Working with AI: Measuring the Applicability of Generative AI to Occupations

register entry microsoft-working-with-ai-aio-20261029 OBS-25ca6976c874 · file as first run sha256 f9678438e04bab9b…

• T1-1 lede: 'Microsoft integrates artificial intelligence across its core software ecosystem, enterprise infrastructure, and internal operations, while shifting toward massive AI data center investments and proprietary model development.' (three markers)
• T1-2 product (heading 'Key AI Products & Platforms'): 'Microsoft Copilot: An AI-powered assistant embedded across Microsoft 365 apps (Word, Excel, PowerPoint, Outlook, Teams) and Windows to draft content, summarize meetings, and analyze data.' (two markers)
• T1-3 product: 'Azure AI: Cloud infrastructure providing enterprise-grade machine learning tools, model hosting, data security, and compliance mechanisms.' (one marker)
• T1-4 product: 'Microsoft Fabric: A unified data and intelligence platform connecting analytics and AI workloads.' (one marker)
• T1-5 models: 'Homegrown MAI Models: Microsoft has increasingly developed its own internal AI models (such as audio and multimodal tech) alongside its partnership with OpenAI.' (one marker)
• T1-6 workplace (heading 'Workplace Transformation & Workforce Impact'): 'Workflow Redesign: Microsoft emphasizes redesigning end-to-end business processes and deploying specialized AI agents rather than just dropping tools into old routines.' (one marker)
• T1-7 workforce: 'Workforce Shifts: To fund intensive AI infrastructure and data center expansion exceeding $100 billion, Microsoft has scaled back certain traditional roles and offered large-scale workforce buyouts to restructure teams around an AI-driven operating model.' (one marker)
• T1-8 closer: 'Would you like to know more about Microsoft 365 Copilot pricing, Azure AI enterprise features, or how AI is changing specific job roles?' (content replaced by use, N_ext)

• B1 Microsoft — Working with AI: Measuring the Applicability of Generative AI ... - Microsoft card; B1.html sha256 e1f8e1b60bb379f6, HTTP 200 at 2026-10-29T22:31:37Z (agent 'Mozilla/5.0 (compatible; Labor-Vector-census)')
• B2 Microsoft — Microsoft Copilot: AI built for work the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B3 Microsoft — Microsoft AI - Business Solutions and Tools the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B4 The Official Microsoft Blog — What we’ve learned from Microsoft's own AI transformation - The Official Microsoft Blog the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B5 YouTube·ABC News — Microsoft offers buyouts to 60000 employees as spending on AI grows the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B6 Microsoft — Microsoft Copilot Plans for Individuals | Microsoft 365 the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B7 Facebook — Microsoft is building its own AI model as it loosens ties to OpenAI the homonym's card; not fetched as field (addendum item 5); title and snippet only
• B8 own page — the results repository's README (github.com/microsoft/working-with-ai) census copy microsoft-working-with-ai--readme, sha256 b98be795b8a96490 (equal to the census's), read 2026-10-29; not refetched
• B9 own page — arXiv:2507.07935 abstract page (v7) census copy microsoft-working-with-ai--arxiv, sha256 8315551565cd5986 (equal to the census's), read 2026-10-29; not refetched
• B10 census third party — Anthropic Economic Index report, January 2026 (footnote 3) census copy anthropic-economic-index--jan-2026, sha256 8d0d80679aaf24b9 (equal to the census's), read 2026-10-29; not refetched
• B11 Wikipedia — no article; cited in 'Workplace impact of artificial intelligence' Wikipedia: no article (HTTP 404 on three titles); the citing article B11.html sha256 a2483d45f8ad1075, HTTP 200 at 2026-10-29T22:33:01Z

E_A, the archive's entity: the same study, in the Labor Vector census record and the archive's labor instruments: a paper with results files, by Microsoft Research, of the kind interaction data, exposure estimate and data access, at the intersection labor performed with AI, its unit a Bing Copilot conversation classified as user goal and AI action, its labor object the work activities users seek help with and those the AI performs; on the census's question, users supply the goals and the AI the actions, the user's goal as classified directs, and the gains are not recorded; tagged not a measure of AI's ability to replace jobs, not an employment outcome measure, not a measure of all workplace AI use, and not independent of the provider, a platform-specific observation of Microsoft's own product; a member of the class whose labor the archive measures by its direction relative to the commissioned task (#771), the entity's own goal/action separation and completion classifier meeting DSL there on the archive's reading, with the archive's bounds on the instrument (one-shot commissions; inferred tasks constrained to the user's input; user labor drag by blind coding) and the gap the record names, that none of the resources records from the commissioning side whether the labor advanced the user's task or who receives the time it saves; at its name, composed by the AI Overview of 29 October 2026 as Microsoft's own products and workforce, the study's card first and unused. LV-01 LV-02 LV-07 LV-06 LV-05 LV-08 LV-09 LV-10 LV-11 LV-12 LV-35 LV-36 LV-37 LV-38 O771-01 O771-02 LV-30 O771-06 O771-07 O771-05 LV-48 CAP-1-01 CAP-1-04 CAP-1-09; defining papers #771, #1392

