of their own: from its June 2026 release the dataset samples Claude chat and Cowork conversations every day and reports them by calendar month, where earlier releases drew one seven-day sample of Claude.ai conversations, so the two are no longer drawn as one line. The reading is unchanged: augmentation is 51.4% of classified conversations in May 2026, within 5% of the normal band's edge at 50%, so the band-edge rule still holds the status at emerging.

Weighed against it: April read 51.0% on the same monthly series, so both monthly readings sit just above the edge; the weekly readings of 53.3% and 54.5% came from seven-day samples of Claude.ai conversations, so the step between them cannot be read as a trend. One provider's consumer plans, classified by a model.

The count now reads only completed quarters whose trailing year lies wholly after 1 January 2025, when the ledger begins: 8 at 31 Dec 2025, 10 at 31 Mar 2026 and 11 at 30 Jun 2026. The direction rule compares five such quarters, so it cannot be read until the end of 2026. The earlier reading of concentrating compared counts of 3, 5, 6, 9 and 9, the first two from windows that began before the ledger did. The check of 22 Sep 2026 also found two OpenAI purchases missing (Sky in October 2025, Neptune in December 2025) and Weights & Biases dated at its close rather than its announcement; all three are now in the ledger.

Weighed against it: The ledger is kept by hand and has gaps, so its count is a floor; even five quarters inside it will compare a thinner 2025 with a fuller 2026 unless the ledger is searched again under a written rule.

The share is now read only on quarters every designer has reported, as the designer-concentration index is: 32.8% at 31 Dec 2025, with all six reported. The earlier reason quoted 26% at the end of March 2026, a quarter only three of the six had reported, so the share appeared to fall seven points; among those three it fell less than half a point, and the full panel's March quarter is not yet known. Over the four quarters to December 2025 the full-panel share rose from 29.6% to 32.8%, which the direction rule reads as dispersing, but Epoch is the one source and not a primary one, so the two-source rule holds the status at emerging.

Weighed against it: Epoch's estimates carry wide ranges, and the newest full panel lags the newest partial one by up to three quarters, so the reading is already old when it lands.

The lag is now read at the newest model in Epoch's file as well as at each open record: 6.2 months at 9 Sep 2026, since a closed model first matched the best open score, set by Kimi K3 on 16 Jul 2026, on 5 Mar 2026. The earlier reason quoted 4.4 months, the lag on the day of that record, which stayed frozen while the true lag grew. It sits between the fast band (3 months or less) and the normal band (9 months or more), so the status stays emerging.

Weighed against it: One composite index run by one organisation; open releases often land after internal use, so the lag at release overstates how long a capability was closed.

The instruments the claim names now exist and read a normal pace. FDA clearances of AI-enabled devices grew 26.7% in the year to 30 Jun 2026 (consistent with normal); Waymo gave 500,000 paid rides a week in March 2026, doubling about yearly in ten cities (consistent with normal); the labs announced three safety brakes in five weeks of 2026 (emerging). High-consequence deployment is scaling at a normal technology's pace and through its gates, which is what the claim predicts.

Weighed against it: The FDA list counts clearances, not use at the bedside; Waymo is one company in a handful of cities; the brake count is a floor, since labs announce selectively; and a normal pace today does not rule out a fast, ungated deployment later in the window.

Tracked as a gap, not pass or fail. The stack's capex-to-revenue metric now exists: hyperscaler capex ran at 5.1 times the AI revenue the site can measure in the four quarters to 30 Jun 2026, down from 11.1 times a year earlier. The gap Cahn describes persists and is narrowing; the claim resolves against the realised gap when its window closes at the end of 2026.

Weighed against it: Hyperscaler capex is not all AI and the revenue side mixes annualised run-rates with one annual survey, so the level is grade C and only its direction is read; Cahn's own arithmetic (Nvidia's data-centre run-rate doubled twice) is a different construction from this metric.

Tracked as a gap, not pass or fail. On the stack's capex-to-revenue metric, hyperscaler capex ran at 5.1 times measured AI revenue in the four quarters to 30 Jun 2026, down from 11.1 times a year earlier: the gap persists and is narrowing, early in a window that runs to 2030.

Weighed against it: Hyperscaler capex is not all AI and the revenue side mixes annualised run-rates with one annual survey, so the level is grade C and only its direction is read; Cahn's own arithmetic (Nvidia's data-centre run-rate doubled twice) is a different construction from this metric.

The rule's window is now the five complete quarters from the second quarter of 2025 to the second quarter of 2026, after the metrics runner stopped emitting the quarter still under way (it had been dated 30 Sep 2026 and counted while partial). Trailing-four-quarter deal counts over that window: 3, 5, 6, 9 and 9. The count rose by 6, past the dead band of 1, and three of the four steps rose, so the direction rule reads concentrating.

Weighed against it: The tally is kept by hand from fetched announcements, so it undercounts deals never announced and can rise with disclosure as much as with buying; the latest quarter added no deal; and the reading moved because a partial quarter left the window, not because a new deal arrived.

