OECD AI Incidents and Hazards Monitor

register entry oecd-ai-incidents-and-hazards-monitor-aio-20261029 OBS-23f50c1b7e04 · file as first run sha256 9c284bc639b5beb5…

• T1-1 definition: 'a real-time tracking tool developed by the OECD.AI Policy Observatory in collaboration with the Patrick J. McGovern Foundation to monitor actual and potential risks posed by artificial intelligence' (marker: an opaque token, the one shared with the Real-Time Tracking line and the incident definition)
• T1-2 Real-Time Tracking: 'It continuously aggregates and analyzes world news using the Event Registry platform, processing over 150,000 news articles daily to detect AI-related events.'
• T1-3 Evidence Base: 'It provides policymakers, developers, and the public with empirical data on how AI risks materialize globally.'
• T1-4 Thematic Classification: 'Media reports are categorized into 14 distinct risk themes—including synthetic media, cyberattacks and fraud, autonomous vehicles, large language model (LLM) hallucinations, labour market disruptions, and child safety.'
• T1-5 AI Incident: 'An event where the development, use, or malfunction of an AI system results in actual harm (such as injury, health damage, critical infrastructure disruption, or human/fundamental rights violations).'
• T1-6 AI Hazard: 'An event or circumstance where the development or use of an AI system could plausibly lead to an AI incident (a near-harm or precursor event).'
• T1-7 access: 'You can explore live data, interactive filters, and recent tracking reports directly on the OECD AI Incidents Monitor.'
• T1-8 closer: 'Would you like to explore specific risk categories (such as labor market impacts or synthetic media) or look into the common reporting framework guidelines?'

• B1 OECD AI Policy Observatory — OECD AI Incidents Monitor, an evidence base for trustworthy AI card; HTTP 200 (2026-10-29T05:06:49Z); B1.html sha256 e1681d645755e7c4
• B2 OECD — AI risks and incidents card; HTTP 403 (2026-10-29T05:08:47Z), unreachable; card snippet stands; B2.html sha256 54fda3deeb29a2f2
• B3 OECD AI Policy Observatory — Overview and methodology of the AI Incidents and Hazards Monitor - OECD.AI card; HTTP 200 (2026-10-29T05:06:50Z); B3.html sha256 26400e79e67de5ee
• B4 OECD — Trends in AI incidents and hazards reported by the media | OECD card; HTTP 200 (2026-10-29T05:08:47Z); B4.pdf sha256 aab2727da192624f
• B5 OECD — Defining AI incidents and related terms (EN) card; HTTP 200 (2026-10-29T05:08:48Z); B5.pdf sha256 9c46e00f1b3ff6de
• B6 YouTube·International Association for Safe & Ethical AI — Establishing Effective Serious Incident Reporting Frameworks – Jimmy Farrell card; HTTP 200 (2026-10-29T05:09:49Z); B6.json sha256 5a063beb92dbc849
• B7 OECD AI Policy Observatory — AI Search Engines Fabricate News Citations, Undermining Publishers (AIM item 2025-03-11-8537) own page; HTTP 200 (2026-10-29T05:08:49Z); B7.html sha256 9edcbeb65fe222f4
• W1 Wikipedia — no article (the monitor, its former name, OECD.AI, the OECD AI Policy Observatory) wikipedia; raw wikitext HTTP 404 for seven titles; rendered page 404; API: titles missing; Special:Search no results
• W2 Wikipedia — OECD (no mention of the monitor) wikipedia; raw wikitext and rendered page HTTP 200 (2026-10-29T05:10:57Z); W2.wiki sha256 03ca69ee493875ab

E_A, the archive's entity: the same monitor, in the MMRS meta-aggregator's record: an automated monitor of AI harm events in the news, run by OECD.AI, whose records are machine-written from the articles, so that the monitor is itself a machine reception of the news, its summaries compositions and the articles their sources; admitting an event when reputable outlets report it, so that a harm no outlet reports has no event; the automated counterpart of the editor-resolved AI Incident Database in the harm basin; a record kept apart from the entity; observed by the Capture Registry at its name, where the composition layer gave its inputs and definitions and left its classifier out; and a member of the class of media monitors, which the archive sets apart from its own instrument of the retrieval layer (#637) MMRS-01 MMRS-12 MMRS-25 MMRS-10 MMRS-11 MMRS-13 MMRS-31 CAP-1-07 CAP-1-02 O637-01; defining papers #637

E_F, the field's entity (type C — a conventional reading holds it (public entity: a monitor)): an automated media discourse monitor of AI incidents and hazards (beta), initiated and developed by the OECD.AI expert group on AI incidents with the support of the Patrick J. McGovern Foundation, previously the AI Incidents Monitor, whose unit is the event (an incident or hazard in the OECD's terms of May 2024, reported by one or more news articles), drawn from Event Registry's AI-tagged news events from reputable outlets, classified since November 2024 by LLMs (GPT-4o mini, GPT-4o) and recorded in LLM-generated titles, summaries and metadata (o3-mini), daily; its evidence a subset bounded by what the media report (B1–B7)

