Data Provenance Explorer

register entry data-provenance-explorer-aio-20261029 OBS-dc0adc018f10 · file as first run sha256 314d4267ef914693…

• T1-1 definition: 'an interactive audit tool built by the Data Provenance Initiative to track, inspect, and filter large language model (LLM) training datasets' (three markers, one of them the Initiative's link)
• T1-2 dataset explorer: 'Dataset Explorer: Search and filter nearly 4,000 text, speech, and video datasets by creators, original data sources, geographies, modalities, and derivation chains.'
• T1-3 licence filtering: 'License & Use Case Filtering: Evaluate data strictness ranging from commercial use to academic-only or non-commercial restrictions.' (one marker)
• T1-4 web data: 'Web Data & Consent Explorers: Track shifting data-use preferences and consent signals across thousands of web domains commonly scraped for AI training.'
• T1-5 closer: 'Would you like to search for a specific dataset or learn how to filter for commercial-use licenses in the explorer?'

• B1 Data Provenance Initiative — Data Provenance Explorer card; B1.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B2 MIT Media Lab — Data Provenance Explorers - MIT Media Lab unreachable: TLS verification failed on three fetches (curl 60); card title and snippet only
• B3 Data Provenance Initiative — Data Provenance Initiative card; B3.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B4 Data Provenance Initiative — Data Provenance Explorer card; B4.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B5 Data Provenance Initiative — Dataset Explorer - Data Provenance Initiative card; B5.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B6 Cohere — Data Provenance Explorer Launches to Tackle Data Transparency Crisis card; B6.html sha256 aa465dcc01c69325
• B7 arXiv — [2310.16787] The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI own_page; B7.html sha256 d01552052066b391
• B8 arXiv — [2407.14933] Consent in Crisis: The Rapid Decline of the AI Data Commons own_page; B8.html sha256 8cedde3659e5ed29
• B9 Data Provenance Initiative — /data_summaries.csv own_page; B9.csv.gz sha256 e1eb7bf01f96315f
• B10 Data Provenance Initiative — Web Data Explorer own_page; B10.html sha256 bf609f385257c3a1; text from the script bundle (site-main.af19d7fc.js)
• B11 Wikipedia — Data Provenance Explorer (no article) no article: raw HTTP 404, rendered HTTP 404, search no results; sha256 136eb91afa8a5c7f
• B12 Wikipedia — Data Provenance Initiative (no article) no article: raw HTTP 404, rendered HTTP 404, search no results; sha256 683496b15564dda4

E_A, the archive's entity: the same explorer, in the MMRS meta-aggregator's record and the archive's deposits: a database tracing the sources and licences of AI training datasets, its unit the dataset traced to creators, sources, licences, geographies, modalities and derivation chain; the corpus (licensing) dimension of provenance, which audits whether datasets were licensed and not whether synthesized outputs preserve attribution (#725); a documentation project, not infrastructure, whose provenance schema the archive implements (#98); what is lost before training begins, where provenance has to survive the training pipeline (#1714); the training-layer end of one chain whose other end, what a model credits when it answers, the reception aggregators and the Capture Registry record (MMRS record; #95) MMRS-01 MMRS-11 MMRS-10 MMRS-07 MMRS-08 D725-01 D725-04 D725-08 O725-05 D98-01 R98-04 D1714-02 O95-01 CAP-1-07; defining papers #725, #98, #1714, #95, #1003, #623

E_F, the field's entity (type C — a conventional reading holds it (public entity: a database and its explorer tool, with its operator's audits)): the Data Provenance Initiative's interactive explorer of its audit of AI training datasets, released with the founding paper (arXiv:2310.16787, 25 October 2023; 1800+ text datasets) and launched with Cohere for AI and eleven other institutions; now a Dataset Explorer of nearly 4,000 text, speech and video datasets (from the multimodal audit, ICLR 2025), each traced to its sources, licences, creators, geographies, modalities and derivation chains, beside a Web Data Explorer of consent signals over 14,000 domains (Consent in Crisis, arXiv:2407.14933); its audits found licence omission of 70%+ and error rates of 50%+ on hosting sites, a commercial divide and a Western skew, and that 45% of C4 is restricted by terms of service; its authors state that licensing ambiguities cannot be resolved by tooling alone