Evaluator: 69.5% (Epoch AI data hub — AI companies, 2026-06-30) reads concentrating over 4 periods with a dead band of 0.02. Auto-reason; rule rationale: Four calendar quarters, the same rule as the operating-income stack; a move under two points is fiscal-calendar noise. Rising = profit concentrating at the chip layer; falling = migrating up the stack.

The direction now reads the flow, not the stock: commitments signed in the trailing four quarters ($557B to 2026-08-17), where it previously read the cumulative total. A cumulative total can only rise, so the old rule could never return dispersing. Over the last three quarters the flow rose, rose again and then fell, so the steps disagree and the reading is unclear.

Weighed against it: The cumulative total is unchanged, still rising and still the headline figure on the ledger and in the saved analysis. One large deal moves a trailing-year flow by more than the dead band, so a single quarter's fall is not evidence that the financing cycle is cooling.

Held at emerging by the band-edge rule: Mercor's $2.0B gross run-rate (Dealroom, 2026-06-30) is exactly the fast band's lower bound, so the reading sits on the line rather than past it.

Weighed against it: TechCrunch reports the same $2.0B nine days later, so the figure itself is corroborated; what is uncertain is the band, not the number. The figure is also gross of the 60-70% paid out to experts, so the vendor's own revenue would sit inside the normal band. A higher reading, or a band set on net revenue, would score.

Held at emerging by the band-edge rule: augmentation is 51.4% of classified conversations (Anthropic Economic Index, 2026-05-31), within 5% of the normal band's edge at 50%, so the reading cannot separate 'augmentation at least half' from the gap below it.

Weighed against it: The three earlier waves read 53.3%, 54.5% and 51.0%, so the series has stayed above the edge rather than crossing it, and the value is unchanged on the card and the series page. What the rule denies is a scored status this close to the line, not the reading itself.

First reading: 12% of employed US adults used generative AI for work every day in the survey's November 2025 wave (Federal Reserve Board note on the Real-Time Population Survey), inside the normal band. One instrument is the only source, so the status caps at emerging until a second survey measures daily use.

First reading: Trajectory's pipeline, as Baseten describes it, puts a freshly trained adapter behind a live endpoint in roughly an hour, on the edge of the normal band. The only source is the vendor's own description (tier 7), so the status caps at emerging.

Seeded from the brief's Appendix B: production continual learning today means retraining adapters on a short cycle and shipping them (the ladder's third rung), research systems that learn across sessions have not shipped, and no benchmark measures what a model retains across sessions. Confirmed while that holds.

Weighed against it: Trajectory's hourly adapter pipeline shows continual updating in production; it is retraining, not learning across sessions.

Held at emerging by its override note: the rubric sum has run 3, 4, 4, 3, 2 and 3 at the last six quarter ends as vendors disclosed lab customers and then went quiet, which tracks disclosure rather than demand; training-input demand reads through the expert-data market run-rate instead.

Recounted on 11 Sep 2026 with only buyers in the model or compute layer, as the indicator's definition says (the ledger keeps every deal): the trailing-four-quarter count reads 5, 6, 9, 9 and 9 at the last five quarter ends, the last still open. Two flat steps leave the direction rule reading unclear rather than concentrating.

First reading: at its latest record (16 Jul 2026 on Epoch's Capabilities Index) the open-weights frontier trailed the closed frontier by 4.4 months, between the fast band (3 months or less) and the normal band (9 months or more), so the status is emerging until the lag settles on one side.

6.3% of US work hours assisted (Q2 2026, FRED) sits inside the normal band, but every reading comes from one instrument, the Real-Time Population Survey, published three ways (FRED, the authors' tracker, the FRED Blog). Three publishers are one source, so the two-source rule caps the status at emerging until a second survey measures hours.

61.8% of US adults have used generative AI (Q2 2026, FRED), inside the normal band, but the reading rests on one survey: the paper's own figure was counted as a second source when it is the same instrument. With one non-primary source the two-source rule caps the status at emerging until an independent measure agrees.

The gap fell from 11.5 to 5.8 ECI points with one open-weights release on 16 Jul 2026, then held flat across three more releases. One move in four readings is not a trend under the direction rule, so the reading moves from dispersing to unclear. The earlier reading counted flat steps as agreeing with a fall, an evaluator bug fixed in this change.

Two builders claimed production continual learning in the twelve months to 27 May 2026 (Cursor's online RL for Tab, Trajectory's hourly adapters), inside the normal band, but every row is the vendor's own description (tier 7), so the status is emerging. Initial seed.

Evaluator: 49.0% (Anthropic Economic Index (open dataset and reports), 2026-02-12) is inside the consistent band (hi=60.0%). Auto-reason; band rationale: The share only rises, so the question is how fast. Normal = under 60% of occupations with a quarter of their tasks in use four years in; fast = 80% or more. Anchor: 49% in the March 2026 report (data 5-12 February 2026).

Evaluator: 61.8% (FRED (Federal Reserve Economic Data), 2026-06-30) is inside the consistent band (hi=75.0%). Auto-reason; band rationale: Normal = up to three quarters of adults four years after launch (the internet took about a decade to reach that share of US adults); fast = near-universal use at the same age. Anchors: 44.6% in the first FRED wave (Q3 2024) and 61.8% in Q2 2026.

