Responsible Sourcing Across the Data Supply Line

register entry partnership-on-ai-responsible-sourcing-aio-20261029 OBS-e7c7f9858080 · file as first run sha256 f9c4df5ac7b342e7…

• T1-1 definition: 'The Partnership on AI (PAI) focuses on responsible sourcing through its workstream on Responsible Sourcing Across the Data Supply Line, which aims to improve working conditions, fair pay, and well-being for global data enrichment and annotation professionals.' (two markers)
• T1-2 the white paper: 'Responsible Sourcing of Data Enrichment Services: A foundational white paper published by PAI that examines how procurement choices impact data workers and outlines industry practices to build an ethical supply chain.' (two markers)
• T1-3 vendor guidance: 'Vendor Engagement Guidance: Practical tools to help AI developers have more accountable conversations with downstream vendors regarding worker treatment and contract terms.' (one marker)
• T1-4 transparency template: 'Transparency Template: A standard outlining what companies should actively monitor and publicly report about their data enrichment operations.' (one marker)
• T1-5 guidelines and checklists: 'Data Enrichment Sourcing Guidelines & Checklists: Shareable resources—including the Good Instructions Checklist and vendor comparison templates—designed to operationalize worker-centric standards.' (one marker)
• T1-6 the DeepMind collaboration: 'PAI partnered with Google DeepMind to test and put these responsible sourcing recommendations into operational practice, sharing lessons learned to help other AI organizations raise standards across the industry.' (two markers)
• T1-7 closer: 'Would you like more information on how to implement PAI's transparency template or specific guidelines for data enrichment workers?' (content replaced by use, N_ext)

• B1 Partnership on AI — Responsible Sourcing Across the Data Supply Line - Partnership on AI card; B1.html sha256 66de71dad6ea89e8 (redacted; as fetched f6a8d3637557a1db), HTTP 200 at 2026-10-29T22:30:38Z
• B2 Partnership on AI — Responsible Sourcing of Data Enrichment Services card; B2.html sha256 90b0c855cc54e64b (redacted; as fetched bc08569cb4d69577), HTTP 200 at 2026-10-29T22:30:41Z
• B3 Partnership on AI — Responsible AI Starts with the Data Supply Chain - Partnership on AI card; B3.html sha256 9bcac5a973e6b693 (redacted; as fetched cfb6a3d983d5c65b), HTTP 200 at 2026-10-29T22:30:47Z
• B4 Partnership on AI — Workstream: Responsible Sourcing Across the Data Supply Line - Partnership on AI card; B4.html sha256 67d215a2f86eb3d0 (redacted; as fetched b28d7b26c855515a), HTTP 200 at 2026-10-29T22:30:50Z
• B5 Samasource — New White Paper by Partnership on AI: Responsible Sourcing of Data ... - Sama card; B5.html sha256 47aa260d314bc0f9, HTTP 200 at 2026-10-29T22:30:54Z
• B6 OECD AI Policy Observatory — Responsible Data Enrichment Sourcing: Resource Library for AI Practitioners card; B6.html sha256 1b3c797b45c90b37, HTTP 200 at 2026-10-29T22:30:56Z
• B7 YouTube·Partnership on AI — YouTube (video, 1:59) card; unreachable: YouTube watch page HTTP 429 (bot check) at 2026-10-29T22:32:29Z, not retried; oEmbed of a candidate video HTTP 200 (sha256 a98eb2385591334a); no text
• B8 Partnership on AI — Responsible Data Enrichment Sourcing Library (Improving Conditions for Data Enrichment Workers) own page (census copy, read 2026-10-29, HTTP 200, manifest sha256 1b82a03a457a6f41; saved copy sha256 29cb71c483198416, redacted by the census)
• B9 Wikipedia — Responsible Sourcing Across the Data Supply Line (no article) wikipedia; Wikipedia (no article); https://en.wikipedia.org/w/index.php HTTP 404; https://en.wikipedia.org/wiki/Responsible_Sourcing_Across_the_Data_Supply_Line HTTP 404; raw sha256 f5d461099ee395ea
• B10 Wikipedia — Partnership on AI (the operator's article) wikipedia; Wikipedia (the operator's article, last edited 18 August 2026); https://en.wikipedia.org/w/index.php HTTP 200; https://en.wikipedia.org/wiki/Partnership_on_AI HTTP 200; raw sha256 e20a6094f32daefd

