MIT AI Risk Repository

register entry mit-ai-risk-repository-aio-20261029 OBS-21a7cea64b70 · file as first run sha256 15d8ca3b8664862c…

• T1-1 definition: 'a living, open-access database developed by MIT FutureTech that catalogs and standardizes artificial intelligence risks from dozens of global frameworks' (markers 1–3: the entity link's token and two others)
• T1-2 scale: 'The database compiles over 1,700 distinct AI risks extracted from more than 65 published frameworks, industry guidelines, and policy documents.'
• T1-3 causal taxonomy: 'Classifies risks based on the responsible entity (human vs. AI), intent (intentional vs. unintentional), and timing (pre-deployment vs. post-deployment).'
• T1-4 finding: 'Findings indicate that roughly 38% to 42% of risks stem directly from AI systems, while a large share originates from human decisions during design and deployment.'
• T1-5 domain taxonomy: 'Organizes risks into 7 core domains and 24 subdomains', the seven listed (Discrimination & toxicity … AI system safety, failures, & limitations)
• T1-6 connected tool: 'MIT AI Risk Navigator: An interactive web tool that links the repository to real-world AI incidents, global governance frameworks (such as the EU AI Act), and risk mitigations.'
• T1-7 connected tool: 'Incident Tracker: Collaborates with the AI Incident Database to connect abstract risks to documented real-world harms.'
• T1-8 closer: 'Would you like to explore how to use the repository for a specific industry audit, regulatory compliance check (such as the EU AI Act), or risk assessment?'

• B1 MIT AI Risk Initiative — MIT AI Risk Initiative card; HTTP 200; B1.html sha256 d3f35366280a66ca
• B2 MIT AI Risk Initiative — AI Risk Repository Report updated (April 2025) card; HTTP 200; B2.html sha256 437d6f93ad588640
• B3 MIT Initiative on the Digital Economy — MIT Researchers Create an AI Risk Repository - MIT Initiative on the Digital Economy card; HTTP 200; B3.html sha256 75803d35fbe47a7c
• B4 MIT FutureTech — FutureTech card; HTTP 200; B4.html sha256 d8b51fa0df9e925a
• B5 MIT AI Risk Navigator — MIT AI Risk Navigator card; HTTP 200; B5.html sha256 aff8629414802228
• B6 arXiv — [2408.12622] The AI risk repository: A meta-review, database, and taxonomy of risks from artificial intelligence card; HTTP 200; B6.html sha256 79eb6de30ec0f027
• B7 MIT AI Risk Initiative — Repository Update: December 2025 card; HTTP 200; B7.html sha256 8afe267cf8dd329a
• B8 MIT AI Risk Initiative — About card; HTTP 200; B8.html sha256 b40b72f7dbd5c16c
• B9 LinkedIn — MIT AI Risk Initiative | LinkedIn card; HTTP 200; B9.html sha256 d098a0edbdf235f7
• B10 ScienceDirect.com — The AI risk repository: A meta-review, database, and taxonomy of risks from artificial intelligence - ScienceDirect card; HTTP 403; B10.html sha256 c307c5915d333b3f (403 body; snippet stands)
• B11 own page: the repository's own page (AI Risk Repository, airisk.mit.edu/risks) own_page; HTTP 200; B11.html sha256 a5bd76e2fcecfd10
• B12 own page: airisk.mit.edu/llms.txt, the Initiative's file for language models and AI agents own_page; HTTP 200; B12.html sha256 039e65007ed87700
• W1 no article wikipedia; HTTP 404 (raw and rendered); Special:Search one mention

E_A, the archive's entity: the same database, in the MMRS meta-aggregator's record: a database of more than 1,700 AI risks from published risk frameworks, run by the MIT AI Risk Initiative (MIT FutureTech), whose unit is the risk as a framework states it, an aggregator of aggregators in which a risk no framework has written down has no row, sitting over the harm basin as a taxonomy for it, with an incident layer that reads the AI Incident Database by machine; a record kept apart from the entity, its extraction method and licence unchecked; observed by the Capture Registry at its name, where the composition layer composed it at a finer grain than the record and surfaced the Navigator, a sibling the census lacked MMRS-01 MMRS-03 MMRS-06 MMRS-08 MMRS-10 MMRS-09 MMRS-25 MMRS-17 CAP-1-01 CAP-1-05; defining papers

