Indicators
Every gauge the site reads, with what it tells you. Unpublished rows have a definition and data but do not yet meet the publishing rules (a status, counterevidence, and either two sources or one primary source), so they carry no status and no page.
112 of 112
| Obs | |||||
|---|---|---|---|---|---|
| Rent concentration in semiconductors Nordhaus's base rate says innovators keep ~2% of surplus; a 70%+ gross margin at the bottom of the stack is the opposite pattern. Whether it holds, rises or erodes is the first capture question. | concentrating | 85 | grade A | 75.0%as of 2026-07-26(2 obs) | 144 |
| Cloud backlog (remaining performance obligations) RPO is the forward view of whether compute demand is contracted ahead of revenue. Its growth is the clearest filed evidence of capex-ahead-of-revenue, and of how much of it is one customer. | concentrating | 80 | grade A | $664Bas of 2026-08-31 | 76 |
| Frontier-lab revenue run-rates Whether the model layer keeps surplus depends on revenue growing faster than the capital it consumes. Run-rate is the only public series, and it is press-sourced. | concentrating | 55 | grade C | $50.0B⚑disputed: An annualised month reported by the press or the company, not booked or audited revenue, and often sourced to the company itself.run-rateas of 2026-09-30 | 37 |
| Lab recoupment ratio The financial form of the return arrow. If deployment revenue is recouping invention capital, the loop is closing; if capital keeps outrunning revenue, the layer is being subsidised. | emerginga model judges: probably stable. The site's readings swing around roughly a third of equity raised: a slump early in 2026 after a large raise, then a recovery by late September to about where the ratio stood at the end of 2025. Read over the rule's three points back to that level, the move sits inside the noise band, so the readings were extended as no clear movement. The strongest point the other way is the sharp climb across the last three readings, which would read as concentrating if it holds. | 50 | grade C | 0.27×as of 2026-09-30(12 obs) | 30 |
| Consumer surplus from generative AI (willingness to accept) Value that users get without paying for it is value no company captures. If what people would need to be paid to give AI up is larger than what every company earns from it, the biggest winner is the user. | dispersing | 65 | grade C | $172Bas of 2026-03-31 | 4 |
| Application layer's share of enterprise generative-AI spend (Menlo) The only published layer split of enterprise dollars. A rising application share means buyers' money is landing above the model layer, where the model-tax question starts. | emerginga model judges: probably concentrating. The two survey waves on the site show applications rising from about a third of enterprise generative-AI spend to roughly one dollar in two, far outside the survey-error band, and the readings were extended on the view that coding and agent products kept pulling money up the stack in 2026. The strongest objection is scope: the survey leaves out chips and embedded features, so the infrastructure side is understated and the shift may be smaller than it looks. | 55 | grade B | 51.4%as of 2025-12-31(2 obs) | 4 |
| Semiconductors' share of stack operating income The capture thesis in one number. "Rents migrate up the stack" means this share falls over four quarters while labs and apps rise; Nvidia holding its share means the bulge stays at the bottom. | concentrating | 70 | grade A | 61.0%as of 2026-06-30(8 obs) | 52 |
| Chips' share of the stack's gross profit Gross profit is what is left after paying to deliver the product, so a layer's share of it is the plainest answer to who keeps the money. If rents migrate up the stack as models commoditise, this share falls; if the scarce part stays scarce, it holds or rises. | concentrating | 60 | grade A | 72.6%as of 2026-06-30(16 obs) | 43 |
| Cloud segment operating margins Whether the cloud layer keeps its margin while capex explodes is the capex-to-revenue question in filed form; depreciation-life changes show up here first. | concentrating | 75 | grade A | 38.7%as of 2026-09-30(2 obs) | 26 |
| Circular financing scale When Nvidia guarantees OpenAI's leases, Microsoft's backlog is half OpenAI, and Amazon funds the customer that commits to AWS, revenue and capital are the same dollars going round the stack. The ledger is where the circularity is filed. | concentrating | 75 | grade B | $998Bas of 2026-09-30(34 obs) | 55 |
| Concentration among covered chip filers (HHI) Rents stay where concentration stays. AMD's data-centre ramp and custom silicon are the routes by which the chip layer's rent would erode; this is the first of them in filed numbers. | stable | 60 | grade A | 0.87 indexas of 2026-06-30(2 obs) | 24 |
| Concentration among covered cloud filers (HHI) Whether the cloud layer stays an oligopoly while labs and apps fragment is the P1 leading indicator of where capture flows next. | stable | 60 | grade A | 0.35 indexas of 2026-06-30(6 obs) | 34 |
| Lab concentration of routed tokens (OpenRouter, HHI) Multi-homing shrinks lock-in, which shrinks the labs' bargaining share; a falling index is the leading sign of model commoditisation in buyer behaviour. | emerginga model judges: probably stable. The weekly readings on the site wobble up and down and the latest sits only just above the first, a move that falls a shade inside the noise band, so they were extended as no clear movement. The level itself describes traffic spread across many labs. Against this, the last week rose sharply, and one more reading like it would tip the rule to concentrating. | 40 | grade B | 0.24 indexas of 2026-11-03(20 obs) | 178 |
| Public cloud software valuation multiples (Clouded Judgement)unpublished The valuation multiple is the market's expectation of durable rent at the application layer. Compression toward low single digits prices in commoditisation by AI; recovery prices in apps keeping their surplus. Cannot refresh unattended: Clouded Judgement's Substack feed refuses the nightly's cloud fetches (HTTP 403 on every run since 10 Sep 2026), so the weekly series would sit stale; the history stays on /series, and the indicator publishes again if a fetchable copy of the weekly table appears. | no status until published stale as of 2026-09-04 a model judges: probably concentrating. Over the two months of weekly prints on the site the median multiple climbed from the high threes to the mid range, well beyond the noise band, so the rule reads concentrating and the readings were extended. The strongest point against is that this is a rebound from a spring trough caused by fears that AI erodes software; over a full year the market has priced less durable rent into incumbent software, not more. | 50 | grade B | 4.40×as of 2026-09-04 | 40 |
| Price per unit of capability (cheapest METR-hour on OpenRouter) Price per unit of capability is the lab layer's capture proxy. Falling means the surplus from better models is passed to applications and users; flat or rising means labs keep it. | emerginga model judges: probably dispersing. The site holds a single price point, so this rests on general knowledge: the price of a given level of model capability has fallen steeply and repeatedly since 2023 as cheaper and open-weight models reach each threshold. That means buyers, not labs, have kept most of the surplus from capability gains. The counterweight is that the newest frontier models launch at firm or higher prices and agents consume far more tokens per task. | 40 | grade A | $0.30as of 2026-09-10(2 obs) | 2 |
