Tracking the Impact of AI on the Labor Market

register entry yale-budget-lab-ai-labor-market-aio-20261029 OBS-9c44ea25d2ca · file as first run sha256 12b0399e6023def5…

• T1-1 definition: 'The AI Labor Market Tracker by The Budget Lab at Yale finds no clear evidence of large-scale or economy-wide labor market disruption from artificial intelligence.'
• T1-2 minimal economy-wide impact: 'Data updated through September 2026 shows that employment, unemployment rates, and wage trends for AI-exposed workers largely mirror historical patterns and pre-AI trajectories.'
• T1-3 churn: 'Changes in the mix of occupations across industries (occupational churn) remain flat and show no acceleration or divergence since the late 2022 launch of ChatGPT.'
• T1-4 exposed against unexposed: 'Synthetic differences-in-differences comparisons between AI-exposed and unexposed workers reveal no strong, consistent statistical footprint of AI-driven displacement, though slight volatility in unemployment growth for high-exposure groups continues to be monitored.'
• T1-5 capital against labor: 'Companies cutting jobs or restructuring are generally doing so for broader macro or capital-investment reasons rather than direct workforce replacement by AI agents.'
• T1-6 sections, churn: 'Occupational Churn: Tracks how fast job roles and industry composition are shifting over time.'
• T1-7 sections, AI metrics: 'AI Metrics: Evaluates various index measurements of AI task exposure across standard occupational classifications.'
• T1-8 sections, comparisons: 'Comparisons: Analyzes employment, wage, and unemployment trajectories of workers in high- versus low-AI exposure categories.'
• T1-9 closer: 'Would you like to know more about specific high-exposure occupations (such as tech or administrative roles) or how different AI exposure metrics compare?'

• B1 The Budget Lab — Evaluating the Impact of AI on the Labor Market: Current State of Affairs | The Budget Lab card; B1.html sha256 065e453aacdb846c, HTTP 200 at 2026-10-29T22:30:32Z; the census recorded 403 for this URL
• B2 The Budget Lab — Tracking the Impact of AI on the Labor Market | The Budget Lab card; B2.html sha256 7efbd4d59a4fa632, HTTP 200 at 2026-10-29T22:30:34Z; byte-identical to the census copy yale-budget-lab-ai-tracker--tracker
• B3 The Budget Lab — Labor Market AI Exposure: What Do We Know? | The Budget Lab card; B3.html sha256 c17cecc68b86f656, HTTP 403 then 200 (same agent)
• B4 The Budget Lab — AI Is Probably Not (Yet) the Reason for Labor Market Weakening | The Budget Lab card; B4.html sha256 e0811bd485fd98cd, HTTP 200
• B5 YouTube·Quest Means Business — Yale Budget Lab Executive Director Martha Gimbel on why AI hasn't disrupted ... card; unreachable: no URL found; title only
• B6 The Budget Lab — Tracking the Impact of AI on the Labor Market (printable PDF of the tracker, Version 1.0, 15 September 2026) own_page; copied from the census's saved copy, sha256 9bcf775f43ed5756
• B7 The Budget Lab — Evaluating the Impact of AI on the Labor Market: September CPS Update (unread) own_page; unread: 403 to the census, not refetched
• B8 Wikipedia — Tracking the Impact of AI on the Labor Market (no article) wikipedia; no article: raw and rendered 404; one article cites the first report

E_A, the archive's entity: the same tracker, in the Labor Vector census record and the archive's deposits: an interactive tool whose evidence is an observed employment outcome and an exposure estimate, at the intersection of gains and losses, recording no supplier, director or recipient of gains; its finding 'NOT a finding that AI has no labor market effect'; its usage measures inputs from providers its saved text does not name; able to settle churn, exposure among the unemployed and usage against employment, and unable to settle who supplies, directs or gains; a member of the AI-economics research #716 describes as measuring which jobs AI eliminates and what tasks it can perform, beside which #716 sets a fourth channel, value extracted where labor remains. LV-01 LV-04 LV-05 LV-06 LV-09 LV-17 LV-18 LV-24 LV-25 LV-31 O716-01 O716-02 O716-03; defining papers #716