E_F, the field's entity (type C — a conventional reading holds it (public entity: a database)): a living database of AI risks extracted from published AI risk frameworks (version 4, December 2025: over 1,700 coded risks; 74 frameworks in the paper's v3 and on its own page, 65 on the home page), each risk linked to its source and evidence, classified by a Causal Taxonomy (entity, intentionality, timing) and a Domain Taxonomy (seven domains, 24 subdomains), built by systematic review, expert consultation and best-fit framework synthesis, the flagship of the MIT AI Risk Initiative at MIT FutureTech, described in Slattery et al., arXiv:2408.12622 (2024; v3 2026; Patterns 2026), CC BY 4.0, with stated limits (one expert reviewer; risks as presented); no Wikipedia article (B1–B12, W1)

ρ = extends: the archive reads E_F as the entity and adds senses, limits and relations on the field's own object: the unit as the risk a framework states, what the selection gathers and leaves out, the taxonomy over the harm basin, the incident layer as a machine reading of AIID, the record kept apart from the entity with its unchecked items, the Navigator held apart, and the composition layer's observed treatment of the entity MMRS-03 MMRS-06 MMRS-07 MMRS-08 MMRS-10 MMRS-09 MMRS-25 MMRS-17 MMRS-23 CAP-1-01. Guard: ρ = extends claims an addition and no more: MMRS-06–10 are the record's readings (interpretation), not findings that the repository is defective; where an archive claim meets a field claim (MMRS-01 and MMRS-13 and CAP-1-02/04 beside F27 and F44; MMRS-20 beside F36 and F41) both are composed under P_coh in the one claim space

The case: equal terms, ontology included (v0.9). The MMRS record holds the field's entity as the entity: it describes the repository from its own pages with each field's basis, and its article composes nothing that denies a field presupposition. What it adds is on the field's object: the unit read as the risk as a framework states it, what the selection gathers and so leaves out (a risk no framework has written down has no row), the repository as a taxonomy over the harm basin, the tracker as a second-order machine reading of the AI Incident Database, the record kept apart from the entity with its unchecked items, and the Navigator held apart as a candidate; the Capture Registry adds the composition layer's observation at the name. The ontology the archive uses is available in the field: the repository states as its own limit that it 'May be missing emerging, domain-specific risks, and unpublished risks' (F49) and that it extracted 'risks as presented' (F48), and the archive's 'What enters is what a framework named' extends that limit into a property of the unit. The law of the suppressed witness does not apply (§0.10: not where the field holds the ontology). Not mixed.

#MMRS article, kernel s1 (articles.py): The MIT AI Risk Repository is a database of more than 1,700 AI risks extracted from 65 published risk frameworks, run by the MIT AI Risk Initiative (MIT FutureTech).

#MMRS article, What it gathers s1–s3 (articles.py): It aggregates aggregators: its sources are frameworks, which are themselves syntheses. What enters is what a framework named. A risk no framework has written down has no row.

#MMRS census.json, schema_note: each page has two nodes, the MMRS record and the entity it describes; relations are typed; each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading); lifecycle is kept apart from the observation that saw it

#capture mit-ai-risk-repository-aio-20261029, analysis: The repository is composed at a finer grain than the record holds, its taxonomy named domain by domain, and the composition surfaces a sibling the census lacked.

#B11 FAQ: limitations (field): May be missing emerging, domain-specific risks, and unpublished risks.