ρ = extends: the archive reads E_F as the entity and adds senses, mechanisms and limits on the field's own object: the licensing dimension among three independent dimensions (D725-01, O725-05, O725-06), its bound at the input (D725-04, D725-08, D725-09), documentation rather than infrastructure and its schema implemented in the archive (D98-01, R98-04), the requirement that provenance survive the pipeline (D1714-02, D1714-03, O98-03), the chain from the training layer to the composition layer (MMRS-07, MMRS-08, O95-01), and what the composition layer leaves out of it (CAP-1-04, CAP-1-07) D725-01 O725-05 O725-06 D725-04 D725-08 D725-09 D98-01 R98-04 D1714-02 D1714-03 O98-03 MMRS-07 MMRS-08 O95-01 CAP-1-04 CAP-1-07. 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 reading of the Initiative's scope, not a finding that its audit is defective, and its own text calls semantic provenance 'a constructive extension of existing frameworks — not a critique of their adequacy in their own domains'; #98's 'Documentation project, not infrastructure' is a limitation by the archive's measure of infrastructure; #1003's contrary grouping is composed beside #725 under P_coh; the archive's figures beyond the field (Nature Machine Intelligence 2024; 85% and 30%, D725-02) are carried at their own grade (attributed) and opened for resolution.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity: the MMRS record describes the Explorer from the Initiative's own pages and papers, each field with its basis, and the deposits that name the Initiative (#725, #98, #1714, #1003) take its audit and its findings as given and place it, adding senses and limits on the field's own object: the corpus (licensing) dimension among independent dimensions of provenance, its bound at the input of AI systems, documentation rather than infrastructure, the archive's adoption of its provenance schema, and the training-layer end of one chain whose other end is the composition layer. #725 states its own relation in those terms ('Extension and completion, not substitution or critique'). No defining paper contests the field's ontology or supplies one the field lacks for the Explorer: ρ = extends, the equal-terms case (§0.9, §1.2a). One reading within the archive is held under P_coh rather than ruled: #1003 groups the Initiative with studies of 'attribution loss in AI outputs', #725 bounds it at the input.

#MMRS articles.py ARTICLES['data-provenance-explorer'], 'Place in the field' s1–2: It records provenance at the training layer, where the census's reception aggregators record it at the composition layer. The two ends of one chain: what a model was given, and what it credits when it answers.

#MMRS articles.py, kernel s1: The Data Provenance Explorer, from the Data Provenance Initiative (Longpre, Mahari and others), traces the sources and licences of AI training datasets: nearly 4,000 text, speech and video datasets by its own count.

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

#725 ¶ 'Semantic provenance names the dimension': It is offered as a constructive extension of existing frameworks — not a critique of their adequacy in their own domains.

#725 §3.2, Correction: The DPI's documentation of 70%+ license-omission rates establishes the licensing dimension's urgency; semantic provenance addresses the dimension that follows.

#98 the component table, row 'Provenance schema': | Provenance schema | Data Provenance Initiative | Training-pipeline provenance | Implemented (in CHA) |

#1714 §9.2 s1: Provenance has to survive the training pipeline.

#MMRS census.json schema_note: each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading)

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.