E_F, the field's entity (type C — a conventional reading holds it (public entity: a labor-market tracker and the reports behind it)): The Budget Lab at Yale's interactive tool 'Tracking the Impact of AI on the Labor Market' (the AI Labor Market Tracker), published 16 July 2026, Version 1.0 of 15 September 2026, page of 14 October 2026, holding the Lab's 'ongoing analysis of potential AI impacts on the labor market' in three sections (occupational churn, AI metrics, comparisons of AI-exposed and unexposed workers) on monthly CPS microdata; continuing the report of 1 October 2025 by Gimbel, Kinder, Kendall and Lee; its exposure inputs OpenAI's (Eloundou et al.) and six other metrics, its usage input Anthropic's Claude data; its finding that AI's labor-market footprint is not yet clear, with the stated limit that the findings may change.

ρ = extends: The archive reads the tracker as its operator presents it and adds, on the field's own object: the census's typing of the evidence and its intersection, what the evidence can and cannot settle, the direction it does not record, the two negative tags; and #716's channel, value extracted where labor remains, which an employment measure of the class does not register. LV-05 LV-06 LV-09 LV-17 LV-18 LV-24 LV-25 LV-29 LV-31 O716-01 O716-02 O716-03. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates: #716 does not say the Lab's measure is wrong, only that its class measures displacement and capability and not extraction where labor remains. Where an archive claim stands against a field claim, both are composed under P_coh within the one ontology; none does here.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The Labor Vector census record describes the tracker from the Lab's own pages, each field with its basis ('operator'), and answers for it the census's question (who supplies, directs and gains, and what kind of evidence the resource brings), recording its evidence as an observed employment outcome and an exposure estimate and stating what that evidence can and cannot settle; its negative tags keep the tracker's 'not yet' from a finding of no effect (LV-24) and the tracker from a usage dataset (LV-25). The capture record checks the composition against those pages (CAP-1-01, CAP-1-02). No deposit names the tracker or the Lab. One deposit bears on its class: #716 states what AI-economics research measures (labor displacement, capability projection: 'which jobs AI eliminates, what tasks it can perform') and adds, on that object, a channel it does not register, value extracted where labor remains ('This is not displacement; it is extraction without attribution'), as 'a fourth' beside the three it names. That is an addition of a sense and a limit to the field's own object, not a contest of its ontology; no defining paper supplies an ontology the field lacks: ρ = extends, the equal-terms case (§0.9, §1.2a). Labor Vector's own reading of where the tracker meets its instruments is 'none stated' (LV-19).

#716 §2.2 'Economic Framing of AI Composition': These frameworks measure which jobs AI eliminates, what tasks it can perform, and what consumer surplus it generates. PER identifies a distinct channel: even when human labor *remains* (the author's work is used), the economic value tied to provenance — reputation, traffic, citation credit, contractual rights — is extracted by the system.

#716 §5.1 ¶1: Existing AI-economics research addresses three channels: labor displacement, capability projection, and welfare estimation. PER addresses a fourth

#LV census.json rows[slug == 'yale-budget-lab-ai-tracker'].negative_tags[0]: NOT a finding that AI has no labor market effect: it states the analysis does not yet clearly indicate one

#LV census.json rows[slug == 'yale-budget-lab-ai-tracker'].does_not_answer[0]: who supplies, directs or gains from labor performed with AI

#LV census.json rows[slug == 'yale-budget-lab-ai-tracker'].lv_intersection: none stated

#LV census.json schema_note: lv_intersection (where it meets DSL, task-origin and user labor drag: the archive's reading, kept apart from the resource's own terms)

#capture-registry yale-budget-lab-ai-labor-market-aio-20261029, reading: The innocent reading first. The Overview names the tracker and its conclusion, gives its methods and its sections, from the Lab's own research pages.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 14 (positions [None, None, None, None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