| Price per unit of capability (cheapest near-frontier model, Artificial Analysis) The same capture proxy as the METR-hour price on an independent capability axis that covers hundreds of models, refreshed daily. Falling means the surplus from better models is passed on; flat or rising means labs keep it. | emerginga model judges: probably dispersing. The site's two readings point the other way: the price per index point rose to about two times its earlier level within September 2026, far beyond the noise band. The judgement discounts them because two readings are a single step, and the cheapest model in a shifting top slice can change simply because a model enters or leaves that slice or the index is re-based. It rests instead on the long record of steep price falls at fixed capability, which hands the surplus to buyers. If labs are starting to hold prices on near-frontier models, this call is wrong. | 40 | grade A | $0.011as of 2026-10-27(2 obs) | 4 |
| Frontier labs' share of tracked venture dollars Capital concentrating in the model layer's incumbents is the installation-phase pattern; capital spreading to applications and infrastructure is what a maturing stack looks like (P1 seed indicator 8). | stable | 50 | grade B | 91.2%as of 2026-09-30(41 obs) | 117 |
| Vertical versus horizontal application funding The brief's bet that value at the application layer accrues to domain-specific products (P1 seed indicator 11); the ratio says where investors are placing it. | unclear | 45 | grade B | 2.03×as of 2026-09-30(8 obs) | 22 |
| Frontier lab valuation to run-rate multiple The market's expected-rent multiple for the model layer. Thirty to forty times says durable rents are priced in; compression toward software multiples says commoditisation is. | emerginga model judges: probably dispersing. The readings on the site fall from a peak early in 2026 to a clearly lower multiple by September, a drop over the last three valuation events that is well outside the noise band, and they were extended. Revenue has been growing faster than valuations, so the market is paying for fewer years of future rent per dollar of sales. Against this, absolute valuations kept rising and run-rates lag, so the multiple can fall while expectations stay just as high. | 45 | grade C | 16.3×as of 2026-10-22(12 obs) | 61 |
| Neolab capital raised per dollar of revenueunpublished Overfunding relative to revenue is the pattern that preceded every prior installation-phase crash. Not measurable from free sources: the neolabs' rounds are filed or reported, but none has a revenue figure in Epoch or a filing, and Leonis Capital's count (evidence below) finds most disclose no revenue at all, so the denominator does not exist. | no status until publisheda model judges: probably concentrating. There are no readings, so this is general knowledge: labs founded since 2023 have raised very large rounds on little or no disclosed revenue, and the rounds grew through 2025 and into 2026 faster than any visible sales. Capital per dollar of revenue therefore looks high and rising. The strongest objection is arithmetic: once revenue starts from almost nothing, even modest sales pull the ratio down quickly, and nobody outside can see the denominator. | — | unmeasured | 0 | |
| Lab vertical integration (acquisitions and acquihires, trailing four quarters) The application layer's exit bell. When the labs buy the layer above them, the surplus that layer was capturing moves down the stack; a rising count is the capture thesis's clearest leading signal. | emerginga model judges: probably concentrating. The trailing-year deal count on the site rises at each of its three quarter-ends, by far more than the noise band, and the readings were extended. Labs and compute companies have been buying coding tools, evaluation software, agents and consumer apps at a steady clip. The main caution is that part of the rise reflects a more thorough search, and many deals are small team hires driven by one buyer. | 55 | grade A | 46as of 2026-06-30(46 obs) | 75 |
| Venture flow to revenue lagunpublished Whether venture money leads or follows revenue tells you whether the flow is a forecast or a reaction. Not yet measurable: the estimate needs at least eight quarters of both venture dollars and revenue for the same sub-layer; the venture series starts in 2023 and revenue exists only for the frontier labs. | no status until published | — | unmeasured | 0 | |
| Open-weights gap to the closed frontier (ECI points) Open-weight quality converging on the frontier is the single best leading indicator of model-layer commoditisation; a widening gap says the closed labs keep their rents. | unclear Evaluator reads stable since 2026-10-31; reason pending | 55 | grade A | 10.2 indexas of 2026-10-17(2 obs) | 149 |
| Non-Nvidia share of AI compute shipped (H100-equivalents) The leading indicator of Nvidia rent decay; TPUs, Trainium and AMD are the routes around the chokepoint. | emerginga model judges: probably dispersing. Across the last year of quarterly readings on the site the non-Nvidia share of cumulative compute rose by a little more than the noise band, so the readings were extended as dispersing, and general knowledge of large custom-chip programmes at the cloud companies points the same way for 2026. The strongest point against is that the rise had flattened to almost nothing by the latest quarters, and much of the non-Nvidia compute is used in-house, not sold. | 50 | grade C | 33.8%as of 2026-06-30(6 obs) | 60 |
| Hyperscaler capex to AI revenue (trailing year) Sequoia's $600B question as a ratio: how far the buildout runs ahead of the revenue that has to repay it. A wide gap can mean the spending is early and revenue will follow, or that the builders will never earn it back; the ratio cannot say which, only whether the gap is closing. | dispersing | 40 | grade A | 4.38×as of 2026-09-30(34 obs) | 98 |
| Hyperscaler server useful lives (10-K estimates) Depreciation is where the capex bet meets the income statement. A lengthening life flatters margins; a shortening one says the hardware is turning over faster than the accountants assumed. | emerginga model judges: probably stable. The site has one reading and its text says no filer shortened its server life in 2026; one company cut a subset the year before while another lengthened, and two held. That is divergence without a common direction, so the reading was extended as no clear movement. The strongest counter is the argument that fast chip release cycles will force shorter lives, which would show up suddenly, not gradually. | 65 | grade A | 5.00 yearsas of 2025-12-31 | 4 |
| Neocloud borrowing cost (CoreWeave facility rates) The compute build is financed at the margin by lenders. If neoclouds pay double-digit rates while hyperscalers borrow at treasuries plus a little, the capture story at the compute layer is a credit story. | emerginga model judges: probably stable. The site gives a single average rate, so direction rests on general knowledge. Newer facilities backed by strong customers have priced below the earliest ones, which pulls the average down, while credit markets grew warier of AI-related debt during 2026, which pushes spreads up; the two roughly offset. The risk to this call is that lender caution has since hardened, in which case the cost is rising and returns are concentrating with lenders. | 60 | grade B | 10.5%as of 2026-06-30(4 obs) | 12 |
| Enterprise multi-model usage share (Menlo) Multi-homing is the model layer's pricing power in one number: buyers who route across providers keep the surplus; buyers locked to one hand it over. | emerginga model judges: probably dispersing. The only numeric reading is from 2023, so this rests on general knowledge and the site's own note that the pattern has spread: large buyers routinely run several model families and add models more than they replace them. That weakens any one lab's pricing power. The strongest point the other way is that spend, as opposed to mere use, remains heavily concentrated in a few labs, so using many models is not the same as splitting the budget. | 45 | grade B | 60.0%as of 2023-11-10 | 5 |