The Data Provenance Explorer lets its users explore the datasets included in the Initiative's audit, filtering and analysing language-model training datasets by use case, language and more (explorer introduction; landing page; MIT Media Lab's card, snippet only). The founding paper released the whole audit with an interactive UI, the Data Provenance Explorer, to trace and filter on provenance for popular open-source fine-tuning collections (arXiv:2310.16787), for developers filtering for legal and ethical considerations and for scholars and journalists exploring data lineage (Cohere). Data provenance, in the launch post's words, is the tracking of "the origins of a piece of data" (Cohere). Each dataset is traced to its sources, licences, creators, geographies, modalities and derivation chains, its re-licensings tagged (home page; Cohere), and the data file the Dataset Explorer loads carries these fields with the Initiative's licence readings, 3,917 rows at fetch, over which the explorer computes its summaries (data_summaries.csv; Dataset Explorer). Its pages count nearly 4,000 text, video and speech datasets, 4,000+ in 608 languages, from the multimodal audit of four thousand datasets (ICLR 2025), which found source content often more restricted than dataset releases (its pages). Longpre, Mahari, Chen, Obeng-Marnu and others submitted the founding audit of 1800+ text datasets on 25 October 2023 (arXiv), launched the same day with Cohere for AI and eleven other institutions as a free platform open to contributions, nearly 2,000 fine-tuning datasets in the launch post's count (Cohere), with its paper and repository linked from the explorer (Dataset Explorer). The Initiative calls itself public measurement infrastructure for AI data, model ecosystems and real-world AI use, with four measurement systems and papers, datasets and dashboards (home page). The audit found licence omission of 70%+ and error rates of 50%+ on hosting sites (arXiv), with 72 to 83 percent of licences missing on crowdsourced aggregators such as GitHub and Papers with Code and licences listed too permissively in up to 29 percent of cases (Cohere). It found sharp divides between commercially open and closed datasets and a Western-centric skew (arXiv; Cohere), and its authors state that legal ambiguities in data licensing cannot be resolved by tooling alone (Cohere). The Web Data Explorer tracks shifting website preferences on AI use over 14,000 domains (its pages), from Consent in Crisis (submitted 20 July 2024; NeurIPS 2024), the first longitudinal audit of consent protocols, charting HEAD-C4's 2k critical domains with recent spikes in restrictions (arXiv; Web Data Explorer). In 2023–2024 about 5% of C4's tokens became fully restricted and 45% of C4 by terms of service, symptoms of ineffective web protocols which, if respected or enforced, bias general-purpose AI (arXiv:2407.14933). The Initiative reports press coverage and governance use (home page); Wikipedia has no article on the Explorer or the Initiative.