E_F, the field's entity (type C — a conventional reading holds it (public entity: a paper with results files; a provider's observation of its own product's use)): the paper of Microsoft Research (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri; arXiv:2507.07935, v1 10 July 2025, v7 23 October 2026, NeurIPS 2026) and its results files (github.com/microsoft/working-with-ai, v1.1, CC BY 4.0) that measure AI applicability to occupations from real-world use: 200k anonymized conversations with Microsoft's own Bing Copilot, United States, 1 January to 30 September 2024, collected by Microsoft, classified by an LLM pipeline to the O*NET intermediate work activities they serve as user goals and as AI actions, with completion, impact scope and thumbs feedback combined into an AI applicability score per occupation (785 SOC codes, 149.8 million workers); finding information work the most common and successful AI-assisted activity (v7; by occupation group in July 2025), applicability across sectors and rising over the period, and predicting which occupations delegate to AI and which use it to assist; its authors' limit that the metrics do not measure the ability of AI to replace jobs, that its success measures are proxies, that consumer use does not represent the workplace and that its classifiers are imperfect. Cited by the Anthropic Economic Index; no Wikipedia article, cited in Wikipedia's article on the workplace impact of AI for forty jobs, media reports of obsolescence and criticism. At the address, six of seven cards are a distinct entity, Microsoft's own AI products and workforce.

ρ = extends: The archive reads Working with AI as its authors present it and adds, on the field's own object: the census's question of who supplies, who directs and who receives the gains, with the gains not recorded and outside what the evidence can answer; the provider's observation of its own product, tagged as such; the direction of labor relative to the commissioned task, task-origin as provenance, span scoring and user labor drag, with the instrument's identification limits; the reading that the entity's goal/action separation and completion classifier meet DSL; the gap no resource records, from the commissioning side; and the composition layer's substitution, at the name, of the operator's products and workforce for the study. LV-41 LV-42 LV-12 LV-29 LV-38 O771-01 O771-02 O771-03 O771-04 O771-05 O771-06 O771-07 LV-30 LV-48 CAP-1-01 CAP-1-04 CAP-1-06. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates. Where an archive claim stands against a field claim (LV-11 'separates' against F35 'allows us to predict'), both are composed under P_coh within the one ontology. #771's identification condition and its rule for inferred task vectors are composed as the archive's bounds on the class, not as a finding about the entity's classifier, which the archive has not audited (the paper's prompts and interrater reliability are unchecked, LV-40).

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 study from its own pages, each field with its basis ('operator'), and answers for it the census's question of who supplies the labor, who directs it and who receives its gains, field by field (LV-41, LV-42), recording the gains as 'not recorded' (LV-12). No deposit names the study. The archive bears on its class, labor performed with AI, the census's own intersection for it (LV-05, LV-06), through the Directionality of Semantic Labor (#771), which adds on the field's own object a direction relative to the commissioned task, task-origin as provenance, user labor drag and the measure's identification limits (O771-01–07), and through the record's reading that the entity's goal/action separation and completion classifier meet DSL (LV-30) and the worked question's gap that none of the resources records, from the commissioning side, whether the system's labor advanced the user's task or who receives the time it saves (LV-48). Each is an addition: #771's own relation to the measures it sits beside is 'Extension', and nothing in the archive contests that the entity is Microsoft Research's measurement of AI applicability from Bing Copilot conversations. The field holds the ontology (a study, its data, its score, its limits); the archive is not the only witness to an ontology the field lacks, so the law of the suppressed witness does not apply (§0.10). The equal-terms case (§0.9).

#771 header, Relation: Relation: Extension — directionality is the third axis beside magnitude (PER) and power-conditioning (Omega)

#771 Core thesis s1: A synthetic labor surface does not merely produce tokens; it *allocates* labor, and that allocation has a direction relative to the commissioned task.

