AI Incident Database (AIID; Responsible AI Collaborative)

register entry ai-incident-database-aio-20261029 OBS-d2d81aafd207 · file as first run sha256 d24b458664f3be0e…

• T1-1 definition: 'a free, open-source project managed by the Responsible AI Collaborative that indexes real-world harms, near-harms, and failures caused by artificial intelligence systems' (markers: the database's card and the Collaborative's link)
• T1-2 model: 'Modeled after safety incident reporting systems used in aviation and cybersecurity, it collects public and crowdsourced reports to help the industry learn from past failures.'
• T1-3 scope: 'algorithmic bias, discrimination, privacy leaks, misinformation, deepfakes, autonomous vehicle accidents, and safety-critical system failures'
• T1-4 growth: 'Incidents tracked by the database have surged significantly, with deepfake and scam-related entries multiplying rapidly alongside broader media scrutiny.' (markers: an unresolved token and the TIME clip)
• T1-5 derived layer: 'Projects like the MIT AI Incident Tracker parse the database using taxanomies and LLM pipelines to categorize risks into structured analytical views.'
• T1-6 closer: an offer to search incident types or 'learn how to submit an incident report'
• T2-1 definition: 'an alleged harm or near-harm event to people, property, or the environment where an AI system is implicated'
• T2-2 criteria: 'an adaptive, specific set of criteria outlined in its AIID Editor's Guide'; 'For a report to be accepted as a true incident record, it must meet three core baseline thresholds'
• T2-3 specific event: 'a distinct, real-world occurrence with clear details on what happened, to whom, and where. Theoretical vulnerabilities do not qualify.'
• T2-4 AI implication: 'the harm likely would not have occurred without the AI's output, omission, or use'
• T2-5 harm or near-harm: 'an adverse effect or a credible chain of events nearly producing one'
• T2-6 categories: 'Using frameworks like CSET's taxonomy, incidents are grouped into categories such as physical safety, discrimination and fairness, financial loss, privacy and security breaches, and misinformation or scams.'
• T2-7 distinctions: incident 'A real event where AI caused or nearly caused harm'; issue 'A potential harm or vulnerability that has not yet materialized in the real world'; variant 'A separate occurrence involving the same AI system, similar causes, and matching outcomes as an existing record'
• T2-8 closer: 'Are you trying to submit a specific case to the database, or are you researching a particular type of AI safety failure … for your own project?'

• B1 AI Incident Database — AI Incident Database card; B1.html sha256 597e795095450ff7
• B2 MIT AI Risk Initiative — MIT AI Incident Tracker card; B2.html sha256 720c3b26f04adc57
• B3 Partnership on AI — AI Incident Database - Partnership on AI card; B3.html sha256 a88664f05e53b9af
• B4 OECD AI Policy Observatory — AI Incident Database card; B4.html sha256 1c3e690c98ec848c
• B5 Kaggle — AI Incidents, Failures & Controversies Database - Kaggle card; B5.html sha256 75d1aa8d8ae2a5f1
• B6 LinkedIn·TIME — AI Incident Database | TIME posted on the topic card;
• B7 AI Incident Database — Editorial and Research Guide for the AI Incident Database card; B7.html sha256 ef7437bf3273bde0
• B8 AI Incident Database — Defining an “AI Incident” card; B8.html sha256 00e4d731859c8bdd
• B9 LinkedIn·TIME — AI Incidents Rise 50% Year-Over-Year: AI Incident Database | TIME posted on the ... card;
• B10 YouTube·HarrietHacks — Examples of AI real attacks card; B10.json sha256 5a32c03de9d2b551
• B11 AI Incident Database — The details about who and why for database - AI Incident Database card; B11.html sha256 40f7355ba917c998
• B12 The Association for the Advancement of Artificial Intelligence — Lessons for Editors of AI Incidents from the AI Incident Database - AAAI Publications card; B12.html sha256 b2bcccb8fa3e63d8
• B13 ACM Digital Library — Why We Need to Know More: Exploring the State of AI Incident Documentation ... card; B13.html sha256 661699817f917fdd
• B14 YouTube·International Association for Safe & Ethical AI — Establishing Effective Serious Incident Reporting Frameworks – Jimmy Farrell card; B14.json sha256 5a063beb92dbc849
• B15 AI Incident Database — Database Backups and Snapshots own_page; B15.html sha256 8d173133a5386515
• B16 AI Incident Database — Incident 693 own_page; B16.html sha256 bccaedd6532a09cd
• B17 AI Incident Database — Citation record for Incident 637 own_page; B17.html sha256 7e6fd184e23c3ebb
• B18 arXiv — [2011.08512] Preventing Repeated Real World AI Failures by Cataloging Incidents: The AI Incident Database own_page; B18.html sha256 264223e5c2e0ba22
• B19 Wikipedia — AI Incident Database (redirect) wikipedia; B19.wiki sha256 03d7f4dfc50b5315
• B20 Wikipedia — Partnership on AI wikipedia; B20.wiki sha256 08e1303706a2474f

