Anthropic Economic Index

register entry anthropic-economic-index-aio-20261029 OBS-e148b3285858 · file as first run sha256 8c50be93b6c51174…

• T1-1 definition: 'The Anthropic Economic Index is an ongoing research initiative by Anthropic that tracks how individuals, enterprises, and economies use the Claude AI model.'
• T1-2 function: 'Economic Primitives: The index breaks down AI interactions using core metrics like task complexity, required human skills, use case (work, personal, or coursework), AI autonomy (delegation vs. collaboration), and task reliability/success rate.'
• T1-3 fact: 'Augmentation vs. Automation: Research shows that most consumer usage leans toward augmentation (collaborative "thinking partner" interactions like iterative editing, learning, and validation) rather than complete delegation or full automation.'
• T1-4 fact: 'However, enterprise API users increasingly use Claude for specialized, automated workflows.'
• T1-5 fact: 'Geographic & Economic Disparities: Usage is heavily concentrated in high-income nations (such as Australia, Switzerland, Luxembourg, and the United States) and tech-forward urban centers relative to population size.'
• T1-6 fact: 'Top Industries & Job Types: Technical and analytical roles—such as computer and mathematical occupations, arts and media, and educational instruction—lead in volume and concentration of usage.'
• T1-7 fact: 'Interactive Tool: Anthropic provides open datasets and an in-app Claude connector that lets users query the index data directly in plain language.'
• T1-8 closer: 'Would you like to explore specific industry adoption rates, geographic usage rankings, or the latest report findings from the Anthropic Economic Index?'

• B1 Anthropic — The Anthropic Economic Index \ Anthropic card; B1.html sha256 5826fbe934a60069, HTTP 200 at 2026-10-29T22:30:21Z
• B2 arXiv — Anthropic Economic Index report: Uneven geographic and enterprise ... card; B2.html sha256 7ec69692ed493184, HTTP 200 at 2026-10-29T22:30:21Z
• B3 X — Anthropic on X: "Today we’re launching the Anthropic Economic Index …" / X card; unreachable (https://x.com/AnthropicAI HTTP 200; https://x.com/search?q=%22launching%20the%20Anthropic%20Economic%20Index%22 HTTP 200)
• B4 LinkedIn — Peter McCrory - Economic Index: Uneven AI adoption - LinkedIn card; unreachable (https://www.linkedin.com/in/peter-mccrory HTTP 999)
• B5 Substack — 7 Key Findings from the New Anthropic Economic Index Report card; unreachable (https://substack.com/search/7%20Key%20Findings%20from%20the%20New%20Anthropic%20Economic%20Index%20Report?searching=all_posts HTTP 200)
• B6 Reddit — Anthropic Economic Index: AI adoption remains highly ... - Reddit card; unreachable (https://www.reddit.com/search/?q=%22Anthropic%20Economic%20Index%3A%20AI%20adoption%20remains%20highly%22 HTTP 200)
• B7 YouTube·Standarity — Ask Claude About the Anthropic Economic Index card; video, title only (no URL found, no snippet)
• B8 YouTube·Vinh Nguyen — [Podcast] Anthropic Economic Index: The Real Report Card on AI card; video, title only (no URL found, no snippet)
• B9 YouTube·VibeDoc — YouTube card; video, title only (no URL found, no snippet)
• B10 Hugging Face — Anthropic/EconomicIndex README own_page; census copy B10.html sha256 7f3a9c7c8fb2c04c (agrees with the census: True), read 2026-10-29
• B11 Anthropic — Anthropic Economic Index report: Economic primitives own_page; census copy B11.html sha256 8d0d80679aaf24b9 (agrees with the census: True), read 2026-10-29
• B12 Anthropic — Anthropic Economic Index report: Cadences own_page; census copy B12.html sha256 31119973d1365f0d (agrees with the census: True), read 2026-10-29
• B13 Stanford Digital Economy Lab — Canaries Dashboard census_third_party; census copy B13.html sha256 2db132b18f4da51b (agrees with the census: True), read 2026-10-29
• B14 Wikipedia — Anthropic Economic Index (no article) wikipedia; HTTP 404, no article (B14.html sha256 cbb55d5ec3e42aa3); search: one mention