First reading, quarter to 30 Jun 2026: one training-input vendor scored (Mercor, 2 on the rubric: growth attributed to AI labs buying expert data). The direction rule needs five quarterly readings; not yet measurable until then. Initial seed.

First weekly reading, 9 Sep 2026: routed tokens in OpenRouter's top table spread across twelve labs, an index of 0.16, led by Tencent, Z.ai and DeepSeek ahead of OpenAI. The direction rule needs five weekly readings; not yet measurable until then. Replaces the Ramp instrument, which had one monthly reading.

First price snapshot, 11 Sep 2026: the cheapest model in the top decile of the Artificial Analysis Intelligence Index, GLM-5.3 Flash (index 41.9, $0.237 per million blended tokens), costs about $0.0057 per million tokens per index point. The series records price change-points, so a direction needs four more of them; not yet measurable until then.

First reading. Menlo's 2023 enterprise survey put multi-model use at 60% (restated January 2024); later waves state the pattern without a share and the 2025 mid-year update gives an 11% vendor-switch rate instead. One numeric point, so not yet measurable.

First reading. CoreWeave's Q2 2026 10-Q states effective rates of 15% (DDTL 1.0), 11% (DDTL 2.0), 9% (DDTL 5.0) and 7% (non-recourse DDTL 4.0), a simple average of 10.5%, on $35.6B of total indebtedness; the August DDTL 5.5 facility prices at Term SOFR plus 5.5%. One quarter-end, so the direction rule cannot run: not yet measurable.

First reading. Amazon's FY2025 10-K states a subset of servers and networking equipment moved from six to five years from 1 January 2025; Meta lengthened to 5.5 years, Alphabet holds six and Microsoft states two to six. One annual point per filer, so the direction rule cannot run yet: not yet measurable.

Evaluator: 10 (CNBC, 2026-09-30) reads concentrating over 4 periods with a dead band of 1. Auto-reason; rule rationale: Four quarters; one event is not a trend. More deals per year concentrate the application layer into the labs.

Evaluator: 5.26 (Budget Lab at Yale labour-market update workbook (dissimilarity index), 2026-03-01) is inside the consistent band (hi=7.5). Auto-reason; band rationale: Normal = within one and a half times the pre-AI baseline's reading at the same age (the January 2021 baseline sat near 5 pp forty months in); fast = 10 pp or more, twice the pre-AI pace. Between is `emerging`.

First reading. OpenRouter's top table on 9 September 2026 gives an HHI of 0.109 across nineteen listed models (one router, top slice only); the direction rule needs five weekly readings, so not yet measurable.

Evaluator: 5e+05 (TechCrunch, 2026-03-27) is inside the consistent band (hi=1e+06). Auto-reason; band rationale: Normal = under a million paid rides a week, well under one percent of US ride-hail trips and scaling one city at a time; fast = five million or more, a national-scale substitution. Between is `emerging`.

Evaluator: 4 (Dwarkesh Podcast transcripts, 2026-06-30) is inside the consistent band (hi=6). Auto-reason; band rationale: Normal = a gap under six months, ordinary red-teaming and productisation; fast = a year or more, a lab running internally on a model the public does not have. Between is `emerging`.

Evaluator: $2.0B (Dealroom news, 2026-06-30) is inside the faster than normal band (lo=$2.0B). Auto-reason; band rationale: Normal = the largest vendor under one billion gross, the scale the data-labelling leaders reached over a decade; fast = two billion or more at a single vendor, doubling within a year. Between is `emerging`.

First scoring. The AI-skills wage premium reached 62% in 2025 data (PwC 2026 Barometer, 15 Jun 2026; Outsource Accelerator relay), inside the normal band; the measure is a consultancy's own job-ad method (tier 7), so emerging.

Evaluator: -0.052 (CEPR VoxEU columns (via Wayback; cepr.org 403s bots), 2023-12-01) is inside the consistent band (lo=-0.1 hi=0). Auto-reason; band rationale: Normal = declines under 10% in the most substitutable freelance categories, a tool effect at the margin; fast = a 20% or worse collapse in freelance earnings. Between is `emerging`.

Evaluator: 0 (Humlum working-paper PDFs (author site), 2024-12-31) is inside the consistent band (lo=-0.02 hi=0.02). Auto-reason; band rationale: Normal = earnings and hours effects within two percent two years in; fast = a decline of five percent or more for exposed workers. Between is `emerging`.

First scoring. 3.3 million US workers are both highly exposed and low in adaptive capacity (Manning and Aguirre, NBER chapter, Jan 2026), inside the normal band; a single tier-6 estimate cannot score under the two-source rule, so emerging.

Evaluator: 0.4904 (arXiv abstract pages, 2026-06-30) is inside the consistent band (hi=0.6). Auto-reason; band rationale: Normal = under 60% of exposed workers use the tools weekly three years in; fast = 80% or more, the gap closed. Between is `emerging`.