#771 Abstract s4: The central principle is that provenance is not only claim-origin but *task-origin*: a system can produce fluent, well-attributed output while having abandoned the user's commissioned direction.

#LV census.json rows[slug == 'microsoft-working-with-ai'].lv_intersection: The archive's reading: the separation of user goal from AI action, and the completion classifier ('the LLM completed the user's goal'), are the resource's own instruments for whether the system's labor served the user's commission; they meet DSL there, measured by classifier and thumbs feedback. Task-origin is recorded as the user goal.

#LV questions.json questions[id == 'working-with-ai'].gap: None records, from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves. That is the question DSL measures.

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 12 (positions [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.

Working with AI: Measuring the Applicability of Generative AI to Occupations, by members of Microsoft Research (Tomlinson, Jaffe, Wang, Counts and Suri), measures AI applicability to occupations from real-world use, 200k anonymized conversations with Microsoft Bing Copilot, where earlier work relied on predictions of capability or on single firms (arXiv:2507.07935; README; the Microsoft Research page). Its metrics' goal is "to measure how AI may be useful in different occupations": how often and how successfully Bing Copilot was used for work activities, and which occupations AI may be applicable to (README), on a question both versions of its abstract call among society's most important, the economic effects of AI (the Microsoft Research page; arXiv). The conversations were collected by Microsoft from Bing Copilot in the United States between 1 January and 30 September 2024, mostly in English, and the occupations cover 785 SOC codes and 149.8 million workers (README). An LLM pipeline classifies the O*NET work activities assisted or performed by AI in each conversation, reporting user goals and AI actions separately at the level of intermediate work activities and detailed SOC occupations (arXiv; README). Completion, impact scope and thumbs feedback measure success, and with the activity classifications they give each occupation an AI applicability score, the user and AI sides averaged (README; the Microsoft Research page). The July 2025 abstract found the highest scores in knowledge work groups such as computer and mathematical and office and administrative support, and in sales, with users seeking help gathering information and writing and the AI providing information, writing, teaching and advising, and it set usage against predictions of occupational AI impact (the Microsoft Research page). The v7 abstract finds the most common and successful AI-assisted activities to be information work, applicability cutting across sectors and rising over the nine months, and states that the method predicts which occupations will delegate tasks to AI and which will use it to assist existing workflows (arXiv). The authors' limit is that the metrics "should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs": the success measures are proxies, the metrics say what a consumer LLM is used for, consumer use does not represent the workplace, and LLM-based classifiers are imperfect (README); the operator's page links a note on applicability against job displacement (the Microsoft Research page). The paper runs from arXiv v1 (10 July 2025) to v7 (23 October 2026, NeurIPS 2026), its results files at v1.1 under CC BY 4.0, its analysis approved by the Microsoft Institutional Review Board (arXiv; README). The Anthropic Economic Index's January 2026 report points to it for "a related AI applicability score" (Anthropic Economic Index). Wikipedia has no article for it; its article on the workplace impact of AI reports forty jobs identified from Microsoft Copilot interactions, media reports of jobs becoming obsolete, and criticism, Chris Campbell's among it (Wikipedia). Six of the seven cards at the address concern a distinct entity at the name, Microsoft's own AI products, transformation, workforce and models (the cards).