| Startups' share of enterprise AI application revenue Whether incumbents with distribution keep the application layer or startups take it is the Teece question at the top of the stack; Menlo's survey is the only published split. | emerginga model judges: probably dispersing. The two survey waves on the site show startups moving from about a third of enterprise AI application revenue to nearly two thirds, far beyond survey error, and the direction was extended into 2026 given how quickly young application companies have grown. The strongest caveat is that the survey excludes AI features embedded in incumbents' existing products, and that the labs themselves are now moving into applications. | 50 | grade B | 63.0%as of 2025-12-31 | 2 |
| Frontier lab gross margin (estimated) Whether the model layer keeps a software margin or a cloud margin decides whether it is a chokepoint or a commodity. | emerginga model judges: probably dispersing. The site's two annual estimates show the margin falling by well over the noise band between 2024 and 2025, as inference spending grew faster than revenue, and that direction was extended. This is one estimator's figure and counts inference compute only. The strongest evidence against is press reporting, behind paywalls, that the labs' margins on paying customers improved over the same period, so the true direction could be the opposite. | 30 | grade C | 38.5%as of 2025-12-31(2 obs) | 4 |
| Compute share of lab spendingunpublished The labs' cost structure is what decides whether falling token prices are passed through or absorbed. Not measurable as a series: Epoch published a one-off range for three companies (evidence below), and no lab discloses its compute share over time. | no status until publisheda model judges: probably concentrating. No series exists, so this is general knowledge: compute already takes the larger part of frontier-lab spending, and the labs' multi-year compute commitments have grown faster than their headcount or revenue. That suggests the share going to compute suppliers is edging up. The main counterweight is the steep rise in pay and share-based awards for researchers, which may be growing just as fast. | — | unmeasured | 0 | |
| Application gross margin net of inference (model tax)unpublished The model tax in one number: whether falling token prices are retained by apps or passed through. Not measurable from free sources: Artificial Analysis now gives list price per token, but a margin needs each app's token volume and revenue, neither public, and list price overstates what large apps pay under committed-use discounts; SemiAnalysis unit costs and The Information's app financials are paid (stubs off). | no status until publisheda model judges: probably stable. There are no readings; from general knowledge, application margins after model costs were squeezed in 2025 as agent-style products burned far more tokens per task, and were then partly repaired by usage-based pricing, cheaper models and in-house models. The net of those forces looks like no clear movement at a thin level. The strongest alternative is that the squeeze is continuing for coding tools, which would read as dispersing away from the application layer. | — | unmeasured | 0 | |
| GPU rental price per hour The cloud layer's price per unit of compute; falling rental prices are the neoclouds' rent eroding, and a crash would be the glut that bandwidth went through in 2001. | emerginga model judges: probably stable. The two months of weekly medians on the site dip slightly and recover to almost exactly where they began, a move far inside the noise band, so they were extended as no clear movement. Rents are above their level of a year earlier and show no glut. The caution is that these are list prices from one aggregator, and contracted prices for large buyers could be moving differently. | 35 | grade B | $3.85as of 2026-10-26 | 174 |
| Secondary-market marks between roundsunpublished Marks between rounds show expected-rent repricing before the next primary round confirms it. Not measurable from free sources: secondary marks are behind Caplight, Forge and EquityZen paywalls; the paid stub is off. | no status until publisheda model judges: probably concentrating. No readings are available, so this is general knowledge: through 2025 and into 2026 shares in the leading labs changed hands at prices above their last primary rounds, and each new round was struck higher again. That points to marks rising between rounds. The strongest caution is that the site shows the labs' valuation-to-revenue multiple falling by September 2026, and thin secondary trades can turn quickly. | — | unmeasured | 0 | |
| Revenue per employee in AI-exposed sectorsunpublished The cleanest capture signal for adopters: output per head rising where AI is used. Not measurable yet: no free source pairs sector revenue with employee counts at a usable cadence; BLS industry mappings to exposure are unverified. | no status until publisheda model judges: probably concentrating. There are no readings; from general knowledge, large technology and software firms grew revenue through 2025 and 2026 while holding or trimming headcount, so revenue per head in the most exposed sectors has been rising. That is consistent with gains accruing to firms and their owners. The strongest objection is that outsourcing, price rises and the unwinding of pandemic over-hiring raise the same ratio without any help from AI. | — | unmeasured | 0 | |
| Model-layer token concentration on OpenRouter (HHI, top table) The only public, daily, cross-vendor usage signal. If routed demand concentrates on one or two models, the model layer holds pricing power; if it spreads, the layer is dispersing from the demand side. | emerginga model judges: probably stable. The first three weekly readings on the site are flat at a low level and only the latest jumps, so the judgement extends the flat run and treats the jump as a table artefact or a single popular release until it repeats. The level describes tokens spread across many models. Taken literally over the whole window the rule would read concentrating, which is the strongest case against this call. | 40 | grade B | 0.14 indexas of 2026-11-03(20 obs) | 178 |
| Labs buying from the training-input layer (enter bell) P1's enter bell: when the labs start buying from the layer below them, that layer's economics turn; read with the exit bell, when they buy the vendors outright. | emerginga model judges: probably concentrating. The site's own score points the other way: it has drifted down over the last year and on its rule reads dispersing. The judgement discounts it because the score is a tally of what vendors chose to disclose, which the site describes as a floor, and it falls as old announcements age out even if buying continues. General knowledge, and the site's reading on the largest expert-data vendor, say the labs are buying far more expert data and training environments than a year ago. If that demand has since cooled, the series is right and this call is wrong. | 30 | grade C | 2.00 scoreas of 2026-09-30 | 5 |
| Nvidia five-year credit default swap spreadunpublished The market's price of the chance that the compute layer's counterparties cannot pay. Not measurable as a series from free sources: five-year CDS quotes come from ICE and Bloomberg terminals; the one free reading is a press-quoted figure on the circular-financing ledger, and a direction rule needs quarterly prints. | no status until published | — | grade C | 82.0 bpsas of 2026-07-27 | 1 |
| Valuation to revenue by layerunpublished Where the market expects the rent to sit. Not measurable across layers from free sources: Epoch's post-money and run-rate rows cover model-layer companies only, so a by-layer multiple collapses to lab_valuation_to_run_rate; application- and infrastructure-layer private multiples need Sacra or PitchBook revenue (paid stubs off), and public EV/revenue needs a market-value source the tracker does not ingest. | no status until published | — | unmeasured | 0 | |