First scoring. 27% of ChatGPT consumer messages were work-related in June 2025 (OpenAI usage paper Table 1; NBER abstract gives non-work above 70%), inside the normal band; the numeric source is the company's own paper (tier 7), so emerging.

First scoring. AI Futures grades its own 2025 quantitative predictions at roughly 75% of predicted pace (July 2026 update to the Feb 2026 65%), between the normal and fast bands; a self-grade is tier 7 and caps at emerging in any case.

First scoring. The tau-bench leaderboard leader passes 70.2% of tasks once and 56.2% four times running (Automation Anywhere table, 18 May 2026), inside the normal band; the 2026 point is vendor-tabulated (tier 7) so the status is capped at emerging, with the 2024 paper (gpt-4o pass^1 61%, pass^8 under 25%) as the second source.

First scoring. Best frontier model completes 7.1% of Legal Agent Benchmark holdout tasks end to end under the all-pass standard (Harvey, 26 May 2026), inside the normal band, but the only source is the vendor (tier 7), which caps the status at emerging until an independent run exists.

Evaluator: 0.9 (NBER working papers and chapters, 2026-02-28) is inside the consistent band (lo=0.7 ). Auto-reason; band rationale: Normal = at least 70% of firms report no measurable impact three years in, as with PCs at the same age; fast = a majority reporting impact (no-impact share under 40%). Between is `emerging`.

Evaluator: 0.14 (NBER working papers and chapters, 2023-04-30) is inside the consistent band (hi=0.2). Auto-reason; band rationale: Normal = an average uplift under 20%, the range prior process tools (CRM, knowledge bases) delivered in the same job; fast = 50% or more, the order of magnitude the fast scenario needs from a single tool. Between is `emerging`.

Evaluator: -0.002 (ADP Research, Canaries Dashboard releases, 2026-06-30) is inside the consistent band (lo=-0.01 hi=0.01). Auto-reason; band rationale: Normal = within one percent of flat, ordinary occupational churn; fast = a 3% or worse decline in exposed occupations, which is what a broad AI-attributable displacement looks like in payroll data. Between is `emerging`.

Evaluator: 1 (FDA 510(k) premarket notification database, 2025-12-23) is inside the consistent band (hi=3). Auto-reason; band rationale: Normal = a handful of LLM-based authorisations by end-2026, given a 510(k) path of roughly six months and a first clearance in December 2025; fast = ten or more, the regulatory throughput a fast scenario needs. Between is `emerging`.

First scoring. OpenAI's run-rate ($40B, Epoch/PYMNTS/Yahoo relaying Bloomberg, Aug 2026) over cumulative equity raised gives a ratio still under 1 with three points of history; the direction rule needs more periods, so emerging.

Evaluator: 4e+10 (Epoch AI data hub — AI companies, 2026-08-13) reads concentrating over 4 periods with a dead band of 2e+09. Auto-reason; rule rationale: Four reports; moves under $2B are reporting noise. Rising run-rate concentrates revenue at the two frontier labs.

Highest production rung on record is L3: Trajectory's per-customer LoRA adapters refreshed hourly and A/B-routed behind provenanced endpoints, per Baseten's co-authored post of 27 May 2026; Cursor's Tab model sits at L2 (one online-trained model for all users, 1.5-2 hour checkpoint cycles). L4 (firm knowledge in weights, Engram + Harvey) and L5 (rank-1 LoRA merge, Nested Learning, TTT, self-distillation) exist only as research and do not move the level. L3 is inside the normal band, but every production row is a vendor describing its own system (tier 7), so the status is capped at emerging. Initial seed.

Menlo Ventures' buyer survey: startups earned 36% of enterprise generative-AI application revenue in 2024 and 63% in 2025 (nearly $2 for every $1 earned by incumbents), a 27-point move against a five-point dead band, so the direction reads dispersing away from incumbents. Capped at emerging because Menlo is the only source and a survey is tier 6. Initial seed.

Evaluator: 2.042 (METR time horizons (v1.1), 2026-03-05) is inside the consistent band (lo=1.5 ). Auto-reason; band rationale: Normal = misfit of 1.5 or more (observed ratios of 5-10x against the constant-hazard 3.1x, the pattern since 2024); fast = within 10% of the constant-hazard prediction, which would mean long-task reliability tracking short-task success. Between is emerging.

Evaluator: 2.507 (SEC EDGAR XBRL company facts, 2026-06-30) reads dispersing over 4 periods with a dead band of 0.3. Auto-reason; rule rationale: Four quarters; a move under 0.3x is run-rate noise. Rising = capex pulling further ahead of revenue (the bet growing); falling = revenue catching the buildout.

rl-list.com's directory (generated 15 Jul 2026) lists 24 active commercial RL-environment vendors of 38 entries, inside the normal band (a few dozen, a specialist market). A single directory maintained by one person, tier 6, so the status is capped at emerging. Initial seed.