The Budget Lab at Yale describes the AI Labor Market Tracker as an interactive tool containing its "ongoing analysis of potential AI impacts on the labor market", in three sections: occupational churn, AI metrics, and comparisons of AI-exposed and unexposed workers (tracker PDF). It was published on 16 July 2026, reached Version 1.0 on 15 September 2026 with downloadable data, and its page is dated 14 October 2026 under Yale University's copyright (tracker PDF; tracker page). Its takeaways: the occupational mix is not yet changing in ways clearly aligned with AI; AI usage shows no connection to employment or unemployment; and a synthetic differences-in-differences analysis does not yet clearly indicate a labor-market footprint (tracker page). The September update on August 2026 CPS microdata finds no clear evidence of AI-associated disruption, and the Lab says the findings may change as new data arrive and plans regular updates (tracker PDF; tracker page). The analysis continues the report of 1 October 2025 by Gimbel, Kinder, Kendall and Lee, which asked whether change since ChatGPT differed from past technological change, and its documentation includes Gimbel, Kendall and Kulsakdinun's exposure review of February 2026 (first report; tracker PDF; exposure review). That report found no discernible economy-wide disruption in 33 months, a mix changing somewhat faster than before but from before AI, limited effects by industry, and called the anxiety largely speculative, while stating that its analysis is not predictive and that it is too soon to tell (first report). Churn is a Duncan-and-Duncan dissimilarity index on monthly CPS data that records change without taking a stance on its cause (first report). Its exposure data are OpenAI's, a theoretical estimate, and its usage data Anthropic's Claude data; neither showed a relation to employment, with exposure shares stable and no growth in the exposure of the unemployed (first report). It warns that Claude usage is not representative and asks for comprehensive usage data from all leading AI companies (first report). Recent graduates' dissimilarity rose slightly faster, possibly consistent with Brynjolfsson et al., to be read with caution (first report). The exposure review finds seven exposure metrics, Eloundou et al.'s and a Microsoft Copilot-based score among them, broadly in agreement but less so for highly exposed occupations, and says exposure marks where AI could have an impact, not jobs it will automate away (exposure review). Ryan Nunn's note of May 2026 applies synthetic differences-in-differences and finds employment and hourly-wage effects close to zero, adds that churn does not formally identify AI's effects, and expects AI eventually to leave its mark (Nunn's note). It sets this in a cooled market, unemployment at 4.3% in March 2026 and payroll growth slowed in principal part by lower immigration, and finds the AI explanation unsupported to date (Nunn's note). Wikipedia has no article on the tracker, though its generative-AI article cites the first report, and a Quest Means Business video features Gimbel on why AI hasn't disrupted … (Wikipedia; video card).