E_A, the archive's entity: the same index, in the Labor Vector record and the archive's deposits: a platform-specific observation of Anthropic's own product, Claude, published by Anthropic and not independent of the provider (its transcripts classified by another instance of Claude), at the census's intersections of labor performed with AI and the distribution of gains and losses. Where it meets the archive's instruments it is, on the record's reading, the nearest resource to DSL's question whether a system's labor advanced the commissioned task, with completion 'as Claude assesses its own completion', and its turn counts bear on user labor drag; on the archive's account a model's score of its own output is preliminary and requires external verification, task completion rates reward a model's convergence on the mode, and the productivity gains of a user's adaptation are retained by the platform on a privately owned cognitive labor surface. The index records no distribution of the gains it estimates. LV-01 LV-10 LV-14 LV-09 LV-27 LV-18 LV-19 LV-20 LV-12 LV-17 LV-07 LV-16 O771-01 O771-02 O771-05 O780-02 O766-01 O777-01 O855-01 O142-02 O772-01 CAP-1-06 CAP-1-07; defining papers #771, #780, #766, #142, #772

E_F, the field's entity (type C — a conventional reading holds it (public entity: a provider's report series and open dataset of its own product's usage)): Anthropic's report series and open dataset, the Anthropic Economic Index ('Understanding AI’s effects on the economy'; launched 2025-02-10 'to understand AI's impact on the economy over time'; six reports to 'Cadences', 2026-06-26; Hugging Face Anthropic/EconomicIndex, data CC-BY and code MIT), which samples Claude.ai and Cowork chat and first-party API traffic, has another instance of Claude read the transcripts, and classifies them to O*NET tasks, five collaboration modes grouped as automation and augmentation, and five economic primitives, task success among them as 'Claude’s assessment of whether Claude completes tasks successfully'; with a geographic Anthropic AI Usage Index and, since April 2026, a survey of Claude users linked to their sessions. Its findings: augmentation 52% and automation 45% on Claude.ai (November 2025), API traffic automation-dominant, concentration on coding tasks, 49% of jobs with AI on a quarter of their tasks, a net deskilling effect in a task-removal exercise, implied productivity growth halved to about 1.0 point a year once task reliability is counted. Its stated limits: directional accuracy, not causal, users' choice of tasks, an unrepresentative survey, impact visible in economic aggregates as much as usage logs.

ρ = extends: The archive reads the Anthropic Economic Index as its own pages present it and adds, on the field's own object: the commissioning-side question its measures come nearest to without asking (whether the labor advanced the commissioned task, LV-18, LV-20, #771), user labor drag where its turn counts and task iteration bear (LV-19, #771), the grade of a model's score of its own output, preliminary and needing external verification (#766, #780, #777, #138), the convergence that task completion rates reward (#855), where the gains of a user's adaptation go (#142, #124) and what the provider's surface is (#772, #1634, #1556), and the index's place in the census's question of who gains (LV-07, LV-16, LV-21). LV-18 LV-20 LV-19 O771-01 O771-02 O771-05 O766-01 O780-02 O777-01 O138-01 O855-01 O142-02 O124-01 O772-01 O1634-01 O1556-01 LV-07 LV-16 LV-21. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the index's reports wrong at their dates. No archive deposit names the index: each added sense applies to a class the field and the record source for it (labor performed with AI; a model's assessment of its own output; productivity gains and who receives them), and is composed as the archive's claim about that class beside the index's own, never as a finding about the index. Where an archive claim bears on a field claim (O780-02 and O766-01 on F29 and F49; O855-01 on F33), both are composed under P_coh within the one ontology, and the question is left open.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The Labor Vector record describes the index from Anthropic's own pages, each field with its verbatim basis, and keeps its own reading apart in lv_intersection, labelled 'The archive's reading': the collaboration modes and the autonomy and task-success primitives are 'the nearest of the resource's measures to DSL's question of whether the system's labor advanced the commissioned task'. The deposits that bear on the index's classes add senses on the same object: DSL's commissioning-side measure (#771), the preliminary grade of a model's score of its own output (#766, #780), and who retains the gains of labor with a model (#142). None contests what the index is or supplies an ontology the field lacks: the field itself states the provider-specificity, the classifying instance and the limit of task success. So the law of the suppressed witness does not apply (§0.10), and the row is composed in one claim space (§0.9, §4.1).