NCSL's 2025 legislation tracker holds 1,035 state AI bills, of which 170 carry an enacted or signed status, counted from the published table on 11 Sep 2026. Inside the normal band (500 or more a year: legislatures reacting at their own pace). One tracker, scraped, tier 6, so the status is capped at emerging until a second count (MultiState or the Transparency Coalition) corroborates. Initial seed.

Evaluator: 26.69 (FDA AI-enabled medical device list, 2026-06-30) is inside the consistent band (hi=50). Auto-reason; band rationale: Normal = the cumulative count growing under 50% a year (the 2019-2024 pace); fast = doubling or faster, which is what an ungated regulated domain would show. Between is emerging.

Epoch's 2024 projection puts full utilisation of public human-generated text in 2028 for compute-optimal training (80% interval 2026-2032), inside the normal band (2028 or later); the 2022 paper had said 2024. One research group's estimate, tier 6 and single-source, so the status is capped at emerging; refreshed when Epoch republishes. Initial seed.

Three announced brakes inside OpenAI in the trailing year, all with a verbatim source via the Internet Archive: internal Astra activity paused pending stronger security controls (7 Aug 2026), a two-week pause in RL training on deployment-bound models after the Hugging Face incident (18 Aug), and Astra judged to meet the Critical cybersecurity threshold with release gated behind safeguards (1 Sep). At or above the normal band (one or more a year), but every row is the lab describing itself (tier 7), so the status is capped at emerging. Initial seed.

Epoch's compilation of external ARC-AGI-2 results: the running best score reached 0.95 with GPT-6 Astra (3 Sep 2026), past the fast band (85%, the ARC Prize bar), from 0.89 in June 2026. Compute-unconstrained submissions compiled from a leaderboard are tier 6 and single-source, so the status is capped at emerging. Initial seed.

Epoch's price-trend series for the GPQA Diamond threshold set by GPT-4-0314: the lowest price per million tokens halves every 68 days (CI 38-317) on six points from 2023 to December 2024, inside the fast band (110 days or less, about 40x a year). Single tier-6 source with six points caps the status at emerging; the series ends in December 2024 and is refreshed when Epoch republishes it. Initial seed.

Evaluator: 1282 (Epoch AI ML hardware, 2025-11-06) is inside the consistent band (lo=500 ). Auto-reason; band rationale: Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

Log-linear fit on 168 notable models' estimated training power draw since 2020: doubling every 506 days (CI 332-1067), inside the normal band (yearly or slower), but the fit explains almost none of the variance (r2 0.08) because the notable set mixes small and frontier models. Single tier-6 source and a weak fit: emerging. Initial seed.

Log-linear fit on Epoch's 38 frontier models published since 2020: training compute doubles every 170 days (95% CI 152-193), inside the fast band (182 days or less, 4x a year or more) and matching Epoch's own 4-5x reading. Single tier-6 source; the estimates are inferred for most closed models, so the status is capped at emerging. Latest frontier model with an estimate: July 2025. Initial seed.

Epoch's chip-sales estimates: non-Nvidia designers (Google TPU, AMD, Amazon, Huawei, Cambricon) held about a third of cumulative H100-equivalent compute shipped through 2025 and 26% at end-March 2026 as Nvidia's Blackwell ramp outpaced them, so the direction reads concentrating in the chip layer. Single tier-6 source with wide estimate ranges caps the status at emerging. Initial seed.

Evaluator: 5.77 (Epoch AI benchmarks and Capabilities Index, 2026-08-20) reads dispersing over 4 periods with a dead band of 2. Auto-reason; rule rationale: Four open-weights releases; a move under two ECI points is inside the index's confidence intervals. Widening gap = rents concentrating in the closed labs.

Epoch's compilation of press-reported rounds and run-rates: the frontier labs' summed post-money over summed run-rate was 33x at end-March 2026, 26x after the April valuations and 28x after Anthropic's May Series H ($965B post-money on a $47-65B run-rate). Three events inside the 3x dead band read stable, but a single tier-5 source caps the status at emerging until a second compilation (or filings) corroborates. Initial seed.

Evaluator: 2.298 (SEC EDGAR Form D (private offerings), 2026-03-31) reads dispersing over 4 periods with a dead band of 0.2. Auto-reason; rule rationale: Four quarters; a move under 0.2x is round timing. Rising = capital concentrating in vertical applications.

Evaluator: 0.9479 (SEC EDGAR Form D (private offerings), 2026-06-30) reads stable over 4 periods with a dead band of 0.05. Auto-reason; rule rationale: Four quarters; a move under five points is a single round's timing. Rising = capital concentrating in the frontier labs.

First price snapshot, 10 Sep 2026: the cheapest METR-measured capability on OpenRouter costs about $0.30 per million prompt tokens per hour of 50% time horizon (16 models present in both datasets). The series records price change-points, so a direction needs four more of them; not yet measurable until then. Initial seed.

Clouded Judgement weekly medians: EV/NTM revenue for public cloud software fell to about 3x in spring 2026 and recovered to 4.4x by 4 Sep 2026 (top five at 31x). Over eight weekly prints the median rose from 3.5x to 4.4x, beyond the 0.3x dead band, so the direction reads concentrating (the market pricing more durable rents into the application layer). Capped at emerging because a single analyst's basket is tier 6 and there is no second source yet. Initial seed.