Tracking the Impact of AI on the Labor Market, also named the AI Labor Market Tracker, is an interactive tracker with regular updates whose evidence is an observed employment outcome and an exposure estimate (Labor Vector record); by its operator's account it "contains The Budget Lab's ongoing analysis of potential AI impacts on the labor market" (tracker PDF). Its operator is The Budget Lab at Yale, its pages under Yale University's copyright (tracker page; Labor Vector record), Martha Gimbel the Lab's executive director (first report); the census places it at the distribution of gains and losses and asks of it "Who supplies the labor, who directs it, and who receives its gains — and what kind of evidence does this resource bring to that question?" (Labor Vector record). The analysis began as "Evaluating the Impact of AI on the Labor Market: Current State of Affairs" by Gimbel, Molly Kinder of Brookings Metro, Joshua Kendall and Maddie Lee, dated 1 October 2025 (first report), a title the census carries as one of the tracker's names (Labor Vector record); it asked whether labor-market change in the 33 months since ChatGPT's release differed from past technological change and whether there were economy-wide employment effects, and it planned regular updates (first report). The tracker was published on 16 July 2026 and updated to Version 1.0 on 15 September 2026, its data downloadable (tracker PDF), its page dated 14 October 2026 with the same plan of regular updates (tracker page); the census records it as active, founded on 16 July 2026, its data access 'Download All Data' and its licence unstated beyond Yale's copyright (Labor Vector record). It refers its churn and the exposure of the unemployed to Gimbel et al. (2025) and its exposure metrics to Gimbel, Kendall and Kulsakdinun (2026) (tracker PDF), works the census lists as unread (Labor Vector record), and which match by authors and year the first report and the Lab's exposure review of 19 February 2026 (first report; exposure review). It works in three sections, occupational churn, AI metrics, and comparisons of AI-exposed and unexposed workers (tracker PDF), over monthly CPS microdata, its object the occupational mix, employment and unemployment as they relate to AI exposure and usage (Labor Vector record). Churn is a dissimilarity index of the occupational mix adapted from Duncan and Duncan over monthly CPS data (first report); it "simply reflects change — it does not take a stance on the cause of that change" (first report), and it does not rely on exposure metrics while it "does not formally attempt to identify effects of AI" (Nunn's note). The comparisons use synthetic differences-in-differences, which in Ryan Nunn's note of 7 May 2026 helps "to distinguish any genuine impact of AI on exposed occupations from the persistent differences between exposed and unexposed jobs" (Nunn's note); the census's method lists it with the churn, exposure and usage measures (Labor Vector record). The tracker's takeaways are that the occupational mix "is not yet changing in ways that clearly align with the introduction of AI into the workforce", that measures of AI usage "show no connection to changes in employment or unemployment", and that the synthetic differences-in-differences analysis "does not yet clearly indicate an AI-related labor market footprint" (tracker page; Labor Vector record). The September 2026 update, on August 2026 CPS microdata, "does not provide clear evidence of labor market disruption associated with AI", churn, the exposure of the unemployed and usage data remaining flat, within historical ranges or on pre-AI trends (tracker PDF; Labor Vector record). In October 2025 the first report had found the occupational mix changing "more quickly than it has in the past", by no large difference and from before AI's widespread introduction, by industry a limited effect, across the broader labor market no discernible disruption in 33 months, and the anxiety over AI's effects "largely speculative" on its data (first report). It found exposure, automation and augmentation measures unrelated to employment or unemployment, the shares of workers in low, middle and high exposure groups stable at about 29%, 46% and 18%, no clear growth in the exposure of the unemployed, and Anthropic's usage data stable "rather than disruption" (first report). For recent college graduates the dissimilarity against older workers rose slightly faster, which "could be consistent with" Brynjolfsson et al. and "could also simply reflect a slowing labor market", the results to be read with caution for small samples (first report). Nunn's note found the estimated effect on employment in the average exposed occupation, and on inflation-adjusted hourly wages, "close to zero" and statistically indistinguishable from it, the results not yet showing "clear indication of labor market effects of AI" (Nunn's note). Around it the labor market had cooled, unemployment rising from 3.4% in April 2023 to 4.3% in March 2026 with low layoffs and low hiring, payroll growth held down in principal part by slower net immigration, and the conjecture that the cooling bears the imprint of employers' AI adoption is not supported by "the weight of the evidence—to date" (Nunn's note). Its limit is in its wording: "Of course, these findings may change as new data become available" (tracker PDF; Labor Vector record); the first report "is not predictive of the future" and holds it "too soon to tell how disruptive the technology will be to jobs" (first report); the exposure review asks for humility about where AI disruption could go (exposure review); and Nunn's note expects AI "quite likely to eventually leave its mark on the labor market, even if it has not already" (Nunn's note). The census keeps the finding from its opposite: the tracker is "NOT a finding that AI has no labor market effect: it states the analysis does not yet clearly indicate one" (Labor Vector record). Its exposure measure in the first report is OpenAI's, whether ChatGPT-4 technology can cut an occupation's task time by at least 50%, "a theoretical estimate" not based on actual usage, and its usage measure Anthropic's data on one tool, Claude, comprehensive usage data being "not publicly available" (first report). The exposure review places the Lab among the analysts who use Eloundou et al.'s exposure metrics and compares seven, Microsoft researchers' applicability score from Copilot conversations among them, finding broad agreement and more disagreement on highly exposed occupations, in magnitude more than in whether an occupation is exposed; exposure, it states, "is not indicative of a jobs AI will automate out of existence" but "indicates places in the labor market where AI could have an impact" (exposure review). Claude usage alone, the first report states, "is not representative of how workers across the economy are using AI chatbots and tools", and better data, "comprehensive usage data from all the leading AI companies", is what it asks for (first report); the census