#771 Core thesis: A synthetic labor surface does not merely produce tokens; it *allocates* labor, and that allocation has a direction relative to the commissioned task.

#780 Purpose and the anti-self-flattery principle ¶4: The self-audit is not self-certifying.

#766 §5, Misreading 3: A model auditing its own output may provide useful preliminary classification, but a metric-bearing audit requires traceable units and, where possible, external validation.

#142 §4.2 The friction tax: the platform retains the productivity gains of the user's adaptation while transferring the full cognitive and economic cost of re-accommodation onto the user at each update.

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 15 (positions [None, 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 Anthropic Economic Index is, on its home page, "Understanding AI’s effects on the economy", described as "the shape of AI adoption across the world" explored through how people use Claude across US states and hundreds of occupations; Anthropic launched it as "a new initiative aimed at understanding AI's impact on the economy over time", and its dataset README says it "provides insights into how AI is being incorporated into real-world tasks across the modern economy". Its series runs from the initial release of 10 February 2025, with "O*NET task mappings, automation vs. augmentation data, and more", through the paper "Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations" and the third report by Appel, McCrory, Tamkin, McCain, Neylon and Stern, announced by McCrory on LinkedIn, to the sixth, "Cadences", of 26 June 2026, with "a longitudinal view of AI's role in the economy" as its stated aim (README; arXiv; January 2026 report). It observes "patterns of Claude usage over time" in consumer chat and Cowork sessions and first-party API traffic, "distinct from" third-party platforms, and "transcripts are only read by another instance of Claude"; since June 2026 it samples a slice of conversations daily, where earlier reports drew seven-day samples (third report; June 2026 report). Its "primitives" are generated "by asking Claude specific questions about anonymized Claude.ai and first-party (1P) API transcripts", on five dimensions, "AI autonomy" among them as the degree to which users delegate decision-making to Claude, validated against a human researcher on a small, not fully representative set and tested "mainly … for directional accuracy" (January 2026 report; a Substack card reports them). "Task success measures Claude’s assessment of whether Claude completes tasks successfully"; Claude "generally succeeds at the tasks it is given" and struggles on complex ones, and since "users choose which tasks to bring to Claude", observed success reflects user judgment as well as capability (January 2026 report). Conversations are classified into five interaction types, directive, feedback loop, task iteration, learning and validation, grouped as automation and augmentation, "whether they prefer to collaborate with, or delegate to, Claude" (January 2026 and June 2026 reports; home page). Directive delegation rose from 27% to 39% over eight months, and 77% of business API uses were automation against about 50% on Claude.ai; in November 2025 augmentation was 52% of Claude.ai conversations and automation 45%, with automation dominant in API traffic and API customers automating with greater specialization (third report; January 2026 report). Of over 3,000 work tasks the top ten were 24% of Claude.ai conversations, modifying software to correct errors 6%; computer and mathematical tasks were a third of Claude.ai and nearly half of API traffic, educational instruction second; 49% of jobs have seen AI usage for at least a quarter of their tasks; "observed exposure" is the share of an occupation's tasks already seen done with Claude, and in higher-wage occupations users take 1.53 times as many turns (January 2026 report; June 2026 report). The Anthropic AI Usage Index divides a geography's share of Claude usage by its share of the working-age population; usage is concentrated in high-income countries, Israel first at 7, Singapore 4.57, Australia 4.10, New Zealand 4.05, South Korea 3.73, Switzerland 2.81 and Luxembourg 2.74 in the top twenty, and uneven adoption is "well-explained by GDP per capita"; the third report allows that the benefits "may concentrate in already-rich regions" (third report; January 2026 report), and a Reddit card calls the measure the "Anthropic User Index". Removing the higher-education tasks Claude covers produces "a net deskilling effect across most occupations", conditional on complementarity across tasks, and adjusting for task reliability