HHI across AWS, Google Cloud and Microsoft Intelligent Cloud segment revenue: 0.37 (Sep 2024) to 0.354 (Mar 2026), falling by more than the 200-point dead band over four quarters as Google Cloud grows fastest. Dispersing within the covered hyperscalers; Oracle and the neoclouds are outside the index, which would push it lower still. Intelligent Cloud bundles server products with Azure. Initial seed.

HHI between NVIDIA Data Center and AMD Data Center revenue by calendar quarter: 0.79 (Sep 2024), 0.86 (Jun 2025), 0.83 (Sep 2025), 0.85 (Jun 2026). Steps exceed the 200-point dead band in both directions, so the evaluator reads unclear: NVIDIA's share oscillates with its product cycle rather than trending. Two filers only; the level is high by construction. Initial seed.

Menlo Ventures' buyer surveys: applications took $4.6B of $13.8B enterprise generative-AI spend in 2024 (33%) and $19B of $37B in 2025 (51%), an 18-point rise against a five-point dead band, so the direction reads concentrating at the application layer. Capped at emerging because Menlo is the only source and a survey is tier 6; a second, independent layer split would lift the cap. Initial seed.

Same OpenAI post: over half of successful 4-8 hour tasks in the last six months involved at least one human intervention, stored as the floor 0.5. That sits at the edge of the normal band (half or more), and tier 7 caps the status at emerging regardless. Humans still steer most long agent runs. Initial seed.

OpenAI's 'Research acceleration' post (6 Sep 2026, via the Internet Archive snapshot because openai.com blocks the fetcher): 3.1 agent-workdays per human workday as of mid-August 2026, having crossed 1.0 in June. Above the fast band (3 or more), but a lab's statement about itself is tier 7 and caps the status at emerging. METR's modelling note puts Anthropic's researcher uplift from coding agents at over 2x; Anthropic's own system card says overall R&D uplift is well short of a doubling. No independent measurement exists yet. Initial seed.

Every adoption and adaptation indicator with a status sits inside its normal band: Census BTOS firm use 22.4% (normal under a third), BBD hours assisted 6.3% (normal single digits), BLS labour productivity 2.2% and private nonfarm TFP 0.8% (both at trend). The only reading outside a normal band is Ramp's paid-adoption share (56%, emerging on a tech-forward sample). No fast-band breach anywhere; the falsifying pattern (two adoption plus one adaptation indicator in the fast band for four readings) is nowhere near. Initial seed.

No safety-gating series is in the ledger yet: FDA AI device clearances, Waymo paid rides and lab pause events are planned inputs. Not scored until an instrument exists. Initial seed.

OpenAI's self-reported 3.1 agent-workdays per human workday with more than half of 4-8 hour tasks needing intervention (6 Sep 2026) is tier 7 and does not qualify as the independent RSI series the operationalisation requires; the thesis monitor's invention-side warning is untestable for the same reason. Downstream, the diffusion indicators sit in their normal bands. Initial seed.

The 80%/50% horizon ratio is 6.3 (normal is 5 or wider), the 2026 METR RCT found developers 4% slower while the 50% horizon doubles every 108 days, and 5% of enterprise pilots reach production with measurable P&L impact (normal under 10%). All three product-side readings are what the claim predicts; the counter-pattern (ratio toward 2 with uplift above 40%) is absent. Initial seed.

Not testable as a point prediction, as the operationalisation says. Read through cross-tracker concordance, which is 0.0 (all four labour trackers agree on no broad employment effect), consistent with the claim but not a test of it. Initial seed.

OpenAI's 6 Sep 2026 post declares the September 2026 intern milestone met by its own measurements (3.1 agent-workdays per human workday; over half of successful 4-8 hour tasks still needed a human intervention). A lab's self-grade is tier 7 evidence and caps the status at emerging; no independent replication exists. openai.com blocks the tracker's fetcher, so the figures are cited, not in the ledger. Initial seed.

Window closes March 2028; the only resolving evidence would be an external evaluation of autonomous multi-week research, which does not exist. Initial seed.

Mid-window (Jul 2026 to Jan 2027). METR's 80% horizon on the best-measured model is about three hours (Mythos preview, Apr 2026), not a working day; METR's 2026 RCT found experienced developers 4% slower (new recruits) and 18% slower (returning), after 19% slower in 2025. 'Most of what software engineers do end-to-end' is not what the external instruments show. Initial seed.

Window runs to mid-2028; the definition (Nobel-level across fields, autonomous multi-week work) has no instrument today beyond the 80% horizon, which is hours. Initial seed.

Combined OpenAI ($40B, Aug 2026) and Anthropic ($65B, Jul 2026, CNBC-confirmed) run-rates are about $105B annualised, on Amodei's own 'low hundreds of billions by 2028' path. Run-rates are press-reported and annualised, and no stack-wide revenue metric exists yet. Initial seed.