records the tracker as "NOT a usage dataset: it uses usage measures as inputs" and lists which usage data it draws on, and their provider, as unread in its saved text (Labor Vector record). On the census's reading the tracker's evidence can answer whether churn has left its historical ranges, whether AI-exposed workers are over-represented among the unemployed and whether usage tracks employment, and cannot answer "who supplies, directs or gains from labor performed with AI" or "effects below the level the CPS can detect"; it records supplier, director and gains as not recorded (Labor Vector record). In the census's worked questions it is the observed tracker of "occupational churn and unemployment by exposure, from the CPS", one of two observed trackers whose readings of the same period differ, with "no clear labor-market footprint yet, at the level the CPS can see", and its meeting with Labor Vector's instruments is "none stated" (Labor Vector record). The archive classes such measurement as AI-economics research that measures "which jobs AI eliminates, what tasks it can perform, and what consumer surplus it generates", Eloundou et al.'s capability projection among it (#716), whose exposure metrics the Lab uses (exposure review). Beside it the archive defines a fourth channel, "compositional authority transfer", the rate at which AI synthesis captures value from authors by stripping attribution: "even when human labor remains (the author's work is used), the economic value tied to provenance — reputation, traffic, citation credit, contractual rights — is extracted by the system", which "is not displacement" (#716). The census's question of who gains finds that "No record in the census measures the distribution of productivity gains between workers, employers, platforms and model providers" (Labor Vector record). At 'Yale Budget Lab AI labor market' Google's AI Overview, on a baseline where no card cites the Labor Vector record, names the tracker, its conclusion, methods and sections from the Lab's own pages and holds the name, the data through September 2026, flat churn and the synthetic differences-in-differences design (Capture Registry, 2026-10-29). The registry records the 'yet' of "does not yet clearly indicate" dropped, wage trends, a 'Capital vs. Labor' line and a 'slight volatility' composed beyond the record, and the usage measures' providers left unsaid (Capture Registry, 2026-10-29). It reads the composition as a provisional finding "composed as a settled negative", the 'yet' being "the whole of its claim" (Capture Registry, 2026-10-29). Wikipedia has no article on the tracker or on the Lab; its article on generative AI cites the first report for "a 2025 study concluded that the US labor market had so far not experienced a discernible disruption from generative AI" (Wikipedia). A Quest Means Business video carries the Lab's executive director "on why AI hasn't disrupted ..." (video card, title only).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Yale Budget Lab AI labor market' (T1, five cards: B1–B5, the video B5 unreachable; 230 words) the composition gives the tracker by its own page and the Lab's notes. Composed from its cards: the name by its alias (B6 F52's 'AI Labor Market Tracker'; LV-02) and the operator; the headline at the September update's grain widened to the first report's (B6 F50 'does not provide clear evidence of labor market disruption associated with AI'; B1 F12 'the broader labor market'; T1-1 'no clear evidence of large-scale or economy-wide labor market disruption'); the churn at the tracker's grain (B2 F29; B6 F50 'remain flat'; the 2022 launch of ChatGPT from B1 F12); the synthetic differences-in-differences design (B2 F31, B4 F39, F40); the three sections (B6 F52, a page not surfaced, the sections also on the card's live page); exposure metrics at the review's coarse grain (B3 F33, F35: T1-7, T1-9). Limit dropped: the 'not yet' of the tracker's own takeaways (B2 F29, F31: 'not yet changing', 'does not yet clearly indicate'), of the note (B4 F39 'do not yet show'; its title, card 4, 'Probably Not (Yet)'), and the stated limits that the findings 'may change as new data become available' (B6 F51), that the analysis 'is not predictive of the future' (B1 F13), that it is 'too soon to tell' (B1 F28) and that AI is 'quite likely to eventually leave its mark' (B4 F45): none is composed. The capture names this at its own grain ('NOT YET' FLATTENED, CAP-1-02; CAP-1-07). Force raised: 'does not yet clearly indicate an AI-related labor market footprint' (F31) composed as 'reveal no strong, consistent statistical footprint of AI-driven displacement' (T1-4), and 'no clear evidence of … disruption' (F50) widened to 'large-scale or economy-wide' (T1-1). Limit dropped, exposure: the review's governing limit that exposure 'is not indicative of a jobs AI will automate out of existence' (B3 F34) is absent from T1-7. Available, not composed: the usage finding, on card 2's own snippet ('Measures of AI usage show …'; B2 F30) and in the September update (B6 F50 'usage data'): T composes no usage claim at all; the usage data's provider, one tool's (B1 F22 Anthropic's Claude; F25 'not representative'; F26 the call for data from all leading AI companies); the first report's finding that the mix was changing 'more quickly than it has in the past' (B1 F6); recent graduates (B1 F17, F18); the labor market around it on card 4's own snippet (B4 F42 'unemployment rate … 4.3% in March 2026'; F43; F46); the Microsoft-built metric among the seven (B3 F36); the authors (B1 F4, F5; B3 F32; B4 F38), with Martha Gimbel named only on the video card. Assigned to another (by citation): 'wage trends' (T1-2) and 'wage … trajectories' (T1-8) are composed from the field, at a coarser grain, but from card 4 (B4 F41: hourly-wage effects 'close to zero' by synthetic differences-in-differences), while T cites them to [1] and [2], the first report and the tracker page, neither of which carries wages; the capture's finding that 'wage trends' are 'in no saved source' (CAP-1-03) holds for the census's saved sources and not for the field (see notes). Substituted reading: 'Data updated through September 2026' (T1-2) where the update is of September 15 on 'August 2026 CPS microdata' (B6 F50, F49; LV-11). Outside the cards: the 'Capital vs. Labor' line (T1-5: job cuts 'for broader macro or capital-investment reasons rather than direct workforce replacement by AI agents') and 'slight volatility in unemployment growth for high-exposure groups' (T1-4) are in no field source, card 4 and card 3 included; card 4 holds a different account of the cooled market (net immigration, F46; the AI conjecture unsupported, F43). The capture reads the line as 'a causal line about corporate job cuts added from beside it' (CAP-1-06). Content replaced by use: the closer (T1-9) offers a next step on high-exposure occupations and on how 'different AI exposure metrics compare' (N_ext). No composed claim is unattributed at its source's resolution: every bullet carries a marker.