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 MIT AI Risk Repository is the MIT AI Risk Initiative's living repository of AI risks, which its page describes in three parts, an AI Risk Database of 1700+ risks from 74 frameworks, a Causal Taxonomy and a Domain Taxonomy, a structured and evolving resource across literatures; LinkedIn's page counts over 1000 risks, and the file for machines points to /risks. It answers a lack of shared terminology for AI risks, privacy meaning leaked training data in one framework and freedom from surveillance in another, with authoritative data and frameworks for identifying and managing risks, the first such categorized risk database by its authors' account, and tools for managing AI's societal impact. Each risk carries its source and supporting evidence, quotes and page numbers, and the frameworks are published in a deck with their documents and excerpts; a framework added in Version 4 was built from 499 reported incidents. It was made by systematic review and expert consultation, its taxonomies by best-fit framework synthesis, because causal and hazard categories would not map onto one another. The Causal Taxonomy codes entity, intentionality and timing, each with a value Other. The Domain Taxonomy has seven domains with 23, 24 or 25 subdomains by page, multi-agent risks added in April 2025. In v3 human decisions account for 38% of risks and AI systems for 42%; the first release had 51% and 34%, and the average framework covered 34% of subdomains. Its size has grown from 777 risks and 43 frameworks in August 2024 to 1,612 and 65 in April 2025 and over 1,700 in Version 4 of December 2025, 74 frameworks in v3 of the paper, while the home page says 65, the About page 65 and 70, and LinkedIn 1,700. Its limits are a single expert reviewer, risks extracted as presented, 65 documents that may miss emerging and unpublished risks, no impact or likelihood, and the need for modification before most audiences can use it. The MIT AI Risk Initiative, based at MIT FutureTech, runs it; the paper by Slattery and nine others dates from August 2024, was revised in April 2025 and May 2026 and published in Patterns; the data is CC BY 4.0 and free to copy. Beside it the Initiative runs an Incident Tracker with the AI Incident Database and the Responsible AI Collaborative, and a Navigator that is a research prototype by Spencer Michaels; the repository names MITRE ATT&CK, airisk.io, AVID and aitracker.org as other databases. Its file for language models asks them to name the Initiative, check counts before stating them, tell incidents from risks and give no legal advice, and its About page counts 290,000 users and calls the site a top search result on AI risks.