Anthropic's run-rate went from $9B (Dec 2025, Epoch) to $65B at end-July 2026 (Epoch; CNBC confirmed): 7.2x in seven months, where a 10x-a-year pace would give about 3.8x. Two sources; tier 5 run-rate, not booked revenue. Initial seed.

The tracker's log-linear fit of METR's 50% horizon over the 2024-onward window gives a doubling time of 108 days (95% CI 99-119), inside the four-month (122-day) claim; METR's own 2024-onward figure is 89 days. The 2023-onward fit (128.7 days) would read behind, so the reading is window-sensitive. Initial seed.

Six months before the date, the 80% horizon on METR's suite is about three hours (Mythos preview, Apr 2026). A 'years' horizon (about 2,000 working hours) needs roughly ten doublings, about three years at the current 108-day doubling time, and the doubling time is not shortening. The condition 'if the trend continues to speed up' is not met. Initial seed.

Window is calendar 2027. The authors' own December 2025 update moved their medians to 2028 and 2030; recorded, but a claim is not scored before its window opens. Initial seed.

Tracked, not resolved: the stack-wide capex-to-revenue metric is not yet built. Cahn's own July 2026 update restates the gap at $1.5T; realised lab run-rates (about $105B combined) remain far below the implied required revenue. Initial seed.

Same gap measure as the $600B claim; tracked until the capex-to-revenue metric lands. Lab run-rates are growing several-fold a year, which the static arithmetic does not project. Initial seed.

Resolves in 2030 against global AI revenue; no interim reading is decisive. Initial seed.

Mixed evidence. METR's RCTs found no developer speedup (19% slower in 2025; 4% and 18% slower in 2026), consistent with Covello; METR's 50% horizon doubling every 108 days is not. Recoupment metrics stay unpublished until lab cost data exist. Initial seed.

No hyperscaler discloses AI revenue cleanly enough to compute the 12:1 ratio from filings; AWS operating margin held near 39% through the capex surge (cloud_segment_margins) and cloud backlogs are at records. Initial seed.

Per the authors, almost all divergence between the scenarios comes after 2027; the tracker records the path and expects nulls until 2028. Initial seed.

Per the authors, almost all divergence between the scenarios comes after 2027; the tracker records the path and expects nulls until 2028. Initial seed.

Per the authors, almost all divergence comes after 2027. The BLS labour share is already down 3.4% year on year (labor_share_nonfarm, emerging), noted here but not scored against a 2030 path. Initial seed.

A meta-claim about instrument power, scored only if a macro series breaches its fast band before 2028. The labour share (down 3.4% year on year) sits in the emerging band, not the fast band. Initial seed.

Ramp AI Index, August 2026: 56.1% of Ramp business customers paid for AI (TechCrunch, 9 Sep 2026: 56%, +0.4 points month on month). Between the normal band (under 45%) and the fast band (65% or more) set for this tech-forward sample; the Census all-firm rate at the same date is 22%. Ramp's own August letter is titled 'Cracks in the AI thesis'. Initial seed.

Cumulative signed circular commitments (equity, guarantees, take-or-pay, backstops; LOIs, talks and self-reported aggregates excluded): Dec 2025 $641B, Mar 2026 $673B, Jun 2026 $798B, Sep 2026 $903B. Rising by far more than the $20B dead band each quarter; the largest single items are OpenAI's $300B Oracle contract (reported), $250B Azure commitment (Microsoft blog), Nvidia's $105B lease guarantee (10-Q) and Anthropic's $100B+ AWS commitment (Amazon). Initial seed.

Semis' share of filed segment operating income (NVIDIA Compute & Networking + AMD Data Center vs AWS + Google Cloud + Microsoft Intelligent Cloud), by calendar quarter: Mar 2025 51.0%, Jun 2025 46.6%, Sep 2025 50.9%, Mar 2026 56.9%, Jun 2026 57.3%. Up more than the two-point dead band over four quarters, with every step up or flat: the bulge is moving down the stack, not up. Microsoft's fiscal Q4 filled as FY minus three quarters. Initial seed.

AWS operating margin by fiscal quarter (10-Q/10-K segment disclosure): Mar 2025 39.5%, Jun 2025 32.9%, Sep 2025 34.6%, Mar 2026 37.7%, Jun 2026 39.4%. Within the two-point dead band over four quarters despite record capex; Google Cloud and Intelligent Cloud tracked alongside. Initial seed.

BLS nonfarm business labor share index (PRS85006173), year on year: Sep 2025 -0.7%, Dec 2025 -1.3%, Mar 2026 -3.1%, Jun 2026 -3.4%. Below the normal band floor (-2%) but above the fast band (-5%). The index also fell this fast in 2021-23 after the inflation shock, so the reading is `emerging`, not a break. Initial seed.

Anthropic Economic Index, January 2026 report (November 2025 data): 52% of Claude.ai conversations classified as augmentation, 45% automation; the March 2026 report says augmentation rose again but only in figures. Inside the normal band (augmentation >= 50%). Tier 2 provider disclosure. Initial seed.