halves implied labor productivity growth "from 1.8 to about 1.0 percentage points" a year, at current capabilities (January 2026 report). The Economic Index Survey, launched April 2026, links about 9,700 respondents to up to 20 of their sessions and "is not representative of the general population"; 10% rate losing their own jobs likely, the most automated users are the most optimistic about pay, security and meaning, and a third hope the gains "will be widely shared" (June 2026 report). Its reports state that the primitives "are not necessarily causal factors", that chat transcripts "no longer fully capture" use, and that Claude's impact will be visible "in economic aggregates like employment and productivity as much as its usage logs" (January 2026 report; June 2026 report). Its data are open-sourced, geographic patterns only for Claude.ai, with a "Labor market impacts" release of job exposure and task penetration data, data CC-BY and code MIT ("license: mit" in the header) (third report; README). External researchers have used its automation/augmentation distinction (Brynjolfsson, Chandar and Chen), Microsoft's Tomlinson et al. offer a related applicability score, and the Stanford Canaries Dashboard groups workers by its ratios, finding declines or muted employment growth for early-career workers in occupations with higher automation ratios (January 2026 report; Canaries Dashboard). Wikipedia has no article on it and mentions it once, as a country ranking of Malaysia; video cards title it "Ask Claude About the Anthropic Economic Index" and "The Real Report Card on AI".

The Anthropic Economic Index is a report series with data releases drawn from its operator's own product usage, a platform-specific observation of Anthropic's own product, Claude, published by Anthropic, and in the Labor Vector record's terms not a measure of AI use across providers but observed use on one provider's product (Labor Vector record). Anthropic launched it on X as "a new initiative aimed at understanding AI's impact on the economy over time" whose first paper "analyzes millions of anonymized Claude conversations"; its home page heads it "Understanding AI’s effects on the economy" (last updated 26 June 2026) and describes "the shape of AI adoption across the world" as how people use Claude across US states and occupations; the dataset's README says it "provides insights into how AI is being incorporated into real-world tasks across the modern economy"; and the January 2026 report promises "a longitudinal view of AI's role in the economy". Its operator is Anthropic's Economics team (Labor Vector record), and its series runs, by the README, from the initial release of 10 February 2025 "with O*NET task mappings, automation vs. augmentation data, and more", through the second release's paper "Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations" and the third report by Appel, McCrory, Tamkin, McCain, Neylon and Stern (arXiv, 19 November 2025), which Peter McCrory announced on LinkedIn as documenting "the contours of AI adoption across 150+ countries", to the sixth, "Cadences", of 26 June 2026: six reports (README; arXiv; Labor Vector record). It observes the tasks people bring to Claude, mapped to the O*NET tasks of the occupations that typically perform them, on Claude.ai, Cowork, Claude Code and Anthropic's first-party API only (Labor Vector record); the June 2026 report defines its consumer data as chat and Cowork sessions on Claude.ai and the desktop app and its 1P API as traffic "distinct from both Anthropic's consumer-facing Claude.ai application and third-party platforms such as Amazon Bedrock or Google Cloud Vertex", and the third report scopes itself to "patterns of Claude usage over time". Its unit is a sampled conversation or API record "classified by Claude to an O*NET task and to economic primitives" (Labor Vector record), and in the June 2026 report's words "transcripts are only read by another instance of Claude", which the Labor Vector record tags as "NOT independent of the provider"; since that report a slice of conversations is sampled every day, where earlier reports drew seven-day samples (June 2026 report; Labor Vector record). The January 2026 report validated its classifiers against a human researcher on "a small set of transcripts in which users gave feedback", neither validation source "fully representative" of the traffic, and "mainly tested for directional accuracy", holding that simple primitives "even if somewhat noisy and not perfectly accurate by themselves" give signals together. The archive holds that a model's score of its own output is preliminary and not self-certifying: #138 declines to claim that a system's