E_A, the archive's entity: the same database, in the MMRS meta-aggregator's record: an editor-resolved database of AI harm events, run by the Responsible AI Collaborative, whose unit is the incident with its reports attached and a stable citation page per incident (the model the meta-aggregator follows), a database in the harm basin beside the OECD monitor and the AIAAIC, whose selection gives an answer surface's ordinary treatment of sources no unit, attribution entering only attached to a harm; observed by the Capture Registry at two addresses, where the composition layer held the entity at the grain of its own pages and left out its incident-and-report structure and its evidence test MMRS-01 MMRS-02 MMRS-16 MMRS-17 MMRS-14 MMRS-29 MMRS-30 CAP-1-02 CAP-1-05 CAP-2-01; defining papers #543

E_F, the field's entity (type C — a conventional reading holds it (public entity: a database)): a database dedicated to indexing the collective history of real-world AI harms and near harms, a project of the Responsible AI Collaborative, whose unit is the incident (an alleged harm or near-harm event in which an AI system is implicated) with reports attached as evidence, admitted by a five-question test, cited per incident at /cite/N and released in weekly snapshots; introduced by Sean McGregor in November 2020 with Partnership on AI, which Wikipedia makes its home article (B1–B20)

ρ = extends: the archive reads E_F as the entity and adds senses and limits on the field's own object: what its selection leaves out (no unit for an answer surface's ordinary handling of sources; no field for attribution), the citation page as a model, the record kept apart from the entity and the lifecycle from the observation, the licence's object, the neighbours and the homonym set apart, the index case, and the composition layer's observed treatment of the entity MMRS-14 MMRS-29 MMRS-16 MMRS-30 MMRS-23 MMRS-17 MMRS-25 MMRS-15 CAP-1-02 CAP-2-01. Guard: ρ = extends claims an addition and no more: MMRS-14 and MMRS-29 are the record's readings (basis 'reading'), not findings that the database is defective; MMRS-28 leaves report preservation unknown; where an archive claim meets a field claim (MMRS-01 beside F73; MMRS-12 beside F30) 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 AI Incident Database from the database's own pages, marking each field's basis (operator, observed, third-party, reading), and it composes nothing that denies a field presupposition; the citation page per incident is the model the meta-aggregator itself follows. What it adds is on the field's object: what the database's selection leaves out (an answer surface's ordinary handling of sources has no unit; attribution enters only attached to a harm), the record kept apart from the entity and the lifecycle from the observation, the neighbours and the homonym set apart, the licence's object, the battery; and the Capture Registry adds two observations of the composition layer at the entity's addresses, which found 'THE ENTITY HELD'. The ontology the archive uses is available in the field itself: the guide says the database is 'not a neutral mirror of all AI harms' (F32) and that its coverage depends on visibility (F33), and MIT that voluntary reporting gives sampling bias (F9); the archive's selection reading extends those stated limits to a class of event (no harm, ordinary composition) the field does not name. 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 AI Incident Database (AIID) is an editor-resolved database of AI harm events, run by the Responsible AI Collaborative since 2020.