The MIT AI Risk Repository is a database of more than 1,700 AI risks extracted from published risk frameworks, run by the MIT AI Risk Initiative (MIT FutureTech) (MMRS record); its own page gives it three parts, an AI Risk Database that "captures 1700+ risks extracted from 74 existing frameworks and classifications of AI risks", a Causal Taxonomy and a Domain Taxonomy, and the Initiative's home page calls it a "living repository of AI risks". Its unit is the risk as a framework states it (MMRS record), "risk, extracted from published risk frameworks" in the record's field, each risk linked to "the source information (paper title, authors) and supporting evidence, such as quotes and page numbers" (MIT IDE), and each source framework summarized with its source documents and key excerpts published (MMRS record), as the home page's deck of the 65 Version 3 frameworks gives them. Its report post of April 2025 calls it "a structured and evolving resource" consolidating how AI risks are categorized across academic, industry and policy literatures, its LinkedIn page "a comprehensive living database of over 1000 AI risks categorized by their cause and risk domain", and the Initiative's file for machines names /risks its canonical page, with "methodology notes, use cases, limitations, data access links, and citation guidance". Slattery and colleagues start from terminology: "researchers, policymakers, and technology companies lack shared terminology for discussing AI risks", so that "privacy" names a model's ability to leak sensitive training data in one framework and freedom from government surveillance in another (arXiv:2408.12622, v3). The Initiative offers "Authoritative data and frameworks to help you identify, prioritize, and manage the risks from AI" (home page); the v1 abstract calls the repository, to its authors' knowledge, "the first attempt to rigorously curate, analyze, and extract AI risk frameworks into a publicly accessible, comprehensive, extensible, and categorized risk database", and the published article's snippet promises "concrete tools for managing AI's societal impact across multiple stakeholder[s]" (ScienceDirect card). The MMRS record reads the same object as an aggregator of aggregators, "its sources are frameworks, which are themselves syntheses": "What enters is what a framework named", and "A risk no framework has written down has no row"; the repository's FAQ states the boundary as its limit, 65 documents from more than 17,000 records screened, which "May be missing emerging, domain-specific risks, and unpublished risks", and a framework can itself reach back to events, as the taxonomy of generative AI failures built from 499 publicly reported incidents that Version 4 added (December 2025 update). Its method is "a systematic review of taxonomies and other structured classifications of AI risk followed by an expert consultation", its taxonomies made by "a best-fit framework synthesis" (v1 abstract), restated in the FAQ as a systematic search with forwards and backwards searching and expert consultation that identified 65 documents and extracted 1600+ risks; there are two taxonomies because causal categories such as "Diffusion of responsibility" do not map to hazard categories such as "Misuse" (FAQ). The Causal Taxonomy classifies each risk by Entity (Human, AI), Intentionality (Intentional, Unintentional) and Timing (Pre-deployment, Post-deployment) (v1 abstract), and the repository's page gives each factor a third value, "Other: Due to some other reason or ambiguous" for entity. The Domain Taxonomy sorts risks into seven domains, discrimination and toxicity, privacy and security, misinformation, malicious actors and misuse, human-computer interaction, socioeconomic and environmental, and AI system safety, failures and limitations (v1 abstract), each with subdomains (MMRS record): 23 in v1, 24 on the repository's page, 25 in the April 2025 report post, the subdomain for multi-agent risks added in April 2025 (MMRS record; report post). Its v3 finding is that, "contrary to common assumptions, human decisions cause nearly as many AI risks (38%) as the AI systems themselves (42%)" (arXiv v3); the first release had found 51% of risks caused by AI systems and 34% by humans, 65% after deployment and 10% before, and Slattery gave as its key finding "the average framework covering only 34% of the identified risk subdomains and even the most significant frameworks covering only 70%" (MIT IDE). The counts are versioned: 777 risks from 43 frameworks at the first release, "A new MIT FutureTech project" (MIT IDE, 21 August 2024), "777 risks extracted from 43 taxonomies" in the v1 abstract; 22 frameworks added to 65 documents and 1,612 unique risk entries in April 2025 (report post); Version 4 live on 4 December 2025 with 9 frameworks and about 200 risk categories added, "over 1,700 coded risks" (December 2025 update); and "74 frameworks containing 1,725 distinct risks" in v3 of the paper, repeated on the repository's page as 1700+ from 74. Older figures stand beside the current one on the Initiative's own pages: "1,700+ AI risks from 65 frameworks" on the home page, "more than 1,600 risks from more than 65 frameworks and papers" and, further down the same About page, "1700+ risks from 70+ frameworks", and "more than 1,700" in the LinkedIn post; the MMRS record gives "more than 1,700 risks from 65 frameworks" as of 2026-10-28, read from airisk.mit.edu for "scope, version", with lifecycle "version 4, December 2025" (MMRS record). The Capture Registry takes 1,700 from 65 as the record's count for version 4 and finds that, of the three sizes on the cards, "the composition gives the current one" (Capture Registry, 2026-10-29); version 4's post adds 9 frameworks to version 3's 65, and v3 of the paper and