| Top-three share of tracked venture dollarsunpublished P1's concentration measure for venture flows. Identical to frontier_capital_concentration on the seed list since Q2 2025 (the top three raisers are the frontier_labs sub-layer), so a separate status would count one fact twice; it publishes when a non-frontier company enters the top three. | no status until published | — | grade B | 91.2%as of 2026-09-30(41 obs) | 117 |
| Median round size by sub-layerunpublished Round size is the price of entry to a sub-layer. Not meaningful on current coverage: one to four seed-list rounds per sub-layer-quarter, so a median restates the venture strip's total and count; it needs a round source with market coverage (aifunding.me is on the read-not-ingested list; PitchBook, Crunchbase and Dealroom are paid). | no status until publisheda model judges: probably concentrating. No usable readings exist, so this is general knowledge: typical AI rounds grew larger at most stages through 2025 and into 2026. Read as the price of entry to a sub-layer, a higher typical round means fewer companies can afford to enter, which is the concentrating direction. The strongest caveat is that very large rounds lift averages far more than medians, and the median within each sub-layer may have moved much less. | — | unmeasured | 0 | |
| Venture flow's lead or lag against sub-layer revenue (quarters)unpublished Capital arriving before revenue is a bet on a sub-layer; capital arriving after it is a response to one. The sign and size of the lag say which. Two things are missing: revenue by sub-layer, which only the labs report and only as press-relayed run-rates, and enough quarters to cross-correlate against. The venture series starts in 2023 and eight quarters of both sides are needed before a lag is more than noise. | no status until published | — | unmeasured | 0 | |
| Stage mix of venture rounds by sub-layerunpublished A sub-layer full of seed rounds is being entered; one full of growth rounds is consolidating. Not measurable from free structured data: Form D has no round or series field, and Epoch's rounds (labs only) name the stage inside a free-text label; a stage mix by sub-layer needs PitchBook, Crunchbase or Dealroom (paid stubs off). | no status until published | — | unmeasured | 0 | |
| Quarters from founding to $100M run-rateunpublished The speed at which new companies reach scale is the application layer's rent in motion. Not measurable: seed/entities.yaml carries no founding dates, and the only run-rate series (Epoch, labs) dates a figure when the press reports it, not when the company crossed a hundred million. | no status until published | — | unmeasured | 0 | |
| Application-company revenue run-rates Whether the application layer keeps a share of the surplus depends on its revenue growing as model prices fall. Until now the tracker had no revenue for any application company. | emerginga model judges: probably concentrating. The readings on the site show the largest application company's run-rate rising by far more than the noise band within a few months of 2026, and they were extended. Several other application companies report similar steep growth. The strongest caveat is that every figure is unaudited and self-selected, and revenue growth says nothing about how much is kept after paying for models. | 35 | grade C | $4.0B⚑disputed: An annualised month reported by the press or the company, not booked or audited revenue, and often sourced to the company itself.run-rateas of 2026-06-05 | 41 |
| Inference-layer revenue run-rates The serving layer sits between the model makers and the applications; whether it keeps margin as token prices fall decides whether it is a durable business or a pass-through. | emerginga model judges: probably concentrating. The two estimates on the site show the largest serving platform's run-rate growing several times over in under a year, far outside the noise band, and the direction was extended. The caveat is strong, though: this is one estimator on one firm, token prices at fixed capability keep falling, and none of these platforms publishes a margin, so revenue passing through the layer may not be profit kept by it. | 35 | grade C | $800M⚑disputed: An annualised month reported by the press or the company, not booked or audited revenue, and often sourced to the company itself.run-rateas of 2026-05-31 | 21 |
| Growth in tokens routed through a public routerunpublished Every other adoption series counts people or accounts. This one counts work: how many tokens are actually being run, and how fast that is growing. Waiting on history: the router's snapshots start in September 2026, and a growth fit on a few days of a trailing-window total would be noise. It publishes once eight weeks of snapshots exist. | no status until publisheda model judges: probably faster than normal. The site has no history yet, so this rests on general knowledge: tokens passing through public routers and model providers multiplied many times over within a single year through 2025 and into 2026, a pace with no close precedent in earlier technologies. The strongest objection is that much of the growth is reasoning and agent loops using more tokens per task, not more users or more work done, and one router is a narrow slice. | — | unmeasured | 0 | |
| METR 50% time horizon The cleanest single measure of how fast the models themselves are improving. Narayanan and Kapoor's argument is that fast improvement here does not, on its own, make firms, jobs or productivity change quickly. | faster than normal | 80 | grade A | 17.4 h(8.5 h–55.1 h)⚑disputed: METR: measurements above 16 hours are unreliable with the current task suiteas of 2026-04-07 | 80 |
| METR 80% time horizon 80% is closer to what a product needs than 50%. If the reliable horizon lags badly, benchmarks flatter products. | faster than normal | 70 | grade A | 3.1 h(1.6 h–6.6 h)as of 2026-04-07 | 26 |
| 50%/80% horizon ratio The ratio is the reliability gap in one number. If it collapses toward 2x while the 80% horizon exceeds eight hours, the normal-technology thesis loses its productisation argument. | consistent with normal stale as of 2026-03-05 | 70 | grade A | 6.34×as of 2026-03-05(2 obs) | 50 |
| US labour productivity, year on year However widely AI is used, it only matters for the economy once output per hour rises. Narayanan and Kapoor expect that to happen at the speed people and firms change, so productivity should stay near its usual trend for years even while the models race ahead. | consistent with normal | 80 | grade A | 2.2%as of 2026-06-30 | 76 |
| US total factor productivity, private nonfarm business TFP is the residual where a general-purpose-technology effect must eventually appear; it strips out capital deepening, so it is the cleanest macro test of adaptation. | consistent with normal | 75 | grade A | 0.8%as of 2025-12-31(2 obs) | 9 |
| Developer productivity uplift (METR RCTs) The most carefully controlled test of whether better models make people faster at real work: randomised trials in which experienced software developers do their own tasks with and without AI. If benchmark gains turned quickly into real gains, it would show here first. | consistent with normal | 70 | grade A | -4.0%⚑disputed: METR itself calls this 'very weak evidence' because developers refusing the no-AI arm biases the sample.as of 2026-02-24 | 2 |
| Enterprise pilots with measurable P&L impact The product-to-adoption valve in one number. If most pilots never pay off, adoption stays broad but shallow. | consistent with normal | 55 | grade C | 5.0%⚑disputed: Denominator and sample disputed; Menlo reports ~47% of pilots reach production on a different definition. No MIT-hosted copy of the PDF exists; this is the mirror Menlo cites.as of 2025-06-30 | 2 |