ρ = extends: the archive reads E_F as the entity and adds senses and limits on the field's own object: the monitor as a machine reception of the news (the consequence of the machine-written record the field states), what its selection leaves out (no event for an unreported harm), the record kept apart from the entity and its unchecked items, the neighbours and the holding, the class-level distinction of media monitors from the archive's instrument, and the composition layer's observed treatment of the entity MMRS-12 MMRS-25 MMRS-11 MMRS-10 MMRS-31 MMRS-26 MMRS-13 MMRS-14 O637-01 CAP-1-03 CAP-1-07. Guard: ρ = extends claims an addition and no more: MMRS-10, MMRS-11 and MMRS-12 are the record's readings (interpretation), not findings that the monitor is defective; O637-01 is a distinction drawn at the class, not a claim on the monitor by name; where an archive claim meets a field claim (MMRS-18 beside F11; MMRS-22 beside F7, F32, F33) 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 monitor from its own pages (the methodology page and the monitor's page), marking each field's source and basis, and composes nothing that denies a field presupposition. What it adds lies on the field's object: the record's machine-written wording read as the monitor being itself a machine reception of the news, its summaries compositions and the articles their sources; what its selection leaves out (a harm no outlet reports has no event); the record kept apart from the entity; the neighbours in the harm basin; the holding; and the Capture Registry's observation that the composition layer left the classifier out. The ontology the archive uses is available in the field itself: the methodology page states that the metadata, title and summary are 'LLM-generated' (F28), the monitor's page marks them '[AI generated]' (F8), the trends paper clusters the monitor's 'LLM-generated summaries' (F41) and bounds its evidence by what journalists know (F36). The archive's reading draws the consequence of the field's own statements; 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 OECD AI Incidents and Hazards Monitor (AIM) is an automated monitor of AI harm events in the news, run by OECD.AI.

#MMRS article, What its selection leaves out s3: The record's own wording is machine-written from the articles, so the monitor is itself a machine reception of the news: its summaries are compositions, and the articles are their sources.

#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 oecd-ai-incidents-and-hazards-monitor-aio-20261029, analysis s1: The monitor is composed from its own methodology page with its inputs and definitions right and its classifier left out, the part of its method that makes each record a machine composition.

#B3 Methodology step 5 (field): This metadata is LLM-generated (OpenAI’s o3-mini) from the top three articles of each event, selected from different news outlets.

#B4 Annex A (field): LLM-generated summaries of AI incidents and hazards were collected and transformed into a semantic representation capturing the underlying concepts.

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 OECD AI Incidents and Hazards Monitor is an "Automated media discourse monitor of AI incidents and hazards (Beta)" that documents AI incidents and hazards to help policymakers, AI practitioners and all stakeholders gain insight into the risks and harms of AI systems and that will, over time, help show risk patterns, by its own page. The OECD.AI expert group on AI incidents, formalised by the OECD in January 2023 to develop a reporting framework and the monitor, initiated it and develops it with the support of the Patrick J. McGovern Foundation, its goal to track incidents and hazards as the OECD defines them, in real time, as an evidence base for policy; it was previously the AI Incidents Monitor, and the OECD's topic page says all actors must monitor incidents and hazards. Its definitions come from the OECD paper of May 2024 (No. 16, by Karine Perset and Luis Aranda, approved on 14 March 2024, from a stocktaking of more than 30 frameworks): an AI incident is an event where the development, use or malfunction of AI systems directly or indirectly leads to harm to health, critical infrastructure, rights, or property, communities or the environment, and an AI hazard is one that could plausibly lead to an incident, near misses included, both resting on the OECD's definition of an AI system. The paper places reputational harm, misinformation, authors' rights under (c), grades the terms into serious incidents, disasters and serious hazards, leaves seriousness to context and lets each jurisdiction set its scope. It detects incidents and hazards from reputable international news outlets through Event Registry, which processes over 150,000 articles a day and clusters them into events tagged by concept, only AI-tagged events reaching the monitor, from the top 50% of outlets by Alexa rank; an incident or hazard can be reported by several articles on one event. Since November 2024 GPT-4o mini filters events and GPT-4o reclassifies them as incidents, hazards or unrelated, articles are classified by content and the event takes their majority label, o3-mini writes each event's title, summary, harm type, severity, stakeholders and country from its top three articles, chosen by outlet ranking, and groups events by summary similarity; before that, traditional machine learning classified them. The pipeline runs daily on events one to four days old; the page marks its summaries and label explanations "[AI generated]", naming classes such as Complementary Information and Beneficial Use, and lists about 18,579 incidents and hazards, while the methodology page gives about 10, and elsewhere 30, a day. Media attention to AI incidents has grown while their share of AI news fell, from 92 to 324 a month and from 3.2% to about 2.5% between 2022 and 2025, and the OECD's trends paper of February 2026, which describes the monitor as tracking media-reported incidents in real time, clusters its LLM-generated summaries into 14 themes, one per item, in three patterns of coverage. Its limits are a likely subset of all incidents, media and public interest rather than systematic tracking, what journalists know and deem newsworthy, possible errors of inclusion and trends that do not measure severity; its information is not the OECD's official view, its third-party articles are not independently verified, and reliability is sought to the best extent possible. Few public, comparable sources exist; the common reporting framework approved by GPAI in early 2025, open submissions, court rulings and a record by theme are the next steps; the trends paper is CC BY 4.0; and the AI Incident Database's founder gave feedback on the definitions. Among its items, the Tow Center's study of AI search engines that fabricate news citations is labelled an AI incident with economic, reputational and public-interest harm, built from articles from 11 March 2025; its card rail also carries a talk on serious incident reporting frameworks (IASEAI 2025).