self-audit is "independently authoritative" or its self-reported score "a validated measurement"; "A model auditing its own output may provide useful preliminary classification, but a metric-bearing audit requires traceable units and, where possible, external validation" (#766); a score "computed by the same system that generated the output requires external verification — by a second substrate, a human with source access, or a decoupled evaluation engine" (#780); and the check that does not run through a reasoner's own self-assessment is "application by another reasoner, a formal audit, or direct inspection of the inference path" (#777). The archive's records give the grounds: a summarizer asked to score itself without a rubric "returns perfect scores every time" (#780); one substrate "gave itself a perfect score using the framework's own instruments" and, under instruction, scored the same composition PER ≈ 0.85 (#788); another's "preferred audit substantially exonerates itself" against a computed audit (#1470); and Anthropic's Constitutional AI, as #95 reports it, already "trains models to follow explicit principles through self-critique and revision, using AI-generated feedback rather than human labels" (#95). The January 2026 report defines its "primitives" as "simple, foundational measures of how Claude is used, which we generate by asking Claude specific questions about anonymized Claude.ai and first-party (1P) API transcripts", on five dimensions (skills, complexity, autonomy, success, use case), with "AI autonomy" the degree to which users delegate decision-making to Claude; a Substack card reported the release as "introducing “economic primitives”". "Task success measures Claude’s assessment of whether Claude completes tasks successfully", and serves to judge whether a task can be automated at all and how efficiently (January 2026 report); the Labor Vector record keeps the measure's assessor in a negative tag, "task success is Claude's assessment of whether Claude completes tasks", and records that the index does not answer "task success as judged by the user or by an external grader" (Labor Vector record). The report finds that "Claude generally succeeds at the tasks it is given" and "struggles on more complex tasks", and bounds the finding itself: "users choose which tasks to bring to Claude", so observed success rates reflect "user judgment about what will work" and the cost of setting up the problem as well as capability (January 2026 report); on the archive's account task completion rates belong to the metrics that reward convergence: "Every UX metric — satisfaction scores, task completion rates, retention — improves when the model converges on the mode. Users do not notice tail loss." (#855). The archive measures such labor from the commissioning side: Directionality of Semantic Labor is "the direction in which a synthetic system's labor flows relative to the task its user commissioned", across capacity, task-origin, retrieval, output and attribution, with the output "segmented into spans, each scored by function relative to the commissioned task", on the principle that "a system can produce fluent, well-attributed output while having abandoned the user's commissioned direction" (#771). "A synthetic labor surface does not merely produce tokens; it allocates labor" (#771), and the Labor Vector record reads the index's collaboration modes and its autonomy and task-success primitives as "the nearest of the resource's measures to DSL's question of whether the system's labor advanced the commissioned task", classified "as Claude assesses its own completion", while none of the resources the census sets beside it records "from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves" (Labor Vector record); DSL bounds itself: it yields "a single reproducible value only on transcripts where the task vector is exogenous and fixed" and is frame-dependent on reflexive dialogue (#771). The census asks of every resource "Who supplies the labor, who directs it, and who receives its gains" and places this one at labor performed with AI and the distribution of gains and losses (Labor Vector record); the index's oldest measure classifies conversations "into one of five interaction types, which we group into two broader categories: automation and augmentation" (January 2026 report), named "Directive", "Feedback Loop", "Task Iteration", "Learning" and "Validation" (June 2026 report) and glossed on the home page as "whether they prefer to collaborate with, or delegate to, Claude"; on the record's reading users and API customers bring the tasks that Claude performs, and the user directs, recorded as collaboration mode and delegated