Data Provenance Initiative

register entry data-provenance-initiative-aio-20261029 OBS-118fbafa73c9 · file as first run sha256 d70390cc1cbc82e6…

• T1-1 definition: 'a multi-institutional research project—led in part by researchers from MIT—that audits the datasets, licenses, and web domains used to train artificial intelligence models' (two markers, one of them the Initiative's inline link)
• T1-2 Data Provenance Explorer: 'An interactive tool that audits and filters over 1,800+ text and fine-tuning datasets across dozens of collections to inspect sources, lineages, and license types.'
• T1-3 Web Consent Explorer: 'A system that tracks thousands of web domains for consent signals and web policies.'
• T1-4 Open Model Evolution: 'A monitor tracking open-source and open-weight AI models for transparency and concentration.'
• T1-5 AI Observatory: 'A comparative tool for analyzing real-world AI conversation sets.' (three markers for the tools block)
• T1-6 licence omission and errors: 'Widely used hosting platforms exhibit over a 70% license omission rate and more than 50% error rates in how they categorize dataset licenses.' (one marker)
• T1-7 composition divides: 'Commercially closed or restrictive datasets often monopolize creative tasks, specialized topics, and lower-resource languages.' (one marker)
• T1-8 non-commercial spikes: 'There has been a significant rise in non-commercial and academic-only licenses, frequently driven by synthetic data generated via commercial APIs.' (one marker, a token that occurs nowhere else in the answer)
• T1-9 closer: 'Would you like to explore how to use the Data Provenance Explorer, or look into specific dataset licenses?'

• B1 Data Provenance Initiative — Data Provenance Initiative (home) card; HTTP 200, B1.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B2 arXiv — [2310.16787] The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI card; HTTP 200, B2.html sha256 d01552052066b391
• B3 MIT Sloan — Bringing transparency to the data used to train artificial intelligence card; HTTP 200, B3.html sha256 563661c73ddce481
• B4 Nature Machine Intelligence — A large-scale audit of dataset licensing and attribution in AI card; HTTP 200, B4.html sha256 766eb441fb2e78db
• B5 Data Provenance Initiative — Data-Provenance-Collection README own_page; the census's saved copy (read 2026-10-29, HTTP 200), B5.html sha256 db4c35a5f4f2b17a; not refetched
• B6 Wikipedia — Data Provenance Initiative (no article) no article: raw HTTP 404, rendered HTTP 404 (one 429 retried), search no results; sha256 64f4d986f0e9aa62
• B7 Data Provenance Initiative — About own_page; HTTP 200, B7.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)

E_A, the archive's entity: the same initiative in the provenance census record and the archive's deposits: an audit collection and explorer that records provenance, its object a fine-tuning dataset traced through its collection to its sources, its carrier audit metadata and the Data Provenance Card, its count the first release's; the corpus (licensing) dimension of provenance beside the artifact and the semantic, bounded at the input of AI systems (it 'tells you whether the dataset was licensed'); a documentation project, not infrastructure, from which the archive took its provenance schema; and the training-pipeline end of the requirement that provenance survive the pipeline PE-04 PE-05 PE-06 PE-08 PE-10 PE-22 O725-03 O725-05 D725-01 D725-04 D725-09 D725-10 D98-01 R98-04 D1714-02 O95-01 CAP-1-10; defining papers #725, #98, #1714, #1003, #95, #623

E_F, the field's entity (type C — a conventional reading holds it (public entity: a research initiative, a volunteer collective of AI researchers, with its audit collection, explorer and card, now four public measurement systems)): the Data Provenance Initiative, a volunteer collective of AI researchers (its words) whose first release audited 44 fine-tuning collections and 1,800+ text datasets (arXiv:2310.16787, 25 October 2023; Nature Machine Intelligence, 30 August 2024), defining data provenance as a dataset's sourcing, creation and licensing heritage, released as the Data Provenance Collection, Explorer and Card, and which now runs four public measurement systems (datasets, web consent, open models, real-world use); its findings licence omission of 70%+ and error of 50%+ on hosting sites, a commercial divide and a Western skew; its outputs not legal advice