| Share of work hours assisted by generative AI Saying you use AI is not the same as working with it. Adoption surveys count anyone who has tried it; this counts the share of working hours actually done with it, which is what any productivity gain has to be built from. | emerginga model judges: probably within the normal pace. The latest reading on the site is a single-digit share of work hours, inside the normal band, and it has risen only gradually over the quarters shown, so the readings were extended. Use is broad but shallow. The strongest evidence the other way is that the share of workers using these tools weekly has climbed quickly, which could pull hours up faster than the level suggests. | 80 | grade B | 5.3%as of 2026-09-30 | 8 |
| US firms using AI (Census BTOS) The broadest official count of whether businesses use AI at all, from a Census Bureau survey sent to 1.2 million firms every two weeks. Every firm counts once, so small businesses dominate; weighted by employees, use is higher. | consistent with normal | 70 | grade A | 24.6%(24.1%–25.1%)as of 2026-10-04 | 25 |
| US businesses paying for AI (Ramp AI Index) The only monthly spend-based adoption series. Paying is a stronger commitment than 'using', and the sample skews to tech-forward firms, so it runs ahead of the Census rate and shows the ceiling first. | emerginga model judges: probably within the normal pace. The latest reading sits between the bands, above the normal ceiling and short of the fast floor, and the monthly gains on the site shrink steadily through 2026 to well under a point. Because the trend is flattening before reaching the fast band, the readings were extended as the pace of an ordinary technology in a tech-forward sample. The case against is the level itself, which already exceeds what the band calls normal for this age. | 60 | grade B | 55.9%as of 2026-09-30 | 46 |
| Agent-workdays per human workday in frontier research (self-reported) The return arrow made numeric. If deployment inside the lab is feeding invention at more than one agent-day per human-day, the loop Narayanan and Kapoor draw but do not instrument is running; whether it substitutes for external bottlenecks is the thesis question. | emerginga model judges: probably faster than normal. The one reading on the site is just above the fast threshold, with agent effort exceeding human effort several times over inside one lab's research organisation, so it was read as fast. The strongest objection is that the figure is measured and published by the lab it flatters, counts agent time and not delivered research, and sits beside another lab's statement of a much smaller overall speed-up and independent studies that find modest gains. | 30 | grade D | 3.10×⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-08-15 | 1 |
| Human interventions on 4–8 hour agent tasks (self-reported) The reliability side of the return arrow. Agent-workdays can rise while humans still steer every long task; the loop only substitutes for people when this rate falls. | emerginga model judges: probably within the normal pace. The single reading says more than one in two successful long agent tasks still needed a person to step in, which sits in the normal band, and it was extended as the present state. Humans still steer most long tasks. The caveat is that the figure is self-reported, rounded and from the earlier part of 2026, and autonomy on long tasks has been improving quickly. | 30 | grade D | 50.0%⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-07-31 | 1 |
| Safety brakes inside the labs (events, trailing year) The paper's safety speed limits are argued from regulated domains; the tracker also watches for them on the invention stage, where a brake inside a lab slows the whole chain before any regulator acts. | emerginga model judges: probably within the normal pace. The site records three announced brakes within the trailing year, which meets the band's test that limits are binding at least occasionally, so it was read as normal. The strongest counterpoint is that all three come from one lab describing itself, and that the paused training run was restarted within a month, so the brakes may be shallower than the count implies. | 40 | grade D | 3as of 2026-11-04(3 obs) | 3 |
| Frontier training compute growth (doubling time) The invention stage's own speed limit. The brief's reading is 4-5x a year; a slowdown here is the first thing the data-and-compute constraint would show. | emerginga model judges: probably faster than normal. The fitted reading on the site has frontier training compute multiplying a little faster than the fast threshold, so it was read as fast, in line with the long-run trend since 2010. It is close to the edge, though, and the newest closed models are missing or inferred with wide error. The strongest alternative is that growth has eased as single training runs hit power and chip limits. | 60 | grade B | 170 daysas of 2026-09-03(67 obs) | 67 |
| Training power draw growth (doubling time) Power is the physical constraint that compute growth eventually hits; the brief's reading is 2.2x a year. | emerginga model judges: probably within the normal pace. The fitted reading on the site has training power draw taking well over a year to grow by the same multiple, inside the normal band, and it was extended. The fit is very loose, however, and the set of notable models mixes small and large systems. General knowledge that the largest training sites have been growing faster than this is the strongest point the other way. | 50 | grade C | 419 daysas of 2026-10-15(216 obs) | 216 |
| Hardware price-performance growth (doubling time) The floor under falling inference prices; the brief's reading is +49% a year. | consistent with normal | 40 | grade B | 1331 daysas of 2026-08-02(42 obs) | 42 |
| Inference price at fixed capability (halving time) Price per unit of capability on a benchmark axis, the lab layer's capture proxy; the brief cites about 40x a year at this threshold. | emerginga model judges: probably faster than normal. The fitted reading on the site sits between the bands but close to the fast edge, and the site notes that its method probably makes the fall look slower than it was, so it was read as fast. General knowledge of very steep price declines at a fixed capability level supports this. The caveat is that the fit has few points on one benchmark threshold and uses list prices. | 45 | grade A | 121 daysas of 2026-04-03(19 obs) | 19 |
| Open-weights lag behind the closed frontier (months) How fast frontier capability becomes something anyone can download and run is the methods stock leaking into products outside the labs' control, and the capture lens's clearest sign of model-layer commoditisation. | emerginga model judges: probably within the normal pace. The readings on the site bounce between roughly a third of a year and two thirds of a year, and the latest is nearer the normal edge than the fast one, so it was read as normal. Open weights trail by something like a model generation. The strongest counter is that each new open release cuts the lag sharply, and one strong release would move the reading to the middle of the range or below. | 45 | grade A | 7.7 monthsas of 2026-10-26(3 obs) | 35 |
| Public text data exhaustion year (Epoch projection) Bottleneck 88 (data exhaustion) as a dated estimate: when the readily available data is spent, continued scaling becomes a business decision about synthetic and licensed data. | emerginga model judges: probably within the normal pace. The site's only reading is a projection made in 2024 that lands exactly on the normal edge, and it was taken as it stands. General knowledge cuts both ways: licensed, synthetic and multimodal data stretch the stock, while the labs' heavy buying of expert data suggests public text is already binding. The interval around the projection is wide enough to cover the fast case. | 40 | grade C | 2028as of 2024-06-06 | 1 |