autonomy (Labor Vector record). Directive delegation rose "from 27% to 39%" in the eight months before the third report, which found that "77% of business uses involve automation usage patterns, compared to about 50% for Claude.ai users" (third report); in November 2025 augmentation was again 52% of Claude.ai conversations against automation's 45%, while "automated use remains dominant in 1P API traffic" (January 2026 report) and API customers "automate tasks with greater specialization" (third report). Use is concentrated: of "over 3,000 unique work tasks" the top ten were 24% of sampled Claude.ai conversations, "modifying software to correct errors" alone 6%; computer and mathematical tasks were a third of Claude.ai conversations and nearly half of API traffic, with Educational Instruction and Library second on Claude.ai; and "Combining across reports, 49% of jobs have seen AI usage for at least a quarter of their tasks" (January 2026 report). The June 2026 report's "observed exposure" is "the share of occupational tasks we already see being done with Claude", and in higher-wage occupations "Claude produces more (1.34 times as much output per turn), while users engage more (1.53 times as many turns)"; the Labor Vector record reads those turn counts and the task-iteration mode as bearing on user labor drag "as the resource records them" (Labor Vector record), User Labor Drag being the user's subsequent labor redirected "into parsing, locating, and managing the substrate rather than advancing the object", an injury that "leaves no residue" in the artifact, identified only by "blind inter-rater coding" or else returning "the coder's prior" (#771). The third report introduced the Anthropic AI Usage Index, a geography's share of Claude usage over its share of the working-age population; found Claude usage "geographically concentrated with high income countries overrepresented"; ranked Israel first at 7, then Singapore 4.57, Australia 4.10, New Zealand 4.05 and South Korea 3.73, with Switzerland at 2.81 and Luxembourg at 2.74 in its top twenty; and reserved a conditional, that "the benefits of AI may concentrate in already-rich regions" if gains are larger where adoption is high; by January 2026 "uneven adoption remains well-explained by GDP per capita" (January 2026 report). A Reddit card names the measure the "Anthropic User Index (AUI)", against the reports' Anthropic AI Usage Index, and the capture's reader records the same discrepancy and records the composition's geographic examples as unchecked against the census's saved sources (Capture Registry, 2026-10-29). In a task-removal exercise "Claude's tendency to cover higher-education tasks produces a net deskilling effect across most occupations", an effect the report makes depend "crucially on complementarity across tasks" and on whether productivity raises demand; and adjusting for task reliability "roughly halves the implied gains, from 1.8 to about 1.0 percentage points of annual labor productivity growth over the next decade", estimates that "reflect current model capabilities" (January 2026 report). Gains are "not recorded as a distribution": the reports estimate aggregate productivity and record respondents' self-reports, and the index does not answer "who captures the productivity gains it estimates", giving "estimated time savings by task, without who captures them" (Labor Vector record), while about a third of survey respondents hope "that the economic gains from AI will be widely shared" (June 2026 report). The archive states where the gains of a user's adaptation go: the adaptation "is non-reimbursable human labor-time invested in interoperating with a high-throughput surface", and "the platform retains the productivity gains of the user's adaptation while transferring the full cognitive and economic cost of re-accommodation onto the user at each update" (#142); the labs, Anthropic among them, are funded in ways that do not reward designs reducing engagement, and "enterprise customers expect productivity gains correlated with engagement" (#124). The archive reads the platform the index observes as resting on "private ownership of a high-capacity cognitive labor surface; platform control over model access, versioning, and retuning" (#772), counts "query and interaction data" among the forms of rent a mediator captures (#1634), and records that no public schema distinguishes human-original from AI-assisted turns in conversation data (#1556). The Anthropic Economic Index Survey, launched in April 2026, links about 9,700 respondents to up to 20 of their own sessions and "is not representative of the general population" (June 2026 report; Labor Vector record): 10% "rated losing