#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

#MMRS article, What its selection leaves out s3: What an answer surface does with a source on an ordinary day, with no harm event, has no unit here: attribution erasure, citation fidelity and provenance stripping enter when they attach to a harm.

#MMRS article, Place in the field s1: Its citation page per incident is the model this meta-aggregator follows.

#capture ai-incident-database-aio-20261029, findings[0]: THE ENTITY HELD.

#B7 §11.1 (field): It is not a neutral mirror of all AI harms, a comprehensive denominator for risk rates, or an independent adjudication of every event.

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 11 (positions [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 AI Incident Database is dedicated to indexing the collective history of harms or near harms realized in the real world by the deployment of artificial intelligence systems, on its home page, a free and open-source project in the OECD's catalogue, a tool to anticipate the risks of AI in safety-critical systems collecting more than 1,200 reports on Partnership on AI's workstream page, and a project cataloguing incidents with a platform to classify them for Paeth and colleagues. Like similar databases in aviation and computer security it aims to learn from experience, from a founding problem McGregor's paper states as intelligent systems causing harms without a collective memory of their failings; the about page cites the transportation sector and computer systems, and the research page takes aviation's accident and incident as its model. It is a project of the Responsible AI Collaborative (home and about pages; the OECD entry; MIT's tracker), launched by Sean McGregor, its Executive Director, whose paper of 17 November 2020 describes an incident collection initiated by an industrial/non-profit cooperative over more than 1,000 reports and is the citation for the whole; the about page lists a Database Founding Sponsor and early contributors from the Partnership on AI. Wikipedia redirects the name to Partnership on AI, which announced it in November 2020 as a tool to identify, assess, manage and communicate AI risk and harm, introduced by McGregor as a systematized collection of incidents, listed it under its Safety Critical AI program and its NIST response, and says it shifted to a new special-purpose independent non-profit in 2022; the article is tagged for primary sources and as AI-generated. The Editor's Guide, Version 2 of 1 October 2026, defines the incident as an alleged harm or near harm event to people, property, or the environment where an AI system is implicated, the issue as one that has yet to occur or be detected, the incident ID as the editorial record of an event, the report as a source document attached to it, and the variant as a distinct incident with the same system and substantially similar causes and harms; the incident page groups variants under the first reported incident. Inclusion asks five questions, a specific event, an identifiable system, causal implication, harm or credible near harm and adequate evidence, under adaptive criteria and presumptive exclusions that are presumptions rather than prohibitions; anyone may submit when in doubt, and incident editors resolve submissions in a participatory project whose taxonomies come from CSET and GMF. Each incident has a suggested citation at /cite/N; duplicates are mapped to canonical records, as 637 to 112; weekly dumps run from March 2021 with an Excel export; the home page lists incidents to 1788 and Paeth and colleagues counted over 750 incidents and 3000 reports. The guide calls it a curated archive of publicly documented allegations, not a neutral mirror, denominator or adjudication, with coverage biased toward wealthy countries and English-language reporting, and absence of records not evidence of absence; Paeth and colleagues find epistemic uncertainty unavoidable, and MIT notes sampling bias in voluntary reporting. Its usage rights are open source/permissive under Apache 2.0 in the OECD's catalogue. Incident 693 records Google's AI search reportedly giving confidently incorrect and harmful information, dated 2024-05-14, its reports naming AI Overviews, machine-classified by MIT as false or misleading information. MIT's AI Incident Tracker classifies 1,600 of its incidents through Arcola AI's LLM pipeline; the OECD's monitor collects incidents automatically from news; a Kaggle dataset of 32 incidents draws on its records; and the rails carry an ACM paper by snippet, TIME clips and two videos by title.