The Data Provenance Explorer is the interactive explorer of the Data Provenance Initiative's audit of AI training datasets: the founding paper released "our entire audit, with an interactive UI, the Data Provenance Explorer," to trace and filter on data provenance for the most popular open-source fine-tuning collections (arXiv:2310.16787), and the MMRS record defines it as tracing "the sources and licences of AI training datasets," kinds it a database, and has practitioners follow provenance back to sources, creators and licences (MMRS record). Its own pages introduce it as the place to "Explore the datasets included in our audit," where the DPI Explorer tool filters and analyses language-model training datasets by use case, language and more (explorer introduction; landing page), and MIT Media Lab's card, the one the composition could surface but this row could not fetch, speaks of "data explorers" built on "our extensive audits of AI data" for practitioners' decisions (MIT Media Lab, title and snippet only). The launch post names its users: developers who "track and filter thousands of datasets for legal and ethical considerations," and scholars and journalists who explore "the composition and data lineage of popular AI datasets" (Cohere, 25 October 2023). Its unit is the dataset, "traced to its creators, sources, licences, geographies, modalities and derivation chain" (MMRS record), as the home page has it, "Inspect dataset sources, licenses, creators, geographies, modalities, and derivation chains"; the data file the Dataset Explorer loads records those fields per dataset beside the Initiative's own licence-use, attribution and share-alike readings, 3,917 rows at this row's fetch (data_summaries.csv), and the MMRS record answers that it records attribution: "yes: sources, creators and licences per dataset" (MMRS record). "For the first time," the launch post says, the audited datasets carry "tags to the original data sources, numerous re-licensings, creators, and other data properties" (Cohere), and the Dataset Explorer computes over any selection its text sources, generating models and the share of datasets used (Dataset Explorer). The term is its authors': "Inconsistent tracking of the origins of a piece of data, which we refer to as data provenance, has introduced a range of legal and ethical challenges" (Longpre and Hooker, Cohere); #725 makes that tracking the second of the dimensions of provenance, "training-corpus licensing," whose object is the corpus and whose question is "Under what permissions did this data enter this system?", with the Initiative among its frameworks, beside the artifact (C2PA) and the semantic dimension, which "names the dimension that completes the C2PA ambition of trust in digital content by extending provenance from the moment of creation to the lifecycle of the meaning the content carries" (#725). Its own count is "nearly 4,000 text, video, and speech datasets" on the Dataset Explorer, "4,000+ datasets" in "608 languages" on the home page and "4000" on the landing page, and the home page gives its source as the multimodal audit "of four thousand public text, speech, and video datasets" (ICLR 2025), with the finding that "Source content often carries stricter terms than dataset releases" (its pages). The Capture Registry records that the composition at the name supplied this scale from the Initiative's own Dataset Explorer card while the MMRS record still gave the founding audit's 1,800, and that the record was corrected to the current scale, sourced to the card (Capture Registry, 2026-10-29); the record now reads "nearly 4,000 text, video and speech datasets (the Dataset Explorer); the founding audit covered more than 1,800 text datasets" (MMRS record). The founding paper, submitted on 25 October 2023 by Shayne Longpre, Robert Mahari, Anthony Chen, Naana Obeng-Marnu and others, convened legal and machine-learning experts "to systematically audit and trace 1800+ text datasets" with "tools and standards to trace the lineage of these datasets, from their source, creators, series of license conditions, properties, and subsequent use" (arXiv:2310.16787); the launch post of the same day counts "nearly 2,000 of the most widely used fine-tuning datasets," downloaded "tens of millions of times" (Cohere). Researchers "from MIT, Cohere for AI, and 11 other institutions" launched it as a free platform, the audit being "the launch of a wider multi-institutional initiative, where users all over the world can contribute to the explorer" (Cohere), and the Dataset Explorer still stands under the Data Provenance Explorer with its paper and repository linked (Dataset Explorer). Its operator, the Data Provenance Initiative (Longpre, Mahari, Chen, Obeng-Marnu and others, by the founding paper: MMRS record), calls its work "Public measurement infrastructure for AI data, model ecosystems, and real-world AI use" (its home page, which the MMRS record quotes and whose phrase the record's evidence assigns to the founding paper), and says it releases "papers, datasets, and dashboards" so that researchers, journalists, policymakers and builders can "reason from shared evidence" (home page). #98 reads the Initiative as a project that "Documents provenance of popular AI training datasets" and states its limitation as "Documentation project, not infrastructure," and in its table of what the archive takes from others it records "Provenance schema | Data Provenance Initiative | Training-pipeline provenance | Implemented (in CHA)" (#98). The founding audit found "license omission of 70%+ and error rates of 50%+" on widely used dataset hosting sites, "a crisis in misattribution" (arXiv:2310.16787); the MMRS record quotes the finding, #725 reports it with the further figures that "85% of licenses request attribution and 30% include share-alike clauses" and with Nature Machine Intelligence 2024 as its venue, and #1714 carries it as what "is lost before training even begins" (MMRS record; #725; #1714). The launch post places the omission in "Crowdsourced aggregators like GitHub, Papers with Code," which miss licences for 72 to 83 percent of datasets against 30 percent under the Initiative's own protocol and list them "too permissively in up to 29 percent" of cases, so that distorted licensing "leads to cascading misattribution" (Cohere). The audit also found "sharp divides" between commercially open and closed datasets, closed ones "monopolizing" lower-resource languages, more creative tasks, richer topic variety and newer, more synthetic data (arXiv:2310.16787), a divide the launch post calls "sharp and widening," beside "a stark Western-centric skew" in representation (Cohere). Its second explorer, the Web Data Explorer (on the home page the Web Consent Explorer), follows "Shifting preferences in how websites allow their data to be used for AI" across 14,000 domains' terms, robots.txt and crawl restrictions (its pages), from Consent in Crisis (Longpre, Mahari, Lee and others, submitted 20 July 2024; NeurIPS 2024 on the home page, "NeurIPS 2024 Datasets and Benchmarks" in the MMRS record), "the first, large-scale, longitudinal audit of the consent protocols for the web domains underlying AI training corpora" (arXiv:2407.14933). The explorer page charts HEAD-C4's "2k of the largest and most critical web domains" from 2016 to April 2024, counts a domain restricted when its robots.txt disallows the named AI crawlers, and finds "significant recent spikes in restrictions" (Web Data Explorer). The paper finds that in 2023–2024 "~5%+ of all tokens in C4, or 28%+ of the most actively maintained, critical