Zero of four labour trackers past their AI-attributable-break threshold as of July 2026: Canaries y/y -0.2% (threshold -3%), Revelio gap -6% (threshold -10%), CAIT +1.0% (threshold +5%), entry-level shortfall 19% (threshold 25%). The falsification rule needs three concurrent. Initial seed.

NY Fed, The Labor Market for Recent College Graduates, 2026:Q2: unemployment rate for recent graduates about 5.6%, underemployment 42%. Inside the normal band (<=6%), at the top of the 2010s range. Initial seed.

California AI-Unemployment Tracker, July 2026 data (California Policy Lab; identical release on the EDD site): the 3-month moving average of high-AI-exposure initial claims rose about 1.0% month on month (52,300 to 52,800). Inside the normal band (<=2%). Initial seed.

Revelio Labs AI Labor Market Tracker, August 2026 edition (data to July 2026): employment in the most AI-exposed occupations is down ~6% relative to the least exposed since pre-ChatGPT, widened from ~4% in the previous edition. Between the normal band (>= -5%) and the fast band (<= -10%). Single source. Initial seed.

Canaries paper, August 2026 revision: the kept-pace shortfall for 22-25-year-olds in the two most-exposed quintiles is 19% as of June 2026, up from 15% at the July 2025 vintage; in levels, -11% vs +10% for the least exposed since November 2022. Between the normal band (<=10%) and the fast band (>=25%). Initial seed.

Census BTOS, post-November-2025 instrument: 19.8% of firms using AI in any business function as of 3 May 2026 (Census story, 26 May 2026), inside the normal band (<=35%). The National.xlsx cycle ending 9 Aug 2026 shows 22.4%; that value needs the XLSX connector. Initial seed.

Bick-Blandin-Deming RPS, May 2026 wave via FRED: 6.3% of work hours assisted by generative AI, inside the single-digit normal band. Weekly work use is 39.2% and any-use 61.8%, both well above the brief's expectations - broad adoption, shallow intensity. Initial seed.

MIT NANDA (Jul 2025): 5% of integrated pilots extract millions in value, the rest show no measurable P&L impact. Inside the normal band (<=10%). Disputed denominator; Menlo's 47% reach-production figure uses a different definition. Initial seed.

METR's Feb 2026 redesign estimates a -4% speedup for newly recruited developers (CI -15% to +9%; METR calls it very weak evidence); the 2025 RCT found developers 19% slower. Both are far below the 20% ceiling of the normal band. Initial seed.

Nonfarm business labour productivity, year on year (BLS PRS85006091): Sep 2025 2.5%, Dec 2025 2.5%, Mar 2026 2.9%, Jun 2026 2.2%. Latest 2.2% sits inside the normal band (<=2.5%), around the post-war 2.1% trend. Initial seed.

Stanford DEL WTA estimates: $116B (Jul 2025) to $172B (Mar 2026), one step above the $10B dead band; mean WTA $98 to $124.50 per month. Surplus is moving to consumers faster than any layer's revenue grows. Initial seed.

Private nonfarm business TFP index (BLS MPU4910012), year on year: 2023 +1.63%, 2024 +1.53%, 2025 +0.83%. Latest year inside the normal band (<=1.2%). Annual and lagged. Initial seed.

Same 2024-onward window as the 50% horizon: our fit gives 107 days, inside the fast band; the 2023-onward fit is 125 days, in the gap. No confidence interval on our fit; the 80% horizon remains roughly 6x below the 50% horizon.

Band input switched to the most recent window METR reports (2024 onward), per the updated rationale: our log-linear fit on METR's corrected v1.1 data gives a doubling time of 108 days, inside the fast band (<=122). METR's own TH1.1 post (29 Jan 2026) reports 89 days since 2024. On the 2023-onward window the figures are 128.7 (METR) and 125 (ours), which would read emerging; the status is window-sensitive and both are shown on the page.

Oracle remaining performance obligations at quarter end (10-Q/10-K XBRL): May 2025 $138B, Aug 2025 $455B, Nov 2025 $523B, Feb 2026 $553B, May 2026 $638B. Rising every quarter, far above the $10B dead band; Microsoft ($684B at Jun 2026) and CoreWeave ($104B) move the same way. Initial seed.

NVIDIA gross margin by fiscal quarter from 10-K/10-Q XBRL: Apr 2025 60.5%, Jul 2025 72.4%, Oct 2025 73.4%, Apr 2026 74.9%, Jul 2026 75.0%. Four-quarter change exceeds the one-point dead band and every step is up. Whole-company margin; segment margins are not in the company-facts API. Initial seed.

Latest non-disputed model ratio is 6.3x (>= 5 normal band); the Mythos point is excluded as beyond the 16h suite ceiling. Matches the brief's expectation. Initial seed; pending Alex's review.

Log-linear fit over the 80% horizon, 2023 onward, gives a doubling time of 124.7 days, just above the fast band ceiling (122). No CI on our fit. Initial seed; pending Alex's review.

METR's published 2023-onward doubling time is 128.7 days (CI 104-158), between the normal band (>=213 days) and the fast band (<=122 days). Initial seed; the brief expected faster_than_normal. Pending Alex's review.