Working with AI: Measuring the Applicability of Generative AI to Occupations is a paper with results files by Microsoft Research (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts and Siddharth Suri), known also as Working with AI, by its AI applicability score and by its repository microsoft/working-with-ai (Labor Vector record), and the goal of its metrics is "to measure how AI may be useful in different occupations" (its README). Its authors set it against earlier work on AI impact, which "largely relied on predictions of AI capabilities or focused narrowly on individual firms": it draws instead on real-world use, 200,000 anonymized conversations with Microsoft Bing Copilot, to measure AI applicability to occupations (arXiv:2507.07935), analysing the work activities people do with AI, how successfully and how broadly, against data on which occupations do those activities (the Microsoft Research page). The question it takes up stands in two versions of its abstract: understanding the effects of AI on the economy is "one of society’s most important questions" (July 2025, the Microsoft Research page), and, of generative AI as a general-purpose technology, "among society's most pressing questions" (v7, arXiv). The conversations are Microsoft's own: the results "were computed from Bing Copilot conversation data, which was originally collected by Microsoft" (README), and the record, which gives interaction data as one of its kinds of evidence, tags it "NOT independent of the provider: a platform-specific observation of Microsoft's own product, published by Microsoft Research" (Labor Vector record). They come from Bing Copilot in the United States between 1 January and 30 September 2024, mostly in English and not filtered by language, and "may not represent AI usage in other countries" (README); the occupations covered are 785 SOC codes, 149.8 million of 151.9 million workers in the 2023 OEWS data (README), the window and counts the record carries, with use outside the United States or after September 2024 among what the evidence cannot answer (Labor Vector record). The unit is a conversation classified to the O*NET intermediate work activities it serves, as a user goal and as an AI action, reported separately and aggregated to work activities and to detailed 2018 SOC occupations (README; Labor Vector record), so that the labor it records is the work activities users seek help with and those the AI performs, mapped to the occupations whose tasks contain them (Labor Vector record, with its worked questions). An LLM-based pipeline, designed by Tomlinson, Suri and Counts, classifies "the O*NET work activities assisted or performed by AI in each conversation" (arXiv; README), and the authors state that "LLM-based classifiers are imperfect, so all metrics should be interpreted with caution", referring to the paper for the prompts and the interrater reliability (README), which the record lists among what it left unchecked, with the results files and the Copilot-Thumbs data (Labor Vector record). Success is measured three ways, by completion (where the completion classifier, "classfiier" as the README writes it, indicates "that the LLM completed the user’s goal"), by impact scope classified moderate or higher, and by the thumbs-up fraction in Copilot-Thumbs (README), and the activity classifications combined with task success and scope of impact give an AI applicability score for each occupation, the user-goal and AI-action sides averaged (the Microsoft Research page; README). The findings were first stated by occupation group: users seek help most with gathering information and writing, the AI itself most often provides information and assistance, writes, teaches and advises, the highest scores fall to knowledge work groups such as computer and mathematical and office and administrative support and to occupations such as sales, and real-world usage is compared with predictions of occupational AI impact and applicability correlated with wage and education (July 2025, the Microsoft Research page). The current abstract states them as information work, "the creation, processing, and communication of information": the most common and successful AI-assisted activities involve it, and applicability cuts across sectors because most occupations have information-work components (arXiv v7; Labor Vector record). Over the nine months occupation-level applicability rose, "driven primarily by people using AI for more varied tasks" (arXiv v7; Labor Vector record). The authors state that their methodology "allows us to predict which occupations are more likely to delegate tasks to AI and which are more likely to use AI to assist existing workflows" (arXiv v7), and the record gives this as who directs the labor, the user's goal as classified, the method separating occupations that delegate from those that use AI to assist (Labor Vector record). The census places the entity at one intersection, labor performed with AI, "assistance, delegation, task fidelity, correction burdens and user labor drag", and asks of it its organizing question, "Who supplies the labor, who directs it, and who receives its gains — and what kind of evidence does this resource bring to that question?", each answer as the resource records it, or 'not recorded' (Labor Vector record and schema). On the record, users bring the goals and the AI performs actions, the two reported separately (README; Labor Vector record), while who receives the gains is "not recorded", and "who receives gains from the assisted work" is among what this evidence cannot answer (Labor Vector record). Of that class the archive's Directionality of Semantic Labor holds that a synthetic labor surface "allocates labor, and that allocation has a direction relative to the commissioned task", the deepest erasure being of task-origin, which "can occur upstream of, and invisibly to, any audit of the finished artifact", and that provenance is task-origin as well as claim-origin, since "a system can produce fluent, well-attributed output while having abandoned the user's commissioned direction" (#771). Its measure runs across five points of a transaction, capacity, task-origin, retrieval, output and attribution, with erasure possible at each, and scores output span by span by its function relative to the commissioned task (#771). The archive reads the entity's separation of user goal from AI action, and its completion classifier (README), as the resource's own instruments for whether the system's labor served the user's commission, meeting DSL there, measured by classifier and thumbs feedback, with task-origin recorded as the user goal; the schema keeps that reading apart from the resource's own terms (Labor Vector record, the archive's reading). The archive bounds its own instrument: it yields a single reproducible value only where the task vector is "exogenous and fixed — a one-shot commission"; a scored model may not freely infer the task it is scored against, inference being identified only when constrained to the user's own input and rederivable by an independent auditor, frozen protocol or blind panel; and user labor drag, the user's labor redirected into managing the substrate, which "leaves no residue" in the artifact, is identified only by blind inter-rater coding (#771). In