| ARC-AGI-2 best score The one serious proxy for out-of-distribution generalisation, the brief's replacement for a 'world-model transfer' dimension. | emerginga model judges: probably faster than normal. The running best score on the site climbed from well under the normal ceiling late in 2025 to above the grand-prize bar by 2026 and has kept rising, so the readings sit firmly in the fast band. The strongest caveat is that these are outside submissions with no cap on cost per task, on a single benchmark, so the score overstates what is practical. | 55 | grade B | 95.0%as of 2026-09-03 | 17 |
| Continual-learning ladder (highest production rung) The crux of the timelines argument. If systems learn on the job, deployment feeds invention automatically and the return arrow stops depending on human-data vendors; Narayanan and Kapoor's speed limits assume it stays unsolved. | emerginga model judges: probably within the normal pace. The highest rung with a shipped system on the site's record is the top of the normal band, per-customer adapters, and nothing on record reaches organisation knowledge held in weights or beyond, so it was read as normal. The rung did rise within the last year. The main risk to the call is that frontier labs may ship more than they describe, and every row is a vendor's own claim. | 45 | grade D | 3.00 rungas of 2026-05-27(2 obs) | 2 |
| FDA-authorised AI-enabled devices, growth The clearest regulated-domain speed limit with an official monthly series; the paper predicts slow diffusion in high-consequence tasks and this is the count of what gets through. | consistent with normal | 70 | grade A | 26.7%as of 2026-06-30(2 obs) | 102 |
| US state AI bills introduced (legislative year) The institutional response the paper expects to lag deployment; volume says attention, enactments say binding rules. | emerginga model judges: probably within the normal pace. The one count on the site is far above the threshold for the normal band, meaning legislatures are reacting at their usual volume, and it was extended into 2026. The strongest point against is the federal push to pre-empt state AI laws, which could thin the flow, and the fact that introduced bills are cheap signals compared with laws enacted. | 55 | grade B | 1035as of 2025-12-31 | 2 |
| Active RL-environment vendors (directory count) The return arrow's supply side: environments are how deployment feedback becomes training signal today, and the vendor count is the only public series on that market. | emerginga model judges: probably within the normal pace. Both directory counts on the site sit well inside the normal band and the count did not grow between them, so it was read as a specialist market. The directory is one person's list and a floor, and general knowledge suggests many more small suppliers exist than it records, which is the strongest reason the true figure could be in the middle range. | 35 | grade B | 23as of 2026-10-05 | 2 |
| Ord half-life misfit (observed 50/80 ratio vs constant hazard)unpublished Reports Ord's reframing alongside the horizons, as the brief asks: a misfit near 1 means success decays at a constant rate with task length; well above 1 means reliability collapses faster at long horizons, which is the capability-reliability gap in one number. A presentation of the 50%/80% horizon ratio, not evidence about it: this is that ratio divided by 3.11, from the same two observations, so a separate status would score one number twice and its bands imply thresholds the ratio's own bands contradict. The misfit is shown on the ratio's page and in its saved analysis. | no status until published stale as of 2026-03-05 | 60 | grade A | 2.04×as of 2026-03-05(2 obs) | 50 |
| FDA-authorised devices built on a large language model The first LLM device through review is the moment the regulated-domain speed limit meets the technology the paper is about; the count after it is the throughput of that limit. | consistent with normal | 60 | grade A | 1as of 2025-12-23(3 obs) | 3 |
| Reliability-science framework adoptionunpublished The 70% problem (70% of tasks always versus every task 70% of the time) only becomes trackable once labs report reliability, not accuracy. Not measurable as a number: adoption is a yes/no per system card, recorded as evidence rows rather than a series. | no status until publisheda model judges: probably within the normal pace. There is no series, so this is general knowledge: system cards from the leading labs report capability and safety tests but, as far as is known, none reports reliability along the full set of dimensions the framework proposes. Slow uptake of a measurement standard is the ordinary pattern. The caveat is that cards have been growing richer, and partial reporting on consistency or calibration may already exist under other names. | — | unmeasured | 0 | |
| Agent calibrationunpublished Calibration is self-monitoring, the reliability dimension that lets a human trust when not to check. HAL now scores calibration for the agents it runs (the Cal column of its reliability table, from agents rating their own confidence after each task), but only for its current table, with no dates, so there is no series to read a direction from; it publishes once HAL dates the scores or keeps a history. | no status until publisheda model judges: probably within the normal pace. No dated series exists, so this rests on general knowledge: agents remain noticeably overconfident about whether they completed a task, and calibration has improved more slowly than raw capability. That fits a technology that still needs human checking. The strongest caveat is that benchmark calibration scores are few and recent, and newer models may have closed more of the gap than is publicly visible. | — | unmeasured | 0 | |
| Agent operational safety incidentsunpublished Operational safety is the fourth reliability dimension and the one that connects to bottleneck 65. Not measurable yet: incidents are anecdotal; no incident registry with a denominator exists. | no status until publisheda model judges: probably within the normal pace. No registry exists, so this is general knowledge: destructive mistakes by deployed agents keep surfacing as anecdotes, and firms respond by restricting permissions and keeping people in the approval loop. That is the ordinary pattern of a technology being fitted with guard rails, not one running unsupervised. The weakness of the call is that without a denominator nobody knows whether the rate per task is falling quickly. | — | unmeasured | 0 | |
| Deployment data share of trainingunpublished The return arrow's throughput in one number. Not measurable: no lab discloses the provenance mix of its training data; recorded only if a system card states it. | no status until published | — | unmeasured | 0 | |
| Employment in AI-exposed occupations, year on year The paper predicts no broad AI-attributable employment decline for years; this is the highest-frequency payroll test of that claim. | consistent with normal | 80 | grade B | -0.5%as of 2026-09-30 | 103 |
| Entry-level employment shortfall in AI-exposed occupations If AI substitutes for the codified knowledge new graduates bring, the first sign is hiring, not separations. This is that sign. | emerginga model judges: probably faster than normal. The two readings on the site sit between the bands, nearer the fast floor than the normal ceiling, and the gap widened by several points in a year, so the direction was extended as fast. The strongest objections are serious: part of the timing predates the chat assistants, hiring freezes and interest rates hit the same occupations, and other labour-market studies find little broad disruption so far. | 70 | grade B | 19.0%as of 2026-06-30 | 4 |