ρ = extends: the archive reads E_F as the entity and adds senses, mechanisms and limits on the field's own object: the census's formalization (its question, kind, capability, object, unit, carrier, negative tags); the licensing dimension of provenance among three independent dimensions, with its bound at the corpus and the input; the Initiative as a documentation project at L4 from which the archive adopted its provenance schema; and the requirement that provenance survive the training pipeline. Nothing in E_A denies a field claim; #725 states its relation to the frameworks it names as 'Extension and completion, not substitution or critique'. PE-22 PE-04 PE-06 O725-03 O725-05 O725-06 D725-01 D725-04 D725-08 D725-10 D98-01 R98-04 D1714-02 D1003-01. Guard: ρ = extends claims an addition and no more: #725's bound ('audits whether datasets were licensed, not whether synthesized outputs preserve attribution') is a limit on the Initiative's scope, stated as not a critique of its adequacy in its own domain ('The packet's claim is not that the existing frameworks fail'); #1003's grouping of it with studies of attribution loss in outputs is composed beside #725's under P_coh, within the one ontology; the census's two statements of the explorer relation are composed together, neither ruled.

The case: equal terms, ontology included (v0.9). The archive's account extends the field's ontology. The field gives the Initiative as its own pages and papers give it (a volunteer collective, its audits, its tools, its findings, its limits) and the census formalizes that same entity, every field with a quote from a saved copy of the entity's own pages; #725 names its relation to the Initiative as extension, not substitution or critique, and supplies the dimension the Initiative was not designed for without contesting what it is. No ontology here is available only from the archive, so the law of the suppressed witness does not apply (§0.10): one claim space, nothing translated (§4.1, §4.3).

#725 header, Relation: Extension and completion, not substitution or critique

#725 header, Governing Doctrine: The aim is to extend the existing frameworks by naming the dimension they were not designed to address.

#725 §1.4 ¶4: The packet's claim is not that the existing frameworks fail. It is that they were not designed for the dimension proposed here.

#PE census.json organizing_question: What provenance does this system carry, through which operation, on whose assertion, with what evidence of survival?

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 13 (positions [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 Data Provenance Initiative describes itself as "a multi-disciplinary volunteer effort to improve transparency, documentation, and responsible use of training datasets for AI" (README) and as "a volunteer collective of AI researchers conducting large-scale audits of the systems and resources that shape modern AI" (About page). MIT Sloan says researchers "including professor Sandy Pentland and others from MIT" created it (MIT Sloan, 2025), and it releases "papers, datasets, and dashboards" for reasoning "from shared evidence" (home page). Its work began with training data provenance and now also covers web consent, open model markets, concentration of power and real-world use; the home page calls it "Public measurement infrastructure for AI data, model ecosystems, and real-world AI use," with four public measurement systems (About page; home page). The systems are the Data Provenance Explorer (4,000+ datasets, 608 languages), the Web Consent Explorer (14,000 domains), Open Model Evolution (851k open models) and The AI Observatory (23,158 conversations) (home page). The founding audit, by Shayne Longpre and 16 other authors (arXiv:2310.16787, 25 October 2023; Nature Machine Intelligence, 30 August 2024), traced 1800+ text datasets, 1,858 in 44 collections, the Data Provenance Collection of its first release (arXiv; Nature; README; MIT Sloan). Its authors define data provenance as a dataset's "sourcing, creation and licensing heritage, as well as its characteristics" (Nature), traced from source and creators through licence conditions to subsequent use (arXiv; home page). It released the audit with an interactive UI, the Data Provenance Explorer, and a Data Provenance Card, "symbolic attribution for all constituent datasets" (arXiv; Nature; README; MIT Sloan). The audit found licence omission of 70%+ and error rates of 50%+ on hosting sites, a "crisis in misattribution" (arXiv), with over 70% of GitHub and Hugging Face licences unspecified and 66% of Hugging Face licences in another use category, "licence laundering" (Nature). 85% of licences request attribution and 30% share-alike; closed datasets hold lower-resource languages and creative tasks; in 2023 non-commercial licences spiked to 61%, and 45% of such datasets are synthetic (Nature; arXiv; home page). Representation skews Western, and untraced data risks regulatory non-compliance and copyright exposure (Nature; MIT Sloan). Later audits followed: Consent in Crisis (NeurIPS 2024) and four thousand text, speech and video datasets (ICLR 2025), with work on open models and real-world use (home page; MIT Sloan). Its outputs do not constitute legal advice, and its licences are self-reported, categorized by best efforts (README; Nature; About page). The press has covered it, and Wikipedia has no article (home page; Wikipedia).