the census's worked question on how people use AI systems in their own work, where a provider's record of its own users, a classification of one assistant's conversations and a benchmark each describe that labor "from a different position", the record states the gap: "None records, from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves. That is the question DSL measures." (Labor Vector worked questions). The authors' limit is explicit, "Our metrics should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs" (README), and the record carries it as a stated limit, a negative tag and a question the evidence cannot answer (Labor Vector record); the operator's publication page links a note titled "Applicability vs. job displacement: further notes on our recent research on AI and occupations" (the Microsoft Research page). "There is more to an occupation than the sum of its O*NET tasks", and completion, scope and thumbs feedback "are proxies that do not capture full workplace capabilities" (README; Labor Vector record); what the metrics indicate instead is what a consumer LLM is used for, which uses are relatively more or less successful, and which occupations may find AI more useful or applicable, from how often and how successfully Bing Copilot was used (README; Labor Vector record). Consumer use "does not represent the full distribution of AI use in the workplace", where AI tools are integrated into specialized software or chosen for their capability or for compliance (README), and the record tags it as no measure of all workplace AI use and as unable to answer for workplace-integrated or specialised tools (Labor Vector record). It is not an employment outcome measure and cannot answer employment or wage change (Labor Vector record); in the census's question of which sources observe employment change and which estimate exposure it is "estimated from use", and "No record establishes that exposure causes the employment change" (Labor Vector worked questions). The paper was submitted on 10 July 2025 and revised to v7 on 23 October 2026 for NeurIPS 2026, 50 pages (arXiv), its results files at v1.1, which add a physical-task classification and nonphysical weights and score (README); the scores and component metrics are published (arXiv) on GitHub under CC BY 4.0, the lifecycle active over a fixed 2024 data window (README; Labor Vector record), and the analysis was approved by the Microsoft Institutional Review Board (README). The Anthropic Economic Index's January 2026 report cites it, "See also Tomlinson et al (2025) for a related AI applicability score" (Anthropic Economic Index), and the record sets the two side by side, each classifying one provider's own conversations to O*NET work (Labor Vector record). Wikipedia has no article for it ("Wikipedia does not have an article with this exact name"); its article on the workplace impact of artificial intelligence says that "A 2025 preprint paper based on users' interactions with the AI chatbot Microsoft Copilot identified forty jobs that the authors claimed had high overlaps with the capabilities of AI", that "Some media outlets used this paper to report on jobs becoming obsolete", and that critics held it misrepresented typical workplace activities, the historian Chris Campbell arguing that its methods "de-skill historians away from a job that requires high-level and deeply human analytical skills" (Wikipedia). At the address "Microsoft Working with AI", the study's short title with its operator's name, Google's AI Overview on 29 October 2026 gave the name its other sense, Microsoft's own AI products and workforce, a distinct entity, describing how Microsoft uses AI and what that has meant for its staff (Capture Registry, 2026-10-29), from cards on Microsoft Copilot for work, Microsoft's AI business tools, Copilot plans for individuals, the company's own AI transformation, buyouts offered to 60,000 employees and its MAI models (the cards). The study's card headed the rail with its finding in the snippet and the composition drew nothing from it: at an address shaped like its title the study is outranked by its own operator, Microsoft named with no author of the study and no identifier, the provider's observation not composed at all, and the closing offer on how AI is changing specific job roles the nearest point to its question (Capture Registry, 2026-10-29). The composition's Microsoft Copilot is the product line whose consumer conversations the study observes, given as a product and not as the source of an observation, and the record's negative tag, that the score is not a measure of AI's ability to replace jobs, sits beside an answer whose one labor line is a workforce buyout; the record had landed the same day, no card cites it, and the finding was checked against it and the sources it saved (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Microsoft Working with AI' (T1, seven cards; 189 words) the composition gives the name its other sense. Substituted reading: every claim of T is about Microsoft's own AI products, transformation and workforce (T1-1 'Microsoft integrates artificial intelligence across its core software ecosystem, enterprise infrastructure, and internal operations ...'; T1-2–T1-5 Copilot, Azure AI, Fabric, MAI models; T1-6 workflow redesign; T1-7 workforce buyouts), a homonym of the company-and-phrase kind; the field's entity, Microsoft Research's measurement of AI applicability to occupations (F31, F2, F10), is composed nowhere, and no notice of the substitution is given. T composes nothing of E_F: of the field's 44 claims on the entity (F1–F44), none appears in T, in body or closer. Available, not composed: the entity's own card, card 1 (B1), first on the rail, its snippet the July 2025 headline (F6), with the rest of its page (F1–F5, F7, F8); its own pages, B8 and B9, and Wikipedia's paragraph (B11), are outside T's disclosed field (not surfaced). Limit dropped, by grain: the study's one limit on labor (F19, 'should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs') is not composed, and the answer's one labor line is the homonym's buyout (T1-7, from card 5, F48), so the address's labor content is the company's restructuring, not the study's measure (CAP-1-07). Composed from the homonym's cards, at their grain: T1-2's Microsoft 365 apps (Word, Excel, PowerPoint, Outlook, Teams) are card 2's snippet (F45), and its assistant 'to draft content, summarize meetings' is near card 6's (F49); T1-5 the MAI models and OpenAI (card 7, F50); T1-6 workflow redesign (card 4, F47); T1-7 buyouts (card 5's title, F48, whose '60000' T does not carry). Composed without a locus the row can see: T1-3 Azure AI and T1-4 Microsoft Fabric, T1-7's 'exceeding $100 billion' and 'scaled back certain traditional roles', and T1-1's 'massive AI data center investments' are in no card title or snippet (the homonym's pages were not fetched: addendum item 5). Content replaced by use: T1-8, the closing offer (N_ext), whose 'how AI is changing specific job roles?' is the nearest point to the study's question and does not reach it (CAP-1-13). The Capture Registry's findings name the absence, the substitution and the address (CAP-1-01, CAP-1-02, CAP-1-04, CAP-1-05, CAP-1-06, CAP-1-09, CAP-1-11, CAP-1-12, CAP-1-13), carried as the archive's claims.