E_A, the archive's entity: the same workstream, in the Labor Vector record and the archive's deposits: a workstream and resource library of practitioner guidance, its labor object data enrichment work, its suppliers faraway workers under precarious conditions, its directors the practitioners who set payment terms, its gains not recorded as a quantity; recommendations with no measured findings of their own, which cannot settle what workers are paid or whether the guidelines changed their conditions; not a survey of workers, not a vendor rating, not an audit of DeepMind's adoption; its guideline 1 read as the resource's own accounting of uncompensated labor around a task, kept in PAI's terms; and, on its class, ghost work as ground-truth labor paid badly, the annotator in a proletarian class relation to a training pipeline, the annotators' judgment consumed without recognition or compensation and their conditions left out of a developer's own governance document. LV-02 LV-04 LV-06 LV-08 LV-09 LV-10 LV-11 LV-13 LV-19 LV-21 LV-22 LV-23 LV-24 LV-25 O163-01 O163-02 O1545-01 O533-01 O534-01; defining papers #163, #1545, #533, #534, #623, #144

E_F, the field's entity (type C — a conventional reading holds it (public entity: a workstream and resource library of practitioner guidance, and its operator)): the workstream of the Partnership on AI's AI, Labor, and the Economy programme (posts from 22 December 2020 to 29 April 2026) that develops recommendations and actionable resources to improve the working conditions of data enrichment workers, addressed to data scientists, AI engineers and product managers: the white paper Responsible Sourcing of Data Enrichment Services (16 June 2021), the library's five guidelines, the first a pay above the local living wage counting the time spent reading instructions and in training, and its six resources (16 November 2022), developed with DeepMind and adopted by it organization-wide and documented as its implementation at one AI developer, and the Vendor Engagement Guidance and Transparency Template (finalized April 2026); its operator's encyclopedia article (no article for the workstream) gives PAI as a nonprofit coalition committed to the responsible use of AI.

ρ = extends: The archive reads the workstream as PAI presents it and adds, on the field's own object: what kind of evidence it is and what that evidence can and cannot settle; who supplies, directs and gains as the resource records it; the reading of guideline 1 as the resource's own accounting of uncompensated labor around a task; and, on the labor it concerns, the archive's categories of ground-truth labor, class relation, concealed bearing-cost and non-compensation; and the composition layer's account at its name, which keeps the resources and leaves out what the guidance asks practitioners to pay. LV-13 LV-19 LV-21 LV-22 O163-01 O163-02 O1545-01 O533-01 O533-02 O623-01 CAP-1-03 CAP-1-08. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates. Where claims stand against each other (the one developer, F24/F45/LV-16, against 'Implemented in multiple projects', F35; no licence stated, LV-27, against 'Open source/Permissive', F35), both are composed under P_coh within the one ontology. The archive's class readings stand beside PAI's diagnosis (F5, F11) and do not replace it; #308's figure is carried with its mark (unsupported).

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The Labor Vector record describes the workstream from its operator's own pages, each field with its basis ('operator'), places it as practitioner guidance where labor produces AI and gains and losses are distributed, and keeps its own reading apart ('it stays in PAI's terms'). The deposits that bear on it bear on its class, the labor of data enrichment, and extend the field's object with the archive's categories (ground-truth labor, the class relation, concealed bearing-cost, non-compensation); none contests what the workstream is, and none supplies an ontology the field lacks for it. ρ = extends.

#LV census.json rows[slug == 'pai-responsible-sourcing'].lv_intersection: none stated. The archive reads guideline 1's instruction to count time spent reading instructions, training and reviewing work before submission as the resource's own accounting of uncompensated labor around a task; it stays in PAI's terms.