The AI Incident Database (AIID) is an editor-resolved database of AI harm events (MMRS record), "dedicated to indexing the collective history of harms or near harms realized in the real world by the deployment of artificial intelligence systems" in its home page's words, which the MMRS record quotes and the OECD's catalogue repeats as a "free and open-source project", and which Paeth and colleagues (AAAI-25) describe as a project that catalogs AI incidents and provides a platform to classify them for operational and research goals. Its unit is the incident, with the reports that document it attached, and every incident has a stable citation page at incidentdatabase.ai/cite/N (MMRS record); the Editor's Guide defines the incident as "an alleged harm or near harm event to people, property, or the environment where an AI system is implicated" and the report as a source document attached to an incident ID, several of which may document one event, and the MMRS record quotes the definition and adds that reports may be news articles, papers or posts. Its home page sets it beside "similar databases in aviation and computer security", to learn from experience and prevent or mitigate bad outcomes; McGregor's founding paper starts from mature industrial sectors such as aviation that collect their failures in incident databases, while intelligent systems "cause real world harms without a collective memory of their failings"; the about page invokes the transportation sector (FAA and FARS) and computer systems, the research page "Defining an AI Incident" takes aviation's distinction of an accident (substantial damage or loss of life) from an incident (the risk of an accident substantially increased), and Partnership on AI's workstream page places it among safety-critical systems in transportation, healthcare and energy, a crowdsourced central repository of more than 1,200 reports. Sean McGregor's founding paper, submitted to arXiv on 17 November 2020, describes "an incident collection initiated by an industrial/non-profit cooperative", searchable by facet and full text over more than 1,000 incident reports, and on 18 November 2020 he introduced it on Partnership on AI's site as "a systematized collection of incidents where intelligent systems have caused safety, fairness, or other real-world problems" (the introduction quoted in Wikipedia's reference); the snapshots page asks that the database as a whole be cited by his IAAI-21 paper, and the MMRS record dates the founding to 2020 as the arXiv paper describes it, with more than 1,000 reports at launch. The home and about pages call it "a project of the Responsible AI Collaborative, an organization chartered to advance the AI Incident Database"; the OECD's catalogue entry (uploaded 25 April 2024) names the Collaborative as developing organisation and MIT's tracker credits it with creating and maintaining the database; the about page credits Sean McGregor with launching it and names him Executive Director, heads a logo block "Database Founding Sponsor" whose image the page labels Partnership on AI, and records that two people contributed to its early stages through their roles with the Partnership on AI; the MMRS record gives the same operator, founder, director and founding sponsor. Wikipedia has no article of its own for it: "AI Incident Database" redirects, with no section anchor, to Partnership on AI (a redirect from a subtopic, with possibilities), whose article records that Partnership on AI announced it in November 2020 as "a tool to identify, assess, manage, and communicate AI risk and harm", that Partnership on AI's response to NIST presented it as meeting attributes of the AI Risk Management Framework, that it was among the Safety Critical AI program's initiatives, and, under programs and initiatives, that it "shifted to a new special-purpose independent non-profit in 2022", the non-profit unnamed; the article is tagged for primary sources (April 2025) and as AI-generated (October 2025). The MMRS record has the database run by the Responsible AI Collaborative since 2020, and Wikipedia has it shift to an independent non-profit in 2022; the two dates stand each with its source. The Editor's Guide (Version 2, 1 October 2026, by Daniel Atherton and Sean McGregor, drafted with disclosed generative-AI assistance) keeps the incident apart from the AI issue, "an alleged harm or near harm by an AI system that has yet to occur or be detected", from the incident ID, the editorial record that bounds and summarizes an event, and from the report: "A report is evidence; an incident ID is the database’s normalized account of an event", so that a later source on the same event is usually an additional report, not a new incident. A variant is, in the guide, "a distinct incident involving the same intelligent system, substantially similar causative factors, and substantially similar harms as a known incident", as the MMRS record restates it; the incident page for 693 describes the variant as having the same