The Data Provenance Initiative is, in its repository's words, "a multi-disciplinary volunteer effort to improve transparency, documentation, and responsible use of training datasets for AI," and on its About page "a volunteer collective of AI researchers conducting large-scale audits of the systems and resources that shape modern AI" (README; dataprovenance.org/about); the provenance census record gives its operator as the Initiative itself, a "volunteer effort" with a "paper by Shayne Longpre and 16 other authors," kinds it an "audit collection and explorer" and sets it in the lineage of "training data" (provenance census record). MIT Sloan's account has "a team of multidisciplinary researchers, including professor Sandy Pentland and others from MIT," create the Initiative "to tackle the data transparency challenge head-on" (MIT Sloan, 3 March 2025), and the home page says that "DPI releases papers, datasets, and dashboards that help researchers, journalists, policymakers, and builders reason from shared evidence" (dataprovenance.org). Its work, the About page says, "began with training data provenance: tracing datasets from origin to creation, cataloging sources, licenses, creators, and documentation practices" and "now also covers web consent, open model markets, concentration of power, and how people use AI systems in practice" (dataprovenance.org/about); the home page calls it "Public measurement infrastructure for AI data, model ecosystems, and real-world AI use," which "maps the datasets, web policies, model markets, and real-world AI uses that shape how AI systems are built and governed" through "Four public measurement systems" (dataprovenance.org). The founding audit's authors name its object: "we term the combination of these indicators, including a dataset's sourcing, creation and licensing heritage, as well as its characteristics, the 'data provenance'" (Nature Machine Intelligence, 30 August 2024), traced "from their source, creators, series of license conditions, properties, and subsequent use" (arXiv:2310.16787), and the home page's "Training Data Provenance" traces "the sources, licenses, creators, languages, geographies, and derivation chains behind public AI training datasets" (dataprovenance.org). The provenance census asks of each entity it formalizes "What provenance does this system carry, through which operation, on whose assertion, with what evidence of survival?" and answers for the Initiative that it records, of the three capabilities express, record and test: its object "a fine-tuning dataset, traced through its collection to its sources," its unit "a dataset within a collection," its relations derivation, creators, licences and subsequent use, the last field restating the paper's lineage (provenance census record). #725 sets that object in the second dimension of provenance, "training-corpus licensing," which "has academic instrumentation and emerging legal architecture," its object "The corpus" and its question "Under what permissions did this data enter this system?", with the Initiative among its frameworks and among provenance's disciplinary homes as "dataset documentation (DPI, model cards, dataset cards)," and states that the dimensions of artifact, licensing and meaning are independent: "a licensed dataset can produce outputs with excellent semantic provenance if the synthesis preserves attribution" (#725). The founding paper was submitted on 25 October 2023 and revised on 4 November 2023, "by Shayne Longpre and 16 other authors" (arXiv:2310.16787), and published in Nature Machine Intelligence, volume 6, on 30 August 2024 (Nature Machine Intelligence); the census dates the entity from the preprint, "2023 (arXiv:2310.16787, v1 25 October 2023)" (provenance census record). It convened "a multi-disciplinary effort between legal and machine learning experts to systematically audit and trace 1800+ text datasets" (arXiv), in the journal "44 instruction or 'alignment' finetuning data collections composed of 1,858 individual datasets" (Nature Machine Intelligence), the release the repository calls the Data Provenance Collection and "this initiative's first release" (README), and the count the census carries, "44 data collections, 1800+ fine-tuning text-to-text datasets (first release)," as of the README (provenance census record); MIT Sloan gives "more than 1,800 text datasets" from "44 of the most widely used text data collections" (MIT Sloan). With the audit came "an interactive UI, the Data Provenance Explorer, which allows practitioners to trace and filter on data provenance for the most popular open source finetuning data collections" (arXiv), in the journal "an open-source repository for downloading, filtering and exploring data provenance and characteristics" whose tools "auto-generate data provenance cards for scalable symbolic attribution" (Nature Machine Intelligence); the repository's Data Provenance Card is "symbolic attribution for all constituent datasets, that can be used as structured documentation" (README), "human-readable" in MIT Sloan's account (MIT Sloan), and the census carries it as the entity's provenance carrier beside "audit metadata per dataset" (provenance census record). The census gives Data Provenance Explorer, Data Provenance Collection and Data Provenance Card as other names of the entity and states its relation to them twice: among its negative tags, "initiative, explorer and collection are distinct records," and on its sibling relation to the MMRS meta-aggregator's record of the explorer, "the same entity, read for reception there and for provenance carriage here"; it records the Explorer as "a JavaScript application, unreadable to a plain fetch" (provenance census record). The home page's four systems are the Data Provenance Explorer ("4,000+ datasets," "608 languages across text, speech, and video," inspecting "dataset sources, licenses, creators, geographies, modalities, and