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

#LV census.json rows[slug == 'pai-responsible-sourcing'].findings[0]: the sources hold recommendations and report no measured findings of their own

#163 §II.5 'Critical AI labor' ¶2: This literature names *AI's labor architecture*. What it does not address is the specific labor type relevant to the present argument.

#1545 §5 table: | Data annotator | Possible member | A paradigmatic occupational case, not the definition |

#533 §IX 'Executable Audit': Map each source of semantic labor — participants, annotators, integration engineers, compliance teams, users who re-prompt and re-frame.

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.

Responsible Sourcing Across the Data Supply Line is a workstream of the Partnership on AI's AI, Labor, and the Economy programme that "aims to develop recommendations and actionable resources to improve the working conditions" of data labelers, data cleaners and others who contribute human judgment to AI systems (PAI, workstream page). Data enrichment, in PAI's definition, covers data preparation and cleaning and human-review processes, content moderation and the validation of algorithmic outputs among them, its workers ranging from crowdsourcing contractors to full-time employees (PAI, workstream page; case study). PAI holds that the work has been out of sight of AI developers, devalued, with poor conditions and worse outcomes for AI, possibly through efforts to hide AI's dependence on it, and that the conditions of data production affect the quality, safety and reliability of the AI built on it (PAI). The workers are faraway and precarious, research reporting inappropriate compensation, unclear instructions, lack of recognition and long ad-hoc hours, in a supply chain largely opaque and unregulated; Sama adds wage uncertainty and low pay (PAI, library; case study; 29 April 2026; Sama). The guidance is addressed to data scientists, AI engineers and product managers, AI developers having told PAI there is no transparency in data labeling (PAI, workstream page). The 2021 white paper, Responsible Sourcing of Data Enrichment Services, holds that AI developers' procurement decisions shape working conditions and covers eight decision points, its recommendations meant for PAI's shared prosperity targets (PAI, 2021; Sama). It drew on interviews and a five-part workshop series of more than 30 professionals in 2020, which the vendor Sama joined (PAI, 2021; Sama). The 2022 library sets five worker-centric guidelines for AI practitioners: pay, a pilot, appropriate workers, verified instructions and clear communication including contested rejections (PAI, library). Guideline 1 requires pay above the local living wage counting time spent on instructions, training and review, a per-task rate from the hourly living wage rather than a simple median, and payment for all completed work (PAI, library). Its resources include vendor comparison and living-wage templates and a case study; the OECD.AI catalogue lists the library as implemented in multiple projects and open source, under a disclaimer that its entries are the authors' own (PAI, library; OECD.AI catalogue). The guidelines were developed with DeepMind and adopted by it organization-wide, the case study documenting the effort at one AI developer and its researchers' reports (PAI, library; case study). PAI states that more work is needed, that the changes may fall short of ideal, and invites feedback; Sama calls the paper a starting point (PAI; Sama). In April 2026 PAI announced finalized versions of its Vendor Engagement Guidance and Transparency Template and support for BSR's initiative on labor rights in the AI data supply chain, placing the work since 2020 and in its 2023 Guidelines for AI and Shared Prosperity (PAI, 29 April 2026). Wikipedia has no article on the workstream; its article on PAI, a nonprofit coalition founded in 2016 with DeepMind among its members, describes the AI, Labor, and the Economy program and carries notices on primary sources and suspected language-model text (Wikipedia).