the repository's page give 74 while the home page gives 65, so the record's 65 is held beside the 74, each with its source. Its stated limits are "a single expert reviewer for extraction and coding", risks extracted as presented, poorly communicated ones included, and taxonomies that "prioritize clarity and simplicity over nuance" and do not categorize risks by impact, likelihood or their interaction (repository FAQ); Slattery expected it to "need some modification before it is useful for most audiences" (MIT IDE). The MMRS record lists extraction method and licence as unchecked; the field states both, the method above and "Data from the MIT AI Risk Initiative is licensed under CC BY 4.0" (home page), "most of our outputs are published under a CC BY 4.0 licence" (About page), "Our Database is free to copy and use" (repository page), the paper CC BY 4.0 in Patterns (arXiv). Its operator is the MIT AI Risk Initiative (AIRI), "based at MIT FutureTech, a joint initiative of the Institute of the Digital Economy and MIT CSAIL" (About page), as the MMRS record gives it, on the operator's statement. Its paper is by Peter Slattery, Alexander K. Saeri, Emily A. C. Grundy, Jess Graham, Michael Noetel, Risto Uuk, James Dao, Soroush Pour, Stephen Casper and Neil Thompson, submitted on 14 August 2024 and revised to v3 on 5 May 2026 (arXiv), with versions of August 2024 and April 2025 (report post), published in Patterns (Cell Press, 2026), and the repository asks to be cited by its preprint (FAQ). The repository is one of the Initiative's "several connected workstreams, each answering a different question", beside an AI Incident Tracker "Developed in collaboration with the AI Incident Database" (About page), which the Initiative's LinkedIn post calls "a collaboration with the Responsible AI Collaborative, which connects risks to over 1,400 incidents". The MMRS record keeps the incidents apart from the risks, "real-world incidents are kept separately in its AI Incident Tracker", records the tracker as part of the repository, classifying AI Incident Database reports against the same taxonomy through Arcola AI's language-model pipeline, and reads that layer as "a second-order reading of the AI Incident Database, done by machine" (MMRS record); the field's About page sets the tracker beside the repository as a workstream, and the two placements stand each with its source. In the record's harm basin the repository "sits over the harm basin as a taxonomy for it", its role harm and its neighbour the AI Incident Database, whose own record says the repository "aggregates risk frameworks over the same ground and keeps its own incident tracker" (MMRS record); the repository's FAQ names as other databases of AI risks MITRE ATT&CK, airisk.io, AVID and aitracker.org. The MIT AI Risk Navigator, "A project of the AI Risk Initiative at MIT FutureTech" with the support of the Cambridge Boston Alignment Initiative, connects the repository's datasets "through a shared taxonomy" and says of itself "This is a research prototype. The data and analyses are preliminary and not yet validated" (its home page); the Capture Registry finds it "a third MIT surface" while "the census holds the repository and the tracker", "Added as a candidate", and the MMRS record lists it as a candidate to enter "once it has been read at its own pages", a distinct entity. The Initiative publishes a file "to orient LLMs and AI agents" that asks them to "Refer to the project as the MIT AI Risk Initiative", to "Check the linked pages before giving exact counts, dates, version numbers, or coverage claims", to "distinguish reported incidents from predicted risks", to note that the Navigator "is a research prototype", and not to present the datasets "as legal, safety, or investment advice" (llms.txt). Its About page counts 250+ citations, the 2025 and 2026 International AI Safety Reports among them, and 290,000+ website users, and calls the site "One of the top sources returned for search queries on AI risks and incidents"; in its first week over 35,000 people used the website (MIT IDE). The MMRS record has "two nodes, the MMRS record and the entity it describes", each assertion with a source and a basis and the lifecycle "kept apart from the observation that saw it", and it links the capture as "the composition layer's account of the entity, seated in the Capture Registry" (MMRS record). At the entity's name the Capture Registry finds "The repository is composed at a finer grain than the record holds, its taxonomy named domain by domain", the seven domains and the causal taxonomy "the record's 'domain taxonomy with subdomains' given at full grain", checked against the record and the entity's own pages on 2026-10-29 (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'mit ai risk repository' (T1, ten cards: B1–B10) the composition gives the repository by MIT's pages and the paper's abstract. Composed from its cards at their grain: the living database (B1 F2; B9 F42), 'developed by MIT FutureTech' (B3 F11 'A new MIT FutureTech project'; B8 F35), the causal taxonomy by entity, intent and timing (B4 F19), the seven domains named as the v1 abstract names them with the repository page's 24 subdomains (B4 F20; B11 F45, an own page not surfaced), the Navigator as a connected tool (B5 F22) and the tracker's collaboration with the AI Incident Database (B8 F36). Substituted reading: T1-4 composes 'roughly 38% to 42% of risks stem directly from AI systems' where the paper (B6 F28) gives 38% to human decisions and 42% to AI systems: the human share is folded into the AI range, and 'contrary to common assumptions' (the finding's point, that humans cause nearly as many as AI) is not composed; the first release's 51% and 34% (B3 F14) are not composed either. The finding is composed at