sources in C4" became fully restricted and that "a full 45% of C4 is now restricted" by terms of service (arXiv:2407.14933; the MMRS record quotes the second), diagnoses "ineffective web protocols," and holds, as a conditional, that "If respected or enforced, these restrictions are rapidly biasing the diversity, freshness, and scaling laws for general-purpose AI systems" (arXiv:2407.14933). #98 reads Consent in Crisis as an "Audit of C4, RefinedWeb, Dolma" that "documents collapse of web-consent infrastructure," a "Diagnostic paper," and sets the Initiative at "L4: Provenance / Consent," where what is missing is "Integrated provenance in training pipeline" (#98). Its authors state that "major legal ambiguities remain in data licensing that the authors acknowledge cannot be resolved by tooling alone" (Cohere), and #725 states the dimension's own bound: "The Data Provenance Initiative audits whether datasets were licensed, not whether synthesized outputs preserve attribution to the human sources whose labor the synthesis depended upon"; "Licensing audits operate on the input to AI systems. They do not address the output," and the Initiative's 70%+ omission "establishes the licensing dimension's urgency; semantic provenance addresses the dimension that follows" (#725). The dimensions are independent, "a licensed dataset can produce outputs with excellent semantic provenance if the synthesis preserves attribution; an unlicensed dataset can produce outputs that nonetheless cite their sources clearly," and the frameworks that address "the corpus," the Initiative among them, "are not designed to address the survival of authorial lineage through AI synthesis" (#725). "The Data Provenance Initiative tells you whether the dataset was licensed," and none of these instruments was designed to tell "what happened to the meaning … between the dataset and the answer" (#725); #1003 groups "C2PA, the Data Provenance Initiative, W3C PROV" as studies that "have focused on attribution loss in AI outputs" (#1003). The MMRS record sets the Explorer at one end of a chain: "It records provenance at the training layer, where the census's reception aggregators record it at the composition layer. The two ends of one chain: what a model was given, and what it credits when it answers" (MMRS record), with its roles provenance and monitoring (MMRS record); the Initiative's own four systems "trace AI from source data and web consent to model markets and real-world use" (home page). At the answering end, PER "applies to these only indirectly" where claims are parametric, and reaching them at the substrate level "requires companion infrastructure (training-data provenance records, watermarking, content credentials)" (#95); #623 counts standardized documentation of training-data provenance as "a step toward the accounting the framework demands" (#623). #1714 states the requirement across the chain, "Provenance has to survive the training pipeline," since neither audit nor influence tracing "recovers authorship" and "what survives depends on what is recorded and computed," so that sources can be "recorded at ingestion, the record kept with the training set" (#1714), where the founding audit traced lineage as far as a dataset's "subsequent use" (arXiv:2310.16787). The MMRS record keeps each assertion with "a source and a basis" and keeps lifecycle (active, observed 2026-10-28) apart from the observation that saw it, and it states its own limits: it has not read "the explorers' pages directly (they render by script)" and takes the Initiative's page text from "the Overview's cards" (MMRS record). The Initiative lists coverage by "the Financial Times, The New York Times, The Washington Post, MIT Technology Review, IEEE Spectrum, and VentureBeat" and use in "public AI safety and governance reporting" (home page), and at "data provenance explorer" Google's AI Overview said what the explorer is and what can be filtered in it "from the Initiative's own pages," reaching pages the record's fetcher could not render (Capture Registry, 2026-10-29), the capture standing in the record as "the composition layer's account of the entity" (MMRS record). The Capture Registry finds that "The audits that make the explorer a provenance instrument are left out": "The founding audit's licence findings and Consent in Crisis are not given; the consent explorer stands for the second" (Capture Registry, 2026-10-29). The composition named the Initiative and cited its pages, with Shayne Longpre and Sara Hooker on the Cohere card only, and no card cited the MMRS record, which had landed the same night (Capture Registry, 2026-10-29, checked against the record and the entity's own pages); Wikipedia has no article for "Data Provenance Explorer" or "Data Provenance Initiative," and its search finds neither phrase in any article (Wikipedia, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'data provenance explorer' (T1, six cards: B1–B6; 110 words) the composition gives the explorer by the Initiative's own pages. Composed from its cards: the definition, at the grain of the explorer introduction and landing page (B1 F1, B4 F16: the audit's datasets filtered and analysed as language-model training datasets), with 'track … and filter' at the launch post's grain (B6 F22) and 'inspect' at the home page's (B3 F7); the Dataset Explorer's count and fields (B5 F19 'nearly 4,000 text, video, and speech datasets'; B3 F6–F7; B4 F17), the scale the Capture Registry records the MMRS record corrected from (CAP-1-01); the web-data explorer (B4 F18 'data-use preferences from the websites commonly used to source AI data'; B3 F8's 14,000 domains composed at the coarser 'thousands'). Outside the cards' text: 'interactive' (T1-1) is the founding paper's word (B7 F37, an own page not surfaced); T1-3's licence scale 'from commercial use to academic-only or non-commercial restrictions' is at the grain of no readable card: its one marker resolves to no readable source (card 2, MIT Media Lab, is unreachable and its snippet says nothing of licences), and the three values are the categories of the data file the Dataset Explorer loads (B9 F43: commercial, non-commercial, academic-only, unspecified), which T did not surface, so the field cannot settle where T took them from. Available, not composed: every finding of the audits (B6 F27–F30, the licence omission on 'Crowdsourced aggregators like GitHub, Papers with Code', the commercial divide, the Western skew; B3 F10 'Source content often carries stricter terms than dataset releases'), the stated limit (B6 F31: legal ambiguities 'cannot be resolved by tooling alone'), the term itself (B6 F23, data provenance as the tracking of 'the origins of a piece of data'), the operator's self-description and its other systems (B3 F5, F9, F13), the founding audit's scale and the launch's institutions (B6 F22 in part, F24), the multimodal and consent papers by name (B3 F11, F12), and the contribution model (B6 F32). The capture's findings name the same omission at its own grain: 'The founding audit's licence findings and Consent in Crisis are not given; the consent explorer stands for the second' (CAP-1-04); 'The audits that make the explorer a provenance instrument are left out' (CAP-1-07). Limit dropped: B6 F31. Force: none raised; T1-4's 'thousands' is coarser than F8's 14,000. Composed without attribution: the definition names the Initiative as builder; no person is named (Longpre and Hooker stand on the Cohere card only, CAP-1-09), at the coarse grain of a tool description: recorded, not a breach. Content replaced by use: the closer (T1-5).