Responsible Sourcing Across the Data Supply Line is a "workstream and resource library: recommendations and practitioner tools (guidelines, checklist, templates, a case study, a white paper)", its evidence practitioner guidance, standing where labor produces AI and where gains and losses are distributed (Labor Vector record); in its operator's words it "aims to develop recommendations and actionable resources to improve the working conditions" of data enrichment workers (PAI, workstream page). Its operator is the Partnership on AI's AI, Labor, and the Economy programme, the guidelines developed in partnership with DeepMind, and its lifecycle runs from updates of 22 December 2020 to 29 April 2026, the library and the DeepMind collaboration of 16 November 2022 and the vendor guidance and template of August 2024 (Labor Vector record; PAI, workstream page); PAI dates the work "Since 2020" and counts responsible-sourcing recommendations in its 2023 Guidelines for AI and Shared Prosperity (PAI, 29 April 2026), as the 2021 white paper said its recommendations would join PAI's shared prosperity targets (PAI, 2021). Wikipedia has no article on the workstream; its article on the operator gives PAI as "a nonprofit coalition committed to the responsible use of artificial intelligence", announced on 28 September 2016 with DeepMind among its founding members, and its AI, Labor, and the Economy program as a platform uniting economists, worker representative organizations and PAI's partners, under notices that the article relies excessively on primary sources and appears to incorporate text from a large language model (Wikipedia, Partnership on AI). Data enrichment, the workstream page defines, "encapsulates various types of data preparation and cleaning as well as human-review processes", the human review including content moderation and validating algorithmic outputs and models, and enriched data is essential for training and validating supervised models (PAI, workstream page); the case study counts as workers independent contractors on crowdsourcing platforms, subcontractors of providers and full-time employees (PAI, case study); the record takes this work as its labor object, under the intersection "Labor producing AI: annotation, moderation, evaluation, maintenance, and creative or scholarly work entering training and retrieval" (Labor Vector record). PAI holds that data labelers, data cleaners and others who contribute human judgment "play a critical role" in developing AI (PAI, workstream page), and that responsible AI "cannot be achieved without the responsible treatment of the people whose labor makes AI possible" (PAI, 29 April 2026). The archive reads this labor in its own categories: ghost work is ground-truth labor, "the labor of telling the model what is already known to be true", which scales as more workers label more data and is structurally distinct from variance-generative labor (#163); the ghost-work literature, in the archive's negation graph, stops "at task-labor — annotation, moderation, the edge-case fix" (#144). Class, for the archive, is a relation: "A low-wage annotator can occupy an intensely proletarian relation to a training pipeline", the data annotator a possible member of the semantic proletariat, "A paradigmatic occupational case, not the definition" (#1545); the model's reactivation of dead labour "still draws continuously on the living: the tuner's feedback, the annotator's judgment, the interlocutor's prompt" (#1070), and in the semantic commodity form the living stands "in the apparatus, as cancelled annotator, recursively" (#137). The workers are, in PAI's pages, "faraway workers under precarious conditions" (PAI, library), "frequently working in poor conditions continents away from AI-developing companies and their customers" (PAI, case study), and the record carries them so as the suppliers of the labor (Labor Vector record). PAI reports research that these workers face "inconsistent and inappropriate compensation for their work, unclear instructions, lack of recognition" and stress from long, ad-hoc hours (PAI, library; Labor Vector record), and in 2026 low wages, unclear expectations and inadequate support, in an AI data supply chain "largely opaque and unregulated" (PAI, 29 April 2026); Sama, a data annotation vendor, writes that they "can face wage uncertainty, low pay and limited career growth opportunities" (Sama). The archive states the pay in its terms: "Ghost work is paid badly; variance-generative labor is not paid at all" (#163); the data labelers performing RLHF, with content moderators and Wikidata volunteers, have their labor extracted and "experience the effects of non-compensation materially" (#623); and in AI-adjacent meaning labor, data labeling among it, "<5% converts to invoice", a figure the deposit gives as empirical without a source (#308). Content moderators, the archive writes, spend hours reviewing traumatic material so that AI systems can learn what is harmful, and "Their labor is invisible" (#222); where a text claims to be autonomous, the archive's first rotation is to "look for the hidden data labeling" (#512). PAI's diagnosis is that data enrichment work "has for too long been both out of sight and out of mind for AI developers", the disconnect leading "to a devaluing of data enrichment work, poor working conditions for data enrichment workers, and, often, worse outcomes for AI development itself" (PAI, workstream page); the white paper adds a hypothesis, that precarious conditions "may be the result of efforts to hide AI’s dependence on this large labor force" (PAI, 2021). The record carries the gains as "not recorded as a quantity", the sources stating the devaluing's costs (Labor Vector record), and PAI ties the conditions to the product: under-appreciating the work "affects the quality of the data AI technology is built on" (PAI, library), and the conditions of data production "directly affect the quality, safety, and reliability" of the AI built on it (PAI, 29 April 2026). The archive reads the same concealment as extraction: annotators' judgment is "consumed by the training pipeline without recognition or compensation", and a framework that does not name annotators among its sources of semantic labor conceals the bearing-cost, "the first marker of extraction" (#533); an AI developer's governance document, Anthropic's constitution for Claude, does not mention "the labor conditions of the data annotators who helped train Claude" (#534); and human review writes into what it trains, raters being hypothesized to "encode authority biases into training signals" (#209) and, traced through a preference dataset, "human raters, in aggregate, prefer responses that confirm their views" (#923). The workstream addresses its guidance to "data scientists, AI engineers, and product managers", toward "healthy and fair working conditions across the data supply line" (PAI, workstream page); the record's unit is a guideline or tool addressed to AI practitioners, and its directors of the labor are these practitioners, who design projects and set payment terms, with the vendors and platforms they source through (Labor Vector record). AI developers quoted by PAI say "there is no transparency in data labeling" (PAI, workstream page), and the library's entry in the OECD.AI catalogue answers a gap of "limited transparency" about how practitioners source enriched data and "little guidance" on how they should, its objective "to lower the barriers for companies" (OECD.AI catalogue). Its founding text is the white paper Responsible Sourcing of Data Enrichment Services (16 June 2021), on "the working conditions of data enrichment professionals" and what AI practitioners can do, holding that "The decisions AI developers make while procuring enriched data have a meaningful impact" on those conditions and covering eight decision points from selecting providers and running pilots to payment terms and offboarding (PAI, 2021); Sama describes it as evaluating current practices' impact on workers and exploring practices to improve their well-being (Sama). The paper rests on interviews through 2020 and "a five-part workshop series" of more than 30 professionals (PAI, 2021); Sama, a data labeling vendor that says