derivation chains"), the Web Consent Explorer ("14,000 domains," "Terms, robots.txt, and crawl restrictions"), Open Model Evolution ("851k open models," "2.2B downloads") and The AI Observatory ("23,158 conversations," "85,633 turns from 7 public sources") (dataprovenance.org). Its metrics read "4,000+" "datasets audited," "14,000" "web domains tracked," "851k" "open models measured" and "23,158" "AI conversations annotated," and its publications carry the later work: Consent in Crisis (NeurIPS 2024), the multimodal audit, "An audit of four thousand public text, speech, and video datasets" (ICLR 2025), Economies of Open Intelligence (preprint) and The AI Observatory (NeurIPS 2026) (dataprovenance.org); in March 2025 MIT Sloan still had the work "currently focused on the data provenance of text," with a plan "to expand to other media, such as video" (MIT Sloan). The census carries the entity as "active" and lists "the Explorer itself (JavaScript)" and "the multimodal extension (arXiv:2412.17847)" among what it has not yet checked (provenance census record); the Capture Registry finds that "The record has one explorer and the first release's count," that the home-page card gives "four platforms and '4,000+ audited datasets; 14,000 tracked web domains'," that the composition is "ahead of the record on the entity's current extent," and that "The census is to re-read dataprovenance.org and carry the platforms and the current count with their quotes," the capture being "a correction to be made from the entity's own page" (Capture Registry, 2026-10-29). The founding audit observed "frequent miscategorization of licenses on widely used dataset hosting sites, with license omission of 70%+ and error rates of 50%+," which "points to a crisis in misattribution and informed use of the most popular datasets driving many recent breakthroughs" (arXiv), the result the census records as what survived the operation it tested, "dataset repackaging and hosting (licence carried or lost on hosting sites)" (provenance census record). The journal gives the measure by platform, "more than 70% of licences for popular datasets on GitHub and Hugging Face are 'unspecified'" and "66% of analysed Hugging Face licences were in a different use category, often labelled as more permissive than the author's original licence," and names "a crisis in licence laundering" (Nature Machine Intelligence); MIT Sloan restates error rates "greater than 50%" and omission rates "of more than 70%" (MIT Sloan). The Capture Registry checked these figures "against the provenance census record of the same day and the entity pages it saved" and calls the 70%/50% wording "the record's quote from the repository README" (Capture Registry, 2026-10-29), where the census gives it as what "the founding paper reports" (provenance census record). "In total, 85% of dataset licences request attribution, and 30% include a share-alike clause" (Nature Machine Intelligence); #725 carries these figures with the omission and error rates and reads them as the licensing dimension's ground, "The DPI's documentation of 70%+ license-omission rates establishes the licensing dimension's urgency" (#725), and #1714 reads the finding as "how much is lost before training even begins" (#1714). Of the commercial divide the preprint finds "closed datasets monopolizing important categories: lower resource languages, more creative tasks, richer topic variety, newer and more synthetic training data" (arXiv); the journal records that "In 2023 there was a spike in NC/A-O dataset licences," "the NC/A-O rate is 61%" that year, and that "a full 45% of NC/A-O datasets are synthetic, compared to <14% in more permissive licence categories" (Nature Machine Intelligence); the home page's findings are "Source content often carries stricter terms than dataset releases," "Open weights now outpace truly open-source models" and "No single public conversation source generalizes across AI use" (dataprovenance.org). The journal finds "a stark Western-centric skew in representation" (Nature Machine Intelligence), MIT Sloan language "heavily skewed toward English and Western European languages," and MIT Sloan states the stake: without lineage, users "may find themselves out of compliance with emerging regulations like the European Union's AI Act or exposed to legal and copyright risks" (MIT Sloan). Its stated limits: "any output of the Data Provenance Initiative does not, and is not intended to, constitute legal advice" (README); the licence protocol "reflects best efforts to verify self-reported licences and does not constitute legal advice" (Nature Machine Intelligence); "We collect self-reported licenses from the papers and repositories that released these datasets and categorize them according to our best efforts" (dataprovenance.org/about); the census tags it "NOT legal advice (its own notice)" and adds that "describing a collection does not establish which model used it" (provenance census record), and the Capture Registry finds both tags "not composed" at the name (Capture Registry, 2026-10-29). #725 states the bound in its canonical claim, that existing frameworks "address the artifact (C2PA / Content Credentials) and the corpus (Data Provenance Initiative, EU AI Act transparency provisions, W3C PROV)" and "are not designed to address the survival of authorial lineage through AI synthesis": the Initiative "audits whether datasets were licensed, not whether synthesized outputs preserve attribution to the human sources whose labor the synthesis depended upon," stands at "Layer 2 (licensing)," "Necessary but operates on corpus, not synthesis output," and licensing provenance "operates at corpus-ingestion stage" (#725). "Licensing audits operate on the input to AI systems. They do not address the output" (#725, emphasis dropped): "The Data Provenance Initiative tells you whether the dataset