| Exposed vs unexposed occupation employment gap (Revelio) An independent, differently sourced read on the same question as the Canaries payroll series. | emerging stale as of 2026-07-31 a model judges: probably within the normal pace. The single reading on the site is only just beyond the normal band's edge, a relative gap of the size occupation groups drift apart in any period, so it was read as normal. The site notes the gap widened noticeably within one month of releases, which is the strongest sign it may be heading for the fast band, and the entry-level reading on the site points that way too. | 65 | grade B | -6.0%as of 2026-07-31 | 1 |
| California unemployment claims from AI-exposed occupations Administrative claims data, not a survey and not a profile scrape; the closest thing to a real-time separations signal. | consistent with normal | 70 | grade A | -4.4%as of 2026-09-30(4 obs) | 119 |
| Recent-graduate unemployment rate (NY Fed) The macro counterpart of the entry-level canaries signal; if AI is closing the door on new graduates, it shows here first among official series. | consistent with normal | 75 | grade A | 5.6%as of 2026-06-30 | 2 |
| Cross-tracker concordance (labour) The thesis's falsification rule requires three trackers breaking concurrently, not one. This is that count, made explicit. | consistent with normal | 70 | grade B | 0as of 2026-09-30(7 obs) | 7 |
| Augmentation vs automation share (Anthropic Economic Index) The paper's bet is that AI diffuses as a tool people use, not a replacement; this is the only large-scale usage classification that speaks to it directly. | emerginga model judges: probably within the normal pace. Both monthly readings on the site sit just above the line where augmentation is the majority of classified use, inside the normal band, and they were extended. The margin is thin. The strongest evidence against is that the series covers one provider's chat products only, and programmatic use, where automation concentrates, skews the other way. | 60 | grade B | 51.4%as of 2026-05-31 | 11 |
| US labour share of income, year on year If AI shifts surplus from wages to capital, this is where it shows in official data. The Anthropic Institute's September 2026 scenarios put the labour share at 60% today and 45% by 2030 under extreme change, and unchanged under modest change. | emerging | 65 | grade A | -3.4%as of 2026-06-30(2 obs) | 38 |
| Customer-support productivity uplift from a generative-AI assistant The best-identified productivity effect on record for a deployed assistant; the normal-technology story predicts modest average gains concentrated in novices, which is exactly what it found. | consistent with normal | 75 | grade B | 14.0%as of 2023-04-30 | 2 |
| Executives reporting no AI impact on their own firm Adoption headlines are broad, but firm-level impact is what adaptation means. Nine in ten firms seeing nothing yet is the productivity J-curve in survey form. | consistent with normal | 70 | grade B | 90.0%as of 2026-02-28 | 2 |
| Frontier models completing expert legal tasks end to end (Harvey LAB) The productisation valve for professional work. A model that passes bar-style questions but completes 7% of real matters is a tool, not a lawyer. | emerginga model judges: probably within the normal pace. The one reading on the site has the best model completing only a small share of expert legal tasks end to end under a strict standard, far inside the normal band, and it was extended. The caveat is that the all-pass standard is deliberately harsh and the vendor scores its own benchmark; partial completions already have practical value. | 40 | grade D | 7.1%as of 2026-05-26 | 2 |
| Agent reliability across repeated runs (tau-bench pass^k) The 70% problem in one number - solving a task once is not solving it reliably. Deployment needs pass^k, not pass^1, and the gap between them is where the normal-technology speed limit lives. | emerginga model judges: probably within the normal pace. The reading on the site has the leading agent passing a little over one task in two on every one of several repeated runs, clearly inside the normal band, and it was extended. Reliability still falls away with repetition. The strongest counter is that the result depends heavily on the domain, with some narrow domains already near the fast threshold. | 55 | grade A | 56.2%as of 2026-05-18 | 110 |
| Progress against the AI 2027 quantitative predictions The scenario the fast lens is calibrated on grades its own forecasts. If progress runs at three quarters of its pace, the timeline stretches by years. | emerging | 50 | grade B | 70.0%as of 2026-07-31(2 obs) | 3 |
| Work-related share of ChatGPT consumer messages The largest usage dataset there is says most use is personal. Diffusion into work is the slower, deeper channel the paper expects. | emerginga model judges: probably within the normal pace. The one reading on the site has about a quarter of consumer messages related to work, well inside the normal band, and the site says the share was falling as personal use grew faster. The reading dates from 2025 and was extended. The strongest caveat is that business and programmatic traffic, where work concentrates, is excluded by construction. | 65 | grade B | 27.0%as of 2025-06-30 | 2 |
| Usage as a share of theoretical exposure The paper's central gap - what the technology could touch versus what people actually use it for weekly. Exposure is a ceiling; usage is the diffusion curve under it. | consistent with normal | 60 | grade B | 0.44×as of 2026-09-30(2 obs) | 10 |
| Workers highly exposed to AI with low adaptive capacity Displacement is a distribution question. The workers who cannot adapt are where a fast scenario turns from a productivity story into a welfare one. | emerginga model judges: probably within the normal pace. The single estimate on the site is a little above the normal ceiling and far below the fast floor, so it was read as nearer normal. It is one modelled estimate that has already been revised upward once, which is the strongest reason for doubt, and a further revision of the same size would put it squarely in the middle range. | 45 | grade C | 6,100,000 workersas of 2026-01-31 | 1 |
| Occupational-mix dissimilarity since ChatGPT (Yale Budget Lab) If AI were reshaping work, the mix of occupations would move faster than in prior technology transitions. The Budget Lab finds it moving at about the historical pace. | consistent with normal | 65 | grade B | 5.56 ppas of 2026-09-01 | 183 |
| Chatbot effect on earnings and hours in exposed occupations (Denmark) The cleanest labour-market null on record, from administrative data covering the whole population. A normal technology shows up in earnings slowly; this is the slow showing. | consistent with normal | 80 | grade B | 0.0%as of 2024-12-31 | 3 |
| Freelance writing earnings after ChatGPT (Upwork) The most substitutable work on the most measurable platform. If displacement shows anywhere first, it is here, and it shows as single digits. | consistent with normal | 70 | grade B | -5.2%as of 2023-12-01 | 3 |
| Wage premium for AI skills (PwC AI Jobs Barometer) The price of scarce complementary skill. A large, stable premium says AI diffuses through people who can use it; a premium that doubles or collapses says the labour market is repricing at speed. | emerginga model judges: probably within the normal pace. The reading on the site falls inside the normal band, a large premium of the kind earlier scarce technical skills carried, and it rose only modestly from the year before, so it was extended. The caveat is that it comes from job advertisements, not wages paid, and that for a narrow group of top researchers pay has risen far more than any such average shows. | 50 | grade C | 62.0%⚑disputed: PwC compares pay in job postings that ask for AI skills with postings that do not: a posting premium, not a measured effect of AI on wages.as of 2025-12-31 | 2 |