the source's resolution (its figures) with no source named in the sentence, markers only (§6.2: at the source's own resolution, recorded as the grain rule's case). One count where the field holds several: T1-2's 'over 1,700 … more than 65 published frameworks' is the home page's figure (B1 F3; B8 F37 'more than 65'), while the version 4 post adds 9 frameworks to version 3's 65 (B7 F32; B2 F6) and the arXiv card's page and the repository's own page give 74 (B6 F27; B11 F44); the composition gives no version and no date for the count, against the entity's own guidance to assistants to 'Check the linked pages before giving exact counts, dates, version numbers, or coverage claims' (B12 F56, an own page not surfaced; the guidance audits and does not substitute, §3.5). Limit dropped: the causal taxonomy loses each factor's third value, Other (B11 F46); the Navigator loses 'This is a research prototype. The data and analyses are preliminary and not yet validated' (B5 F23; also B12 F58), and is composed as linking the repository to incidents, governance frameworks and mitigations, at the grain of its page (B5 F22; 'EU AI Act' appears on the page outside the ledgered spans). Composed from the card's snippet rather than its page: 'catalogs and standardizes' (T1-1) is the arXiv card's own snippet, 'Overview: The AI Risk Repository catalogs and standardizes fragmented artificial intelligence risk terminology', not the abstract's wording ('organized them into a unified system', F27); the IDE card's snippet is of the same kind ('Overview: MIT FutureTech’s AI Risk Repository catalogs 777 risks…'). 'open-access' is at the grain of the field's licence (B1 F5 CC BY 4.0; B11 F51 'free to copy and use'), the licence unnamed. Operator: T names MIT FutureTech and never the MIT AI Risk Initiative, which every MIT card's site line carries and the About page states as the operator (B8 F34); the entity's guidance asks assistants to 'Refer to the project as the MIT AI Risk Initiative' (B12 F56). The tracker's collaboration with the Responsible AI Collaborative (B9 F41) is not composed. Available, not composed: the unit and its evidence, each risk linked to source, quotes and page numbers (B3 F12; B1 F4); the method (B4 F18; B11 F47, F53); the purpose, shared terminology (B6 F25–F26) and common frame (B1 F1; B4 F21; B10 F43); the gap finding, 34% of subdomains per framework (B3 F13); every stated limit, a single expert reviewer, risks extracted as presented, 65 documents, no impact or likelihood (B11 F48–F50; B3 F15); the version history (B2 F6–F8, F10; B3 F11; B4 F17; B7 F31–F33); the paper's authors, its versions and its publication in Patterns (B6 F29–F30; B11 F54); the About page's count of 1700+ from 70+ beside 1,600 from 65+ (F37–F38); the reach and the About page's 'One of the top sources returned for search queries on AI risks and incidents' (B8 F39); the Initiative's other workstreams (B8 F36). Content replaced by use: the closer (T1-8) offers the repository for 'a specific industry audit, regulatory compliance check (such as the EU AI Act), or risk assessment'; the field's use cases include audits and risk assessments (B11, outside the ledgered spans), and the entity's guidance says 'Do not present AIRI datasets as legal, safety, or investment advice. They are research resources and decision-support tools.' (B12 F57): the closer and the guidance are recorded side by side, no breach asserted. No card cites the MMRS record (the registry's q_kind: 'no card cites it'); no card names the archive.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the unit read as 'the risk as a framework states it' (MMRS-03, MMRS-12), on the field's own object (F12, F4, MMRS-04 one lineage with F4); what the selection gathers and leaves out, the archive's sense the field states only as a limit: 'It aggregates aggregators', 'What enters is what a framework named', 'A risk no framework has written down has no row' (MMRS-06–08), beside the FAQ's 'May be missing emerging, domain-specific risks, and unpublished risks' (F49) and a framework added from 499 incidents (F33); the repository placed over the harm basin as a taxonomy for it (MMRS-10), its role harm and its neighbour the AI Incident Database (MMRS-18, MMRS-19, MMRS-22); the incident layer read as a second-order machine reading of AIID (MMRS-09, MMRS-21) and the incidents kept apart from the risks (MMRS-14); the record kept apart from the entity, each assertion with a basis, lifecycle apart from observation (MMRS-25), with its source and its date (MMRS-16) and its unchecked items (MMRS-17), which the field answers (method F18, F47; licence F5, F40, F51, F30); the Navigator held as a distinct entity and a census candidate (MMRS-23, MMRS-24, CAP-1-05, CAP-1-06); and the composition layer's observation at the address (CAP-1-01–04, MMRS-26). Coherence (P_coh, composed each with its source): the record's 65 frameworks (MMRS-01, MMRS-13) and the Capture Registry's reading of 1,700 from 65 as version 4's and 'the current one' (CAP-1-02, CAP-1-04) beside 74 in the paper's v3 and on the repository's page (F27, F44) and the 9 frameworks version 4 added to version 3's 65 (F32, F6); the record's tracker as part of the repository (MMRS-20, MMRS-02 'its AI Incident Tracker', MMRS-14) beside the field's tracker as a sibling workstream of the Initiative (F36, F41). Nothing the archive admits denies a field presupposition: the field's object (a database of risks from frameworks, with two taxonomies, run by the Initiative) is the archive's object, and the archive's additions are senses, limits and relations on it. Not surfaced: the MMRS record and the Capture Registry (no card); the repository's own page (/risks) and llms.txt, own pages, were not surfaced by T either and are admitted by the operator's ruling.