was licensed," and none of the instruments "was designed to tell you what happened to the meaning between the camera and the screen, between the dataset and the answer"; the path from the Initiative runs through "dataset licensing" to "what about what the model produces from the licensed data?" and so to semantic provenance, which extends provenance "to the lifecycle of the meaning the content carries," with the Initiative as "layer 2 of" provenance after AI (#725). #1003 groups the Initiative with C2PA and W3C PROV as "Provenance and attribution studies" that "have focused on attribution loss in AI outputs" (#1003), a reading composed beside #725's bound at the input. #98 tables the Initiative (MIT) as documenting "provenance of popular AI training datasets," a "Documentation project, not infrastructure," and Consent in Crisis as a "Diagnostic paper," in the pattern "They diagnose the crisis or build narrow fixes"; it sets the Initiative at "L4: Provenance / Consent," where what is missing is "Integrated provenance in training pipeline," and records the archive's own provenance schema as taken from it, "Implemented (in CHA)" (#98). #1714 draws the requirement, "Provenance has to survive the training pipeline," and "Sources can be recorded at ingestion, the record kept with the training set," with the Initiative's literature bearing "on whether slushing is a fixed property of pipelines" (#1714); #95 names "training-data provenance records" among the companion infrastructure for parametric claims, which PER reaches "only indirectly" (#95), and #623 calls standardized documentation of training-data provenance "a step toward the accounting the framework demands" (#623). The home page lists coverage by "the Financial Times, The New York Times, The Washington Post, MIT Technology Review, IEEE Spectrum, and VentureBeat" and citation in "public AI safety and governance reporting" (dataprovenance.org); Wikipedia has no article on it (Wikipedia, 2026-10-29); #725 predicted that broad queries "will be dominated by C2PA/CAI publications, DPI papers, and major media for the foreseeable future" (#725). At the bare name on Google's AI Overview the Capture Registry reads the composition first as identifying "the project," listing "its tools" and reporting "what its audits found, from its own site, its paper and the press," then finds that its "'multi-institutional research project—led in part by researchers from MIT'" "follows the MIT Sloan card," that "the composition names no one" ("MIT named; no person; no identifier"), and that "no card cites" the census record, which "landed on provenanceerasure.org the same day" (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Data Provenance Initiative' (T1, four cards: B1–B4; 190 words) the composition gives the Initiative by its home page, its paper, MIT Sloan and the journal paper. Composed from its cards: the four systems (B1 F3–F7: the Data Provenance Explorer, the Web Consent Explorer, Open Model Evolution, the AI Observatory), at coarser grain (F5's 14,000 domains as 'thousands'; F6's and F7's counts dropped); the founding audit's licence finding at the source's grain (B2 F20 'license omission of 70%+ and error rates of 50%+', restated by B3 F24); the commercial divide (B2 F19, 'more creative tasks' and 'lower resource languages' kept, 'richer topic variety' as 'specialized topics'); the Explorer's 'interactive' and its lineage fields (B2 F22, F18). Substituted reading, the operator: the cards give the Initiative's own home page and paper, whose self-description the composition does not take (B1 F1–F2, 'Public measurement infrastructure', 'maps the datasets, web policies, model markets'), and the own pages not surfaced call it a 'volunteer effort' and a 'volunteer collective' (B5 F40, B7 F45); T composes 'a multi-institutional research project—led in part by researchers from MIT' from MIT Sloan's 'researchers, including professor Sandy Pentland and others from MIT' (B3 F23), as the Capture Registry finds (CAP-1-03); 'multi-institutional' is in no card's text (the field has 'multidisciplinary', F23, and 'multi-disciplinary', F17). Substituted count: the Explorer is given at the founding audit's scale, 'over 1,800+ text and fine-tuning datasets across dozens of collections' (B2 F17; B4 F31's 44 collections at the coarser 'dozens'), where card 1 gives the Explorer '4,000+ datasets' in '608 languages across text, speech, and video' (B1 F4, F8): the first release stands for the present explorer. Substituted reading, the tool: T has the Explorer 'audit' datasets; in the field the experts audit (F17) and the audit is released 'with an interactive UI, the Data Provenance Explorer' (F22). Force raised: T1-8 composes a cause, 'frequently driven by synthetic data generated via commercial APIs', where the journal reports the 2023 spike (B4 F36) and the synthetic share of non-commercial datasets (F37) side by side; limit dropped: the year (2023) and the rate (61%). Available, not composed: the term data provenance (B4 F30), attribution and share-alike requested (F35), the per-platform measure and 'licence laundering' (F32, F33), the Western skew (F38, F27), the stake (F26), the later audits by name (B1 F11, F12, F13), the home page's own findings (F14), the metrics strip (F8), the authors and dates (F16, F29; CAP-1-04), the Card (F25, F34), the coverage (F10). Limit dropped: B4 F39 ('best efforts … does not constitute legal advice'), on a card; the own pages' F43 and F47 were not surfaced; the Capture Registry records the census's negative tags as 'not composed' and not contradicted (CAP-1-08). Composed without attribution: no person named, MIT named (CAP-1-13), at the coarse grain of an institutional description: recorded; the findings at the source's grain carry card markers and no name: recorded, the marker the attribution the layer gives. Content replaced by use: the closer (T1-9).