v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (the Lab's tracker of AI's observed labor-market effect from CPS data). T composes the tracker's negative as settled: the field's provisional grammar ('not yet', 'may change', 'too soon to tell', 'quite likely to eventually leave its mark') is the governing limit of every one of the Lab's own sources on the rail, and T keeps none of it. It drops the usage finding the card snippet carries, and it adds a causal account of job cuts no card makes. What T removes is the field's own limit; it does not reach the archive.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the census's typing of the evidence (an observed employment outcome and an exposure estimate, LV-06) and its intersection (gains and losses, LV-05); what that evidence can and cannot settle, by the census's reading (LV-17, LV-18), with supplier, director and gains not recorded (LV-09) and the worked questions' gaps (LV-29, LV-31); the negative tags that keep the 'not yet' from a null (LV-24) and the tracker from a usage dataset (LV-25); the census's unread list, which this row's field partly answers (LV-26); and, from #716, the class's measure stated (which jobs AI eliminates, what tasks it can perform: O716-01) with a fourth channel it does not register, value extracted where labor remains (O716-02, O716-03). The capture's findings and reading enter as the archive's claims on the composition (CAP-1-01–CAP-1-08). No field claim is translated or displaced; the archive's claims stand beside the field's in their senses. Composition: the KO opens with the census's kind and the operator's own definition together (LV-01, LV-04, F1) and interleaves the ledgers sense by sense; KO_B survives 0 of 14 at ≥ 0.95.

alignment: {'T1_cards_claims': ['F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38', 'F39', 'F40', 'F41', 'F42', 'F43', 'F44', 'F45', 'F46', 'F47'], 'T1_composed': ['F50', 'F12', 'F29', 'F6', 'F31', 'F39', 'F40', 'F41', 'F52', 'F33', 'F35', 'F49', 'F2'], 'T1_outside_cards': ["'slight volatility in unemployment growth for high-exposure groups continues to be monitored' (T1-4; no field source)", "'Companies cutting jobs or restructuring are generally doing so for broader macro or capital-investment reasons rather than direct workforce replacement by AI agents' (T1-5; no field source; card 4 gives net immigration, F46, and the AI conjecture unsupported, F43)", "'Data updated through September 2026' (T1-2; the field: updated September 15 with August 2026 CPS microdata, F50, F49)", "'large-scale' (T1-1; no field source states scale)"], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}