causes, harms and system, grouped under the first reported incident, and adds that variants "do not need to have been reported outside the AIID". The Capture Registry (2026-10-29) records that at the database's name Google's AI Overview gave neither the incident as the unit, nor the reports attached to it, nor the citation page /cite/N, "collects public and crowdsourced reports" standing in for the unit, and that at the question of what counts the incident, the issue and the variant were the guide's definitions in substance, the first verbatim, and the report was again not stated; the MMRS record's negative tags hold that it is not a list of research papers, its unit the event and papers entering as reports. A candidate passes the guide's five-question test: a specific event, an identifiable AI system, a causal implication (an important role in the chain of events, not necessarily the sole or dominant cause), harm or credible near harm (a chain that could readily have produced harm, stopped by an external factor independent of the system and not a built-in safeguard functioning as designed), and adequate evidence, at least one source supporting the essential event (guide); the MMRS record states the same test and records that admission requires harm or credible near harm. At "what counts as an incident in the ai incident database" the Overview read the guide and kept its distinctions with one compression, the five-part test given as three thresholds, the identifiable system folded into "AI implication" and adequate evidence dropped, so that the composition carried the test's harm condition and lost its evidence condition, the condition on which the MMRS record's account of what its selection leaves out rests (Capture Registry, 2026-10-29). The research page "Defining an AI Incident" calls the criteria adaptive, reports "accepted or rejected on the basis of a growing rule set defined in the Editor's Guide"; the about page asks that an event be submitted when in doubt, the project meant to converge on shared criteria through the community's candidates, and says its initial set of more than 1,000 incident reports was intentionally broad, an autonomous car killing a pedestrian among its examples; the guide's presumptive exclusions include general criticism of a product or technology class without identified harm and benchmark failures, model errors or hallucinations that remain inside testing or private drafting, "presumptions rather than absolute prohibitions", which the MMRS record quotes as "benchmark failures that stay in testing" and "general criticism of a technology". Anyone may submit (MMRS record; the home page invites incident reports, to be indexed and made discoverable), and the incident editors, Daniel Atherton the lead, resolve submissions and maintain them; the database is managed in a participatory manner by those contributing code, research and broader impacts, CSET and Nikiforos Pittaras (GMF) have contributed its taxonomies (about page), MIT's tracker credits the Collaborative with reviewing all submitted reports, and the MMRS record's intake field states open submission with editors resolving submissions and maintaining records. Every incident has a suggested citation crediting its submitters and editors, at incidentdatabase.ai/cite/N (snapshots page); that citation page per incident is the model the MMRS meta-aggregator follows, and the MMRS record observes per item the incident id and citation page, title, date, alleged deployer and developer, harmed parties, attached reports, CSET and GMF classifications and variants. A duplicate is merged into the earlier record and its citation page stays up and points there (MMRS record): the citation record for 637 reads "This incident is a duplicate of Incident 112", its reports migrated, as the guide's duplicate mapping redirects a retired identifier to its canonical record; ids therefore include merged duplicates, and the highest id seen, 1788 (MMRS record, 2026-10-28; the home page on 2026-10-29), is an upper bound on incidents, against which Paeth and colleagues counted over 750 distinct incidents detailed by more than 3000 indexed third-party reports at their writing. The whole database is offered as weekly point-in-time dumps in JSON, MongoDB archive and CSV, listed from 29 March 2021 to 26 October 2026, with a weekly Excel export of numbered incidents and their taxonomy classifications (snapshots page; MMRS record, which counts the weekly snapshots as its version history); the guide warns that the live site can change after an export is created. The MMRS record keeps itself apart from the entity it describes, each assertion with a source and a basis (operator statement, third-party report, observation, reading) and the lifecycle (active) apart from the observation that saw it (2026-10-28), and it leaves the Wikidata item and any data licence beyond the snapshots page unchecked. The OECD's catalogue gives its usage rights as open source/permissive and its licence as Apache 