| Expert-data market run-rate (Mercor) The RL-environment and expert-data market is the input the fast scenario needs to keep scaling. A single vendor doubling to two billion gross in six months is that input arriving. | emerginga model judges: probably faster than normal. The reading on the site puts the largest expert-data vendor at the fast threshold, a scale that data-labelling leaders took about a decade to approach, and it was read as fast. The strongest caveats are that the figure is a press-sourced gross run-rate, most of which passes straight through to the experts, and that lab demand could fall quickly if synthetic data substitutes. | 70 | grade C | $2.0B⚑disputed: An annualised month reported by the press or the company, not booked or audited revenue, and often sourced to the company itself.run-rategrossas of 2026-06-30 | 6 |
| Internal-to-public deployment gap at the frontier AI 2027 turns on labs keeping their best models internal. A gap that grows is the scenario's signature; a gap of months is ordinary product testing. | consistent with normal | 45 | grade B | 4 monthsas of 2026-06-30 | 2 |
| Waymo paid rides per week The paper's clearest example of a regulated, physical-world deployment that diffuses city by city. The number is the speed limit made visible. | consistent with normal stale as of 2026-03-27 | 65 | grade C | 500,000 rides per weekas of 2026-03-27 | 2 |
| US adults who have used generative AI The extensive margin. The paper expects breadth to move fast and depth to lag; this is the breadth half of that claim, read against hours assisted. | emerginga model judges: probably within the normal pace. The latest reading on the site is a clear majority of adults but still well inside the normal band, and the quarterly path would not reach the fast floor for some years, so the readings were extended. The strongest point the other way is that the most recent quarter's gain was the largest in the series, and that even this pace is quicker than the internet managed. | 70 | grade B | 60.8%as of 2026-09-30 | 18 |
| Occupations with a quarter of their tasks using Claude Depth across the occupational map rather than across people; whether AI use is reaching most of the work in most jobs or a narrow band of them. | consistent with normal | 55 | grade B | 49.0%as of 2026-02-12 | 1 |
| US workers using generative AI for work weeklyunpublished The work margin between ever-used and hours assisted. Folded into two published indicators: the same FRED series is shown on bbd_work_hours_assisted and is the numerator of exposure_vs_usage_gap, so a separate status would count one survey twice in the early-adoption bucket. | no status until published | — | grade B | 34.8%as of 2026-09-30 | 10 |
| US workers using generative AI for work every day Daily use is where depth shows up first. | emerginga model judges: probably within the normal pace. The one reading on the site, from late 2025, is a little under the normal ceiling, so it was read as normal. It is close to the edge, and the same survey's weekly-use series has kept climbing since, so daily use has probably risen into the gap between the bands by now; that is the strongest reason this could be wrong by a step. | 30 | grade B | 12.0%as of 2025-11-30 | 1 |
| Vendors claiming shipped continual learning The breadth behind the ladder's top rung. One vendor shipping per-customer learning is an anecdote; six is a product category. | emerginga model judges: probably within the normal pace. The site counts two builders claiming shipped continual learning in the trailing year, inside the normal band, and no frontier lab has branded a release that way, so it was read as normal. The strongest caveat is that labs may ship such methods without saying so, which this count would miss, and that every row is a vendor's own description. | 35 | grade D | 2as of 2026-11-04(2 obs) | 3 |
| Per-customer adapters per productunpublished The scale at which systems learn from each customer separately. Not measurable from public sources: Trajectory describes an adapter per customer refreshed hourly without a count, and Decagon states a share of traffic on self-trained models, not adapters. | no status until published | — | unmeasured | 0 | |
| Retention after continual updatesunpublished Continual learning without forgetting is the hard part; retention is how to tell. Not measurable yet: no production system reports retention after updates; the research rows on the ladder report their own tasks, and no shared forgetting benchmark scores deployed systems. | no status until published | — | unmeasured | 0 | |
| Learning-curve slope in deploymentunpublished The six-month-employee test is a slope, not a level. Not measurable: the only curve on record (cl.engram_harvey.closed_book_accuracy.pt) is a research result across training steps, not lift over deployment time. | no status until published | — | unmeasured | 0 | |
| Fastest production refresh of a continually updated model (hours) How quickly deployment feeds back into the model is the return arrow's speed; a refresh measured in minutes is continual learning in all but name. | emerginga model judges: probably within the normal pace. The single reading on the site is exactly at the edge of the normal band, an hourly retrain-and-redeploy pipeline at one vendor, and it was read as normal. It describes a retrained adapter, not a model learning between sessions. The risk is that the reading sits on the boundary and one vendor describing a quicker path would move it into the gap. | 30 | grade D | 1.00 hoursas of 2026-05-27 | 1 |
| Enterprise sessions allowed to train vendor modelsunpublished Continual learning needs deployment data; enterprise contracts that forbid training on sessions close the arrow for most paid use. Not measurable: the policy is set contract by contract, and no vendor publishes the share of enterprise seats whose sessions may train its models; default opt-outs are recorded as evidence when a vendor changes them. | no status until publisheda model judges: probably within the normal pace. There are no readings; from general knowledge, the standard business terms of the main model vendors exclude customer sessions from training unless the customer opts in, and large buyers insist on it. So only a small share of enterprise seats feed the vendors' models, which is the slow, contract-bound pattern of an ordinary technology. The counter is that some consumer and individual plans have moved toward training by default. | — | unmeasured | 0 | |
| A benchmark for learning on the jobunpublished Without a benchmark the top rung cannot be observed, so the ladder tops out by construction. A yes/no, not a series: Epoch's CL-bench rows score learning from material supplied within one task (domain reasoning, rule application, procedure, empirical discovery), not retention across sessions, so they do not meet the L6 six-month-employee test. | no status until published | — | unmeasured | 0 | |
| Enterprise AI pilots reaching production (surveys)unpublished The step between a pilot and reorganised work is where diffusion stalls; the P&L measure beside it (MIT NANDA) asks a harder question, so the two are read together. The surveys measure different things (production, abandonment, scaling). Two survey families now measure the share of projects that reach production (IDC for Lenovo, two editions; Gartner, one, for late 2023), and the indicator publishes once both have two years. | no status until publisheda model judges: probably within the normal pace. The one reading on the site has a little under one pilot in two reaching production, and the surveys alongside report a third or more abandoned. That is the familiar pattern of enterprise technology, where many trials stall on integration and trust, so it was read as normal. The strongest counter is that the surveys define their terms differently, and some suggest AI purchases convert faster than ordinary software did. | — | grade B | 47.0%as of 2025-11-25 | 12 |