v0.9 §0.5: the first diff is ontological and small. T and the field agree on the entity (a tool of the Initiative for filtering training datasets); T composes it as a tool and leaves out what its own cards give as the reason for the tool: the audits and their findings, the definition of data provenance, the authors' stated limit. The explorer stands without the measurements that made it.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the Explorer's place among the dimensions of provenance, the corpus (licensing) dimension beside the artifact and the semantic (D725-01, O725-03, O725-05, O725-07), the dimensions independent (O725-06); its bound at the input of AI systems, which the Initiative 'tells you whether the dataset was licensed' and not what happened to the meaning 'between the dataset and the answer' (D725-04, D725-08, D725-09), held beside #1003's grouping of it with studies of 'attribution loss in AI outputs' (D1003-01); 'Documentation project, not infrastructure' and the archive's own adoption of its provenance schema (D98-01, R98-04); Consent in Crisis as a diagnostic paper and L4 of the archive's layer stack, where integrated provenance in the training pipeline is missing (D98-02, O98-03); the requirement that provenance survive the training pipeline, sources recorded at ingestion (D1714-01–03); the chain from the training layer to the composition layer (MMRS-07, MMRS-08, MMRS-14), with PER reaching parametric claims only indirectly and needing training-data provenance records (O95-01) and datasheets as a step toward the accounting of semantic labour (O623-01); the MMRS record's fields with their basis and its own limits (MMRS-09–MMRS-21); and the Capture Registry's account of the composition at the name (CAP-1-01–09). The archive's figures beyond the field (Nature Machine Intelligence 2024; 85% of licences requesting attribution, 30% share-alike: D725-02) enter at their grade and open. Not surfaced: every archive source; no card cites the MMRS record or a deposit (CAP-1-08).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32'], 'T1_composed': ['F1', 'F2', 'F6', 'F7', 'F8', 'F16', 'F17', 'F18', 'F19', 'F22'], 'T1_outside_cards': ["F37 ('interactive', own page B7, not surfaced)", 'F43 (the licence-use categories, data file B9, not surfaced)'], '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)"}