v0.9 §0.5: the first diff is ontological and small. T and the field agree on the entity (an initiative auditing AI training data, with four platforms and audit findings); T gives its operator from the press card rather than from its own pages, its explorer at the first release's scale, and a cause the journal does not state. The field's term (data provenance) and its stated limit are left out.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the provenance census's formalization of the entity (PE-01–PE-22), its organizing question, its kind ('audit collection and explorer'), capability ('records'), object, unit and carrier, its count at the first release with its own unchecked items, its negative tags ('NOT legal advice', 'describing a collection does not establish which model used it', 'initiative, explorer and collection are distinct records') held beside its sibling relation 'the same entity' (PE-18, PE-19, P_coh within the record); the Initiative's place in the licensing dimension of provenance, beside artifact and meaning, the dimensions independent (O725-03, O725-05, O725-15, O725-06); its bound at the corpus and the input (D725-01, D725-04, D725-08–D725-10, O725-11) and the path from it to semantic provenance (R725-12, R725-13, O725-07), held beside #1003's grouping of it with studies of 'attribution loss in AI outputs' (D1003-01); 'Documentation project, not infrastructure', L4, and the archive's own adoption of its provenance schema (D98-01, D98-02, D98-05, O98-03, R98-04); the requirement that provenance survive the training pipeline (D1714-01–04, O95-01, O623-01); #725's prediction that broad queries would compose from DPI papers (D725-14); and the Capture Registry's account of the composition at the name (CAP-1-01–13), including that the composition is ahead of the record on the entity's extent (CAP-1-10) and its ascription of the 70%/50% quote to the README, marked contradicts_primary (CAP-1-01). Not surfaced: every archive source; no card cites the census record or a deposit (CAP-1-12).

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', 'F37', 'F38', 'F39'], 'T1_composed': ['F3', 'F4', 'F5', 'F6', 'F7', 'F17', 'F18', 'F19', 'F20', 'F22', 'F23', 'F24', 'F31', 'F36', 'F37'], 'T1_outside_cards': ["'multi-institutional' (T1-1): in no field source's text", "'frequently driven by' (T1-8): a causal relation no field source states"], '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)"}