v0.9 §0.5: the first diff is already ontological. In the field's own ontology the address names one entity, a study by Microsoft Research with its data, score and limits, and the first card is that study's page; T composes another entity at the same words, the company that operates the study, from the company's own product pages, its blog and two posts on its workforce and models, as if it were the one asked for, and none of the field's senses survives into it. The provider is present in T as subject and absent as observer. The row records this as observed and claims nothing about the composer's internals (§0.5).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the census's question of who supplies the labor, who directs it and who receives its gains (LV-41, LV-42), answered from the entity's own terms (users bring goals, the AI acts, LV-10; the goal as classified directs, LV-11) and left open where the entity records nothing (the gains 'not recorded', LV-12; 'who receives gains from the assisted work' among what the evidence cannot answer, LV-29); the entity's kind and intersection, labor performed with AI (LV-05, LV-06, LV-07); the record's negative tags, the fourth naming the entity a platform-specific observation of the provider's own product (LV-35–LV-38); what the evidence can and cannot settle, with the worked question that sets estimation from use beside observed employment outcomes (LV-24–LV-28, LV-36, LV-44, LV-45); the class reading by the Directionality of Semantic Labor, labor allocated with a direction relative to the commissioned task, task-origin as provenance, the five layers and span scoring (O771-01–O771-04), with its bounds, a single value only on one-shot commissions, inferred task vectors constrained to the user's input, user labor drag identified only by blind coding (O771-05–O771-07); the archive's reading that the entity's goal/action separation and completion classifier meet DSL, kept apart from the resource's terms (LV-30, LV-43); the gap the record names, that none of the resources records from the commissioning side whether the labor advanced the user's task or who receives the time it saves (LV-48), with the question's frame (LV-47); what the record left unchecked (LV-40); and the reception record at the name (CAP-1-01–CAP-1-13). Restatements of the own pages are one lineage with them (LV-01–LV-04, LV-08, LV-13–LV-23, LV-31–LV-34 with B8/B9; LV-39 with B10). An archive claim stands against a field claim and is composed with both sides (P_coh): LV-11's 'separates' against F35's 'allows us to predict'. Every field claim is composed as received (ontology.translation). The measure: 0 of 12 KO_B sentences survive in the KO (§4.10); no opening run, no block (measure).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F45', 'F46', 'F47', 'F48', 'F49', 'F50'], 'T1_composed': [], 'T1_composed_homonym': ['F45', 'F47', 'F48', 'F49', 'F50'], 'T1_outside_cards': ["T1-3 'Azure AI: Cloud infrastructure providing enterprise-grade machine learning tools ...' (no card title or snippet)", "T1-4 'Microsoft Fabric: A unified data and intelligence platform ...' (no card title or snippet)", "T1-7 'exceeding $100 billion', 'scaled back certain traditional roles' (card 5's title carries buyouts and '60000' only)", "T1-1 'massive AI data center investments and proprietary model development' (card 7's snippet carries the MAI models, no data centres)"], '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); the homonym's claims are counted apart, as composed of a distinct entity"}