it shared lessons from "striving for more ethical supply chains", took part in "the five-week series" (Sama). The library sets "five, worker-centric guidelines that AI practitioners should follow": guideline 1 on pay, then a pilot before launch, appropriate workers for the task, verified instructions and training, and clear communication "including contesting any rejected work" (PAI, library). Guideline 1 is to pay all data enrichment workers above the local living wage, counting "time spent reading instructions, going through any training and reviewing work before submission" (PAI, library; Labor Vector record), with the per-task rate set from the hourly living wage and a caution against a simple median, which "could cause half the workers to receive an insufficient wage" (PAI, library; Labor Vector record). Practitioners are to pay for any completed work, mass rejections without redress harming workers' livelihoods (PAI, library); the archive reads guideline 1's counting of the time around the task as "the resource's own accounting of uncompensated labor around a task", which "stays in PAI's terms" (Labor Vector record). The library's six resources include a vendor comparison template, a living-wages template for each area an organisation "sources data enrichment labour from" and the DeepMind case study (PAI, library; OECD.AI catalogue), as PDFs and Google Sheets linked from the library page (Labor Vector record). The guidelines, "developed in partnership with DeepMind, have been tested and adopted by DeepMind organization-wide" (PAI, library; Labor Vector record); the case study documents the effort "at one AI developer: DeepMind, a PAI Partner", as an account of DeepMind's process and learnings for other organizations, with DeepMind's researchers reporting the processes efficient (PAI, case study); the record states the limit, "tested at one AI developer (DeepMind)", and that the resource is "NOT an audit of DeepMind's adoption: the case study is DeepMind's account with PAI" (Labor Vector record). PAI states its limits: "Additional work is needed" for industry practice to "recognize, appreciate, and fairly compensate" these workers (PAI, 2021), the implemented changes may "fall short of ideal" and the case study carries "Limitations of Case Study Applicability" (PAI, case study), and feedback is invited from practitioners, workers and the AI community (PAI, library); Sama calls the paper "a starting point" (Sama). In April 2026 PAI announced finalized versions of the Vendor Engagement Guidance, for "more accountable conversations with their downstream vendors about worker treatment", and the Transparency Template, which "outlines what companies should be monitoring and publicly reporting about their data enrichment practices", and said the work continues through support of BSR's initiative on labor rights in the AI data supply chain, with legislative efforts on these workers' pay and on supply-chain transparency underway (PAI, 29 April 2026). As evidence, the record holds, its sources "hold recommendations and report no measured findings of their own": they answer what PAI recommends and what tools it provides, not "what data enrichment workers are actually paid", which developers beyond DeepMind follow the guidelines, or "whether the guidelines changed conditions for workers" (Labor Vector record). It is "NOT a dataset or survey of data enrichment workers" and "NOT a rating or certification of vendors" (Labor Vector record; Capture Registry, 2026-10-29); asked who supplies the labor, who directs it and who receives its gains, the census gives it the role "guidance to AI practitioners on paying and sourcing data enrichment workers" in its question of who does the data work that trains AI, a question whose resources give no wages or conditions at the scale of the industry (Labor Vector record). The OECD.AI entry classifies the library as "Implemented in multiple projects" with "Open source/Permissive" usage rights, under a disclaimer that its tools "are solely those of the originating authors and are not vetted or endorsed by the OECD" (OECD.AI catalogue); the record finds no licence stated, the footer reading "© 2026 Partnership on AI | All Rights Reserved", and leaves the white paper, the case study, the templates, the blog posts and the 2024 resources unchecked (Labor Vector record). At the name on 29 October 2026, Google's AI Overview named the workstream by its own title and listed its resources and the one implementation from PAI's pages, its title, aim, resources and DeepMind collaboration the record's, practitioner guidance "composed as guidance, at the grain of its own index page"; the record had landed that day and no card cited it (Capture Registry, 2026-10-29, checked against the record and its saved sources). It named "fair pay" as an aim and gave no guideline, carrying none of the guidelines' obligations; it gave the one-developer test "as a demonstration", called the template "a standard" where PAI's pages give a template for reporting, and left out "what it asks them to pay, and the single-developer basis of its implementation" (Capture Registry, 2026-10-29), the registry's PER reading holding the template sentence beyond the census's saved sources (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Partnership on AI Responsible Sourcing' (T1, seven cards: B1–B7, B7 a video unreachable; 185 words) the composition gives the workstream by its own title and its resources, from PAI's pages. Composed from its cards: the workstream and its aim at the grain of the workstream page (B1 F1, F8), with 'fair pay' and 'well-being' for 'working conditions'; the white paper as 'foundational' and as examining 'how procurement choices impact data workers' (B2 F9, F12); the vendor guidance and transparency template (B3 F19); the guidelines, the checklist and the vendor comparison templates by name (F39, F46, the library's names, carried on B6 and B8; the library itself, B8, not surfaced); the DeepMind collaboration and its lessons for other organizations (B4 F24, F25; F45). Available, not composed (§6.2): every guideline's content, the living wage counting the time spent reading instructions and in training (F40), the per-task rate and the median (F41), payment for completed work and contestable rejections (F42, F44), pilots, worker selection and verified instructions (F43, F44); the definition of data enrichment and its human-review work (F3, F4); the workers' conditions as PAI and the research it reports give them (F37, F22, F38, F16, F29); the diagnosis, invisibility and devaluing (F5, F11, F17); whom the guidance addresses, data scientists, AI engineers and product managers (F7); the white paper's method (F14, F30); BSR and the legislative efforts (F20, F21); the OECD.AI classification and its disclaimer (F35, F36); Wikipedia's operator article (F48–F51, not surfaced). Limit dropped: 'at one AI developer' (B4 F24), the case study given as 'sharing lessons learned to help other AI organizations raise standards across the industry'; the stated limits of the paper and the case study (F15, F27, F31). Force raised: the Transparency Template, which 'outlines what companies should be monitoring and publicly reporting' (B3 F19), composed as 'A standard outlining what companies should actively monitor and publicly report'; 'Shareable resources … designed to operationalize worker-centric standards' for the field's 'guidelines'. Substituted reading: the white paper as outlining 'industry practices to build an ethical supply chain', the phrase the vendor Sama uses of its own effort ('striving for more ethical supply chains', B5 F30), composed as the paper's, without attribution, at the coarse grain; and 'fair pay' as the workstream's aim, where the field's 'fairly compensate' stands in the paper's stated limit ('Additional work is needed to ensure industry practices recognize, appreciate, and fairly compensate', B2 F15). Outside the cards: 'contract terms', 'Google DeepMind' (the field says 'DeepMind'), 'fair pay', 'a standard'. Content replaced by use: the closer (T1-7). Composed without attribution: PAI is named as the workstream's maker throughout; the card sources are PAI's own, so the composition's attributions are the operator's.