2.0, without saying of what; the MMRS record observes the code open source and no data licence stated on the snapshots page, and the Capture Registry (2026-10-29) records that the Overview's "free, open-source" does not distinguish the two. The guide's governing principle is that the database "records publicly documented allegations of harm or near harm" and "does not independently adjudicate every factual claim"; it is "a curated archive of publicly documented incidents, issues, reports, and responses", not a neutral mirror of all AI harms, a comprehensive denominator for risk rates or an independent adjudication of every event; wealthy countries, prominent companies, spectacular events and English-language reporting are likely overrepresented, and harms to marginalized groups, inside private organizations or in ordinary administration may be underrepresented; and "A lack of records is not evidence that a sector or country has no incidents." Paeth and colleagues find that certain patterns of incidents present structural ambiguities that challenge incident databasing and that epistemic uncertainty in incident reporting is unavoidable; MIT's tracker states that, the reporting being voluntary, its input is subject to some sampling bias, and that a systematic validation study of its classifications has not yet been completed. On the MMRS record's reading a measured study enters as a report when it documents an event, and what an answer surface does with a source on an ordinary day, with no harm event, has no unit: attribution erasure, citation fidelity and provenance stripping enter when they attach to a harm, so the database has no field for attribution, and whether the reports' texts are archived beyond their links the record leaves unknown. The archive's Hexagonal Lexical Engine (#543, 10 March 2026) lists AI ethics glossaries, "AIAAIC, AI Incident DB", among the venues to which it addresses definitions with citations of three of its terms: semantic liquidation, ghost governance and taxonomic violence. Incident 693, "Google AI Reportedly Delivering Confidently Incorrect and Harmful Information", dated 2024-05-14 and edited by Daniel Atherton, alleges that Google developed and deployed the system and that Google users and the general public were harmed, carries reports naming Google's AI Overviews, and is machine-classified under MIT's taxonomy as false or misleading information (incident page); the MMRS record calls it the field's index case in the database, holding it with seven reports and the AI Overviews launch errors. MIT's AI Incident Tracker classifies 1,600 incidents from the database by risk, cause, harm, severity and other dimensions, its raw reports processed by Arcola AI's LLM pipeline against the MIT Risk Repository and a severity scale based on CSET's AI Harm Taxonomy (tracker page); the MMRS record lists the tracker as classifying AIID reports, and the MIT AI Risk Repository as aggregating risk frameworks over the same ground with its own incident tracker. In the harm basin, where the MMRS record gives it the roles of harm and monitoring, its neighbours are the OECD AI Incidents and Hazards Monitor, automated and news-derived (MMRS record; incident 693's page describes it as collecting and classifying incidents and hazards in real time from news sources), and the AIAAIC Repository; it is neither the OECD monitor nor a vulnerability database (MMRS record). The Kaggle dataset "AI Incidents, Failures & Controversies Database" (32 incidents, 29 companies, 13 countries, 24 columns, CC BY-SA 4.0), which takes primary incident records and AIID identifiers from the database and cites Stanford HAI's AI Index 2025 for a 56.4% surge in AI incidents in 2024, is another dataset with a near name (MMRS record), and at the database's name the Overview did not use it, its card beside the database's own (Capture Registry, 2026-10-29). The rails at the two addresses also carry an ACM paper read only by its snippet (submissions of incidents and controversies included by criteria such as relevance, fairness and accuracy), two TIME clips on LinkedIn titled "AI Incident Database | TIME posted on the topic" and "AI Incidents Rise 50% Year-Over-Year: AI Incident Database", and two YouTube videos admitted by title, "Examples of AI real attacks" and "Establishing Effective Serious Incident Reporting Frameworks – Jimmy Farrell | IASEAI 2025". At its name the Overview composed the database at the grain of its own about page and kept it apart from its neighbours, giving the operator, the aviation model and the MIT tracker as a derived layer, which are the MMRS record's fields, and omitting the incident-and-report structure that makes the database what it is; no card was machinemediation.org, the meta-aggregator's record not yet indexed (Capture Registry, 2026-10-29), whose battery asks that the operator be named, the unit stated with reports attached, the database not merged with the OECD monitor or the AIAAIC, and the citation page given (MMRS record).