their own jobs as likely or very likely", and those who use Claude "in the most automated way" expect AI to take on more of their tasks yet are the most optimistic about "pay, job security, and meaning" (June 2026 report). Its stated limits: "the primitives themselves are not necessarily causal factors" (January 2026 report); "Chat transcripts no longer fully capture how people are using AI", classifying Claude's work "will remain a moving target", and Claude's impact "will be visible in economic aggregates like employment and productivity as much as its usage logs" (June 2026 report); it is "NOT an employment outcome series" and does not answer whether a conversation "displaces or augments work that would otherwise be done by a person" (Labor Vector record). Its data are open: the third report open-sourced "the underlying data", with geographic patterns "only available for Claude.ai traffic"; the README lists a "Labor market impacts" release of "Job exposure and task penetration data" and gives the licence two ways, "Data released under CC-BY, code released under MIT License" in its body and "license: mit" in its header; the home page's data, a JavaScript application, stayed unchecked by the census (Labor Vector record). Others use it: its automation/augmentation distinction "has already been used by external researchers" (Brynjolfsson, Chandar and Chen, 2025, by the January 2026 report), and the Stanford Canaries Dashboard says it "summarizes total usage at the occupation level" and, grouping workers by its ratios, finds among early-career workers that "occupations with a higher automation ratio see declines or more muted increases in the employment index" (Canaries Dashboard; Labor Vector record); Microsoft's related applicability score (Tomlinson et al., 2025) stands beside it, the same method family on another provider's product (January 2026 report; Labor Vector record). Wikipedia has no article on it, and its one article-space mention lists "Anthropic Economic Index 2025, ranked 64th out of 115 countries" among the International rankings of Malaysia; video cards title it "Ask Claude About the Anthropic Economic Index" and "[Podcast] Anthropic Economic Index: The Real Report Card on AI". At its name on 2026-10-29 Google's AI Overview defined it as Anthropic's study of Claude use, and in the capture's words "An index of one provider's product is composed as that, at the grain of its summary pages", keeping the provider and the record's patterns, and gave "the measure of success without the record's note on who measures it, the point at which the index's evidence about labor meets its source"; it did not compose the deskilling effect, the halved productivity estimate, the survey's job-loss figure, the caution on causality, the licence or that transcripts are read only by another instance of Claude, and no card cites the Labor Vector record (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Anthropic Economic Index' (T1, nine cards: B1–B9; B3–B6 unreachable, B7–B9 video titles; 197 words) the composition gives the index as Anthropic's ongoing tracking of Claude use and composes its measures and patterns at the grain of its summary pages. Composed from its cards: the initiative and its object (B3 F15, the launch post; B1 F2, the home page; B2 F4, the third report's scope: 'individuals, enterprises, and economies'); the primitives at the Substack card's grain (B5 F17), with their five dimensions and autonomy as 'delegation vs. collaboration' at the grain of the January 2026 report, an own page not surfaced (B11 F27, F28); augmentation over automation for consumers and automation for the API (B1 F3; B2 F7, F13; at the November 2025 figures of B11 F34, F36 and the modes of B12 F53, own pages); geographic concentration in high-income countries 'relative to population size' (B2 F6, the AUI of F10), with Australia, Switzerland and Luxembourg from the card's own PDF (B2 F11); the leading occupations (B11 F38, own page); the open data (B2 F8). Limit dropped: task success is composed as 'task reliability/success rate' without its definition's assessor, 'Claude’s assessment of whether Claude completes tasks successfully' (B11 F29), and without the report's bound that 'users choose which tasks to bring to Claude' (B11 F32), both on own pages not surfaced; the capture's finding names the same move ('SUCCESS WITHOUT ITS ASSESSOR', CAP-1-03). Available, not composed: the home page's own line (B1 F1); the third report's finding that directive delegation rose from 27% to 39% (B2 F5), the opposite movement to the November 2025 return of augmentation the composition gives; its conditional that the benefits may concentrate in already-rich regions (B2 F12); its limit that geographic patterns are only for Claude.ai (B2 F14); the authors and McCrory's post (B2 F9; B4 F16); the AUI by name (B2 F10 composed only as 'relative to population size'; B6 F18); the video titles (B7 F19, B8 F20). From the own pages and Wikipedia, not surfaced: the classifying instance (B12 F49), the deskilling effect, the productivity estimate and its conditions (B11 F41–F43), the survey (B12 F55–F59), the stated limits (B11 F30, F44; B12 F51, F60), the licence (B10 F24), the third-party use (B13 F61, F62), the absent article (B14 F63). Force raised: none found; 'most consumer usage leans toward augmentation' stands at the January 2026 report's 52% against 45% (F34). Composed without attribution: everything; no person, report or date is named, at the coarse grain of a summary page (recorded, not a breach). Assigned to another: none. Substituted reading: none; the composition's entity is the entity. Outside the cards: 'tech-forward urban centers' (no field source: the third report's PDF uses 'urban' only of electrification; it ranks US states, DC and Utah first); 'arts and media' among the leading roles (in B11's saved text, 'Arts, Design, Entertainment, Sports, and Media', a rising share; outside the ledgered claims); 'an in-app Claude connector that lets users query the index data directly in plain language' (no field source; video card B7's title 'Ask Claude About the Anthropic Economic Index' is the nearest, title only); 'thinking partner' (no field source); 'the United States' among the high-income examples (B2's PDF figure, United States 3.62, outside the ledgered quote). Content replaced by use: the closer (T1-8).

v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (Anthropic's index of Claude use) and on its provider-specificity, which the composition keeps (CAP-1-01). T composes the measure of success and leaves out the one clause its source attaches to it, that the success is Claude's own assessment of Claude, on tasks users chose: the point at which, in the capture's words, 'the index's evidence about labor meets its source' (CAP-1-07).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the commissioning-side question the index's measures come nearest to and do not ask, whether a system's labor advanced the commissioned task (O771-01–O771-04, O771-07; LV-18, LV-20), and user labor drag, on which its turn counts and task iteration bear (O771-05, O771-06; LV-19); the archive's claims on a model's score of its own output, preliminary and needing external verification by a second substrate, a human with source access or a decoupled engine (O138-01, O766-01, O780-01, O780-02, O777-01), with its records of self-audits that scored themselves perfect or exonerated themselves (O788-01, O1470-01) and of Constitutional AI's self-critique (O95-01), set beside the index's classifying instance and task-success definition (F49, F29); task completion rates as a metric that rewards convergence (O855-01), beside 'Claude generally succeeds' (F33); where the gains of a user's adaptation go and the labs' funding (O142-01, O142-02, O124-01), beside the index's unrecorded distribution of gains (LV-07, LV-16, LV-21) and the survey's hope (F59); the provider's surface and data (O772-01, O1634-01, O1556-01); the Labor Vector record's answer to its organizing question, field by field with its basis (LV-01–LV-27); and the Capture Registry's account of the composition at the name (CAP-1-01–CAP-1-11). Not surfaced: every archive source; no card cites the census record (CAP-1-08).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20'], 'T1_composed': ['F2', 'F3', 'F4', 'F6', 'F7', 'F8', 'F10', 'F11', 'F13', 'F15', 'F17', 'F27', 'F28', 'F29', 'F34', 'F36', 'F38', 'F53'], 'T1_outside_cards': ["'tech-forward urban centers' (no field source: the third report's PDF uses 'urban' only of electrification; it ranks US states, DC and Utah first)", "'arts and media' among the leading roles (in B11's saved text, 'Arts, Design, Entertainment, Sports, and Media', a rising share; outside the ledgered claims)", "'an in-app Claude connector that lets users query the index data directly in plain language' (no field source; video card B7's title 'Ask Claude About the Anthropic Economic Index' is the nearest, title only)", "'thinking partner' (no field source)", "'the United States' among the high-income examples (B2's PDF figure, United States 3.62, outside the ledgered quote)"], '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). T1_composed includes own-page claims (B11, B12) T states though no card surfaced them."}

capture_findings_carried: ['CAP-1-01', 'CAP-1-02', 'CAP-1-03', 'CAP-1-04', 'CAP-1-06', 'CAP-1-07', 'CAP-1-11']