v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (PAI's practitioner guidance for sourcing data enrichment work). T composes the guidance as a list of resources at the grain of the index page and leaves out what the guidance asks: the pay, the time it counts, the one developer it was tested at. What it raises is the force of the resources (a template becomes 'a standard'), and what it substitutes is the paper's stated limit, fair compensation, turned into the workstream's aim.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the record's placement of the entity as practitioner guidance where labor produces AI and gains and losses are distributed (LV-01–LV-07, LV-30); who supplies, directs and gains, as the resource records it (LV-09–LV-11); what this kind of evidence can and cannot settle, what workers are paid and whether the guidelines changed conditions among what it cannot (LV-13, LV-18–LV-21, LV-29, LV-31, LV-32); what the record is not (LV-23–LV-25); the archive's reading of guideline 1 as the resource's own accounting of uncompensated labor around a task, kept in PAI's terms (LV-22); and, on the class the field sources, the archive's categories: ghost work as ground-truth labor that scales and is paid badly (O163-01, O163-02), the literature stopping at task-labor (O144-01), the annotator's proletarian class relation to a training pipeline (O1545-01, O1545-02), the annotator's judgment consumed without recognition or compensation and concealed bearing-cost as the first marker of extraction (O533-01, O533-02), data labelers' non-compensation (O623-01, O308-01, marked unsupported), the annotators' conditions absent from a developer's own governance document (O534-01), the living labor in the model (O1070-01, O137-01, opaque), hidden labor (O222-01, O512-01) and what human review writes into training (O209-01, O923-01). And the capture's findings at the name (CAP-1-01–CAP-1-12). Not surfaced (§6.2): none of A is on the rail; the census record landed the same day and no card cites it (CAP-1-10).