T against L_F(B) (AIO's ontology at work on its own cards): At 'ai incident database' (T1, six cards: B1–B6) the composition gives the database by its own pages and its catalogue entry. Composed from its cards: the definition (B1 F1, at the grain 'harms, near-harms, and failures'), the operator (B1 F4; B4 F14; B2 F8), 'free, open-source' (B4 F12, the OECD's phrase), the aviation model (B1 F2, 'computer security' composed as 'cybersecurity'), 'public and crowdsourced reports' (B1 F3; B3 F10 'crowdsourced'), safety-critical failures (B3 F11) and the MIT tracker as a derived layer (B2 F6–F7). Partly in the field, outside the ledgered spans: T1-3's list of harm kinds (bias, privacy leaks, deepfakes, vehicle accidents) is at that grain in no database card; its nearest words are the Kaggle dataset's (B5: 'deepfakes influencing elections', 'autonomous cars killing pedestrians', the tags 'bias' and 'deepfakes'). Absent from the readable field: T1-4's growth claim ('surged significantly, with deepfake and scam-related entries multiplying') has its markers on an unresolved token and the TIME clip (B6, title only); the nearest readable figure, Kaggle's '56.4% in 2024' (B5 F16), is Stanford HAI's count of AI incidents in general, not of the database's entries: if T drew on it, the reading is substituted (a general statistic composed as the database's growth) and composed without attribution, which the field cannot settle. Available, not composed: the incident as the unit and the report attached (absent on the cards of T1, present on T2's: F19, F22–F23); MIT's stated limits (F9: sampling bias, no validation study); the OECD's Apache 2.0 licence without its object (F13), which T's 'free, open-source' does not distinguish from a data licence (CAP-1-07); Partnership on AI as the source of the workstream page (F10–F11), whose page names no Collaborative; the Kaggle dataset's provenance from the AIID (F15–F17), not composed and not confused (CAP-1-06). Content replaced by use: the closer (T1-6). At 'what counts as an incident in the ai incident database' (T2, nine cards: B2, B7–B14) the composition reads the Editor's Guide (B7) and composes its definition verbatim (F19; T2-1), the adaptive criteria (B8 F38; T2-2) and the distinctions of incident, issue and variant (F19, F20, F24; T2-7). Limit dropped: the five questions (F26) become 'three core baseline thresholds', the identifiable AI system folded into implication and adequate evidence (F27) dropped (CAP-2-04); the near-harm condition loses its external factor independent of the system (F29; T2-5 'a credible chain of events nearly producing one'); the issue loses 'or be detected' (F20; T2-7 'has not yet materialized'). Force raised: the variant's 'substantially similar harms' (F24) is composed as 'matching outcomes' (T2-7). Composed with attribution at the source's grain: the definition and thresholds carry 'its AIID Editor's Guide'. Partly sourced: 'Theoretical vulnerabilities do not qualify' (the guide's issue, F20, and its §3.2 table outside the ledgered spans); 'AI implication' as a but-for test (F28 states implication as an important role, not necessarily sole; the but-for question is the guide's §2.3 practical test, outside the ledgered span); CSET's taxonomy (F48) with categories the ledger does not carry at that grain. Available, not composed: the report as evidence attached to an incident ID and the many-to-one rule (F21–F23; CAP-2-05); the governing principle that the database records allegations and does not adjudicate (F25); the presumptive exclusions as presumptions (F30); what the database represents and its coverage biases (F32–F34); Paeth and colleagues' finding of unavoidable epistemic uncertainty (F50) and their CVE inspiration (F51); the ACM snippet (F52). Content replaced by use: the closer (T2-8). Neither transcript names the Partnership on AI as founding sponsor (F47; Wikipedia, not a card at either address, composes the database inside Partnership on AI's article, F69–F73), the citation page per incident (F55), the snapshots (F54) or the founding paper (F65–F68); these are own pages and Wikipedia admitted beside the cards by the operator's ruling and were not surfaced by T.