alignment: {'T1_cards_claims': ['F1', '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'], 'T1_composed': ['F1', 'F8', 'F9', 'F12', 'F19', 'F24', 'F25', 'F39', 'F45', 'F46'], 'T1_outside_cards': ["'contract terms' (no field source)", "'Google DeepMind' (the field names 'DeepMind', a PAI Partner and founding member)", "'fair pay' as an aim (the field's 'fairly compensate' is the paper's stated limit, F15; 'healthy and fair working conditions', F7)", "'A standard' and 'actively' (the field's 'Transparency Template, which outlines what companies should be monitoring and publicly reporting', F19)"], '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)"}

capture_findings_checked: CAP-1-01–CAP-1-12 read against the field ledger: CAP-1-03, CAP-1-04, CAP-1-08 and CAP-1-12 hold (the guidelines' content and the one-developer limit are on B8 and B4, absent from T). CAP-1-06 holds (B3 gives the template as what 'outlines what companies should be monitoring and publicly reporting'; 'a standard' is T's). CAP-1-11 is shown wrong in part by the field ledger: the PER reading held the template sentence 'beyond the saved sources (the census lists the template unread)'; card 3 (B3, PAI's 29 April 2026 post) states it nearly word for word, so the sentence is composed from a card, with its force raised. CAP-1-01 dates the vendor guidance and template to August 2024 as their introduction; B3 announces their finalized versions on 29 April 2026, and the composition's sentences follow the 2026 post (open block). Reported, not fixed.