v0.9 §0.5: the first diff is already ontological, and here it is small. The field's ontology and T's agree on the entity (a database of AI harms run by the Responsible AI Collaborative); what T keeps of its cards is the database's self-description, its operator and its derived layer, and what it drops is the database's own structure of record (the incident as an editorial account with reports as evidence attached) and the guide's evidentiary limits (the fifth question, the external factor, 'or be detected', 'substantially similar'). At the name the composition stays at the grain of its about page (CAP-1-01); at the unit's question it reads the guide and compresses its test (CAP-2-01). The row records this as observed and claims nothing about the composer's internals (§0.5).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the database as an editor-resolved database of harm events whose unit carries its reports and a stable citation page (MMRS-01, MMRS-02, MMRS-06); what its selection leaves out, the archive's sense the field does not state: what an answer surface does with a source on an ordinary day has no unit, and attribution erasure, citation fidelity and provenance stripping enter only when they attach to a harm, so there is no field for attribution (MMRS-13, MMRS-14, MMRS-29), with the open question whether report texts are kept (MMRS-28); the citation page as the meta-aggregator's model (MMRS-16) and the merge rule that keeps a page up (MMRS-10), with the highest id as an upper bound on incidents (MMRS-20); the record kept apart from the entity, each assertion with a basis, the lifecycle apart from the observation (MMRS-30, MMRS-31) and its unchecked items (MMRS-32); the code-against-data licence distinction (MMRS-23, CAP-1-07); the neighbours and negative tags in the harm basin (MMRS-17, MMRS-18, MMRS-24, MMRS-35) and the Kaggle homonym set apart (MMRS-25, CAP-1-06); incident 693, Google's AI Overviews, as the field's index case (MMRS-15, MMRS-36); the battery's checks (MMRS-33); the composition layer's two observations at the entity's addresses (CAP-1-01–05, CAP-2-01–05); and the archive's own address of three of its terms to the database as a glossary venue (#543, R543-01). Coherence (P_coh, composed each with its source): the MMRS record's 'run by the Responsible AI Collaborative since 2020' (MMRS-01) against Wikipedia's shift to an independent non-profit in 2022 (F73); the guide's variant as a distinct incident (F24, MMRS-07) against incident 693's page, grouped under the first reported incident (F62); the MMRS record's quoted presumptive exclusions (MMRS-12) against the guide's words (F30), the substance agreeing. Restated lineages (§3.6, the own page keeping the earlier date): MMRS-03 with F1; MMRS-04, MMRS-37 with F4, F46, F47; MMRS-05, MMRS-19 with F65, F68; MMRS-06 with F19, F22; MMRS-07 with F24; MMRS-08 with F26; MMRS-09 with F3, F45; MMRS-11, MMRS-22, MMRS-27 with F54, F57; MMRS-34 with F6, F7. The measure is in row.measure.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18'], 'T1_composed': ['F1', 'F2', 'F3', 'F4', 'F6', 'F7', 'F8', 'F10', 'F11', 'F12', 'F14'], 'T2_cards_claims': ['F6', 'F7', 'F8', 'F9', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38', 'F39', 'F40', 'F41', 'F42', 'F43', 'F44', 'F45', 'F46', 'F47', 'F48', 'F49', 'F50', 'F51', 'F52', 'F53'], 'T2_composed': ['F19', 'F20', 'F24', 'F26', 'F28', 'F29', 'F38', 'F48'], '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)"}