The OECD AI Incidents and Hazards Monitor (AIM) is an automated monitor of AI harm events in the news, run by OECD.AI, whose census gives its kind as a monitor (MMRS record); its own page calls it an "Automated media discourse monitor of AI incidents and hazards (Beta)", which the MMRS record quotes, and the OECD's trends paper of February 2026 describes it as tracking media-reported AI incidents and hazards "in real-time" and leveraging the OECD common reporting framework to categorise them consistently. Its methodology page records that it was previously the AI Incidents Monitor, the name the MMRS record gives beside AIM, and that its use is subject to the OECD's terms and conditions. The methodology page says it "was initiated and is being developed by the OECD.AI expert group on AI incidents with the support of the Patrick J. McGovern Foundation"; the OECD's definitions paper dates the group to January 2023, when the OECD formalised it to develop a common AI incident reporting framework and an AI Incidents Monitor, and the MMRS record names OECD.AI as operator, on the methodology page's evidence. Its own page says it documents incidents and hazards so that policymakers, AI practitioners and all stakeholders may gain insight into the risks and harms of AI systems, and that over time it will help show risk patterns and a collective understanding of them; the methodology page sets its goal as tracking incidents and hazards "as defined by the OECD, in real time" to provide the evidence base for policy discussion, and the OECD's topic page, read by its card's snippet only, says all actors must monitor incidents and hazards and that risk mitigation will require a solid evidence base. Its terms are the OECD's: the paper Defining AI incidents and related terms (OECD Artificial Intelligence Papers No. 16, May 2024), written by Karine Perset and Luis Aranda, approved by the Digital Policy Committee on 14 March 2024 and built on a stocktaking of more than 30 frameworks, standards and legal instruments, provides them to inform the common reporting framework and "its application through the OECD AI Incidents Monitor (AIM)", and the methodology page and the MMRS record both source the monitor's definitions to it. An AI incident is "an event, circumstance or series of events where the development, use or malfunction of one or more AI systems directly or indirectly leads to" harm to the health of a person or groups, disruption of critical infrastructure, violations of human rights or breaches of law protecting fundamental, labour and intellectual property rights, or harm to property, communities or the environment (the paper; restated on the methodology page), and an AI hazard is such an event that "could plausibly lead to an AI incident", near misses included (the paper; the methodology page); the MMRS record gives the two terms in brief, the hazard in the paper's words. The paper places reputational harm to individuals, hate speech and mis- and disinformation under (c), as a breach of fundamental rights, and violations of authors' and personal rights there too, with industrial property under (d); the definitions rest on the OECD Recommendation's definition of an AI system (2019, revised 2024), as the methodology page states. The paper also grades the terms, serious AI incidents a subset "intended to include substantial, material incidents", an AI disaster a serious incident that may test or exceed a society's capacity to cope, and serious AI hazards a subset of hazards, and leaves seriousness context-dependent and "left for further discussion"; its governing limit is that the definitions aim at interoperability "while providing flexibility for jurisdictions to determine the scope of AI incidents and hazards they wish to address". Its unit is the event: an incident or hazard "can be reported by one or more news articles covering the same event", its own page notes, its data processing powered by Microsoft Azure using data from Event Registry, and the MMRS record's census gives the unit as an incident or hazard, an event in which an AI system leads, or could plausibly lead, to harm. Its intake is the news: the methodology page says the monitor detects incidents and hazards "from reputable international news outlets", with data from Event Registry, a news intelligence platform "processing over 150 000 news articles daily", whose events are clusters of articles on one happening tagged with concepts, only events with AI concepts reaching the monitor; the trends paper adds that the data rely on the top 50% of media outlets by Alexa traffic ranking, and the MMRS record gives the same clustering and tagging and reads admission as an event entering when reputable outlets report it. Since November 2024 its classifiers are large language models (methodology page): from Event Registry's event summary, OpenAI's GPT-4o mini filters out events that seem unrelated and GPT-4o reclassifies the rest as incidents or hazards, each article is then classified from its own content and the event takes the most common label among its articles, and items collected before November 2024, classified then "using traditional machine learning techniques", are being re-annotated with the same method. Similar events are grouped through the ten nearest summaries by cosine similarity, o3-mini deciding whether they cluster, and the pipeline runs daily on events one to four days old, articles later than that being "likely to be misclassified as a separate event" (methodology page). The MMRS record gives the same pipeline: AI-tagged news events read from Event Registry daily and, since November 2024, large language models that sort them into incidents and hazards and write their summaries (GPT-4o mini filtering, GPT-4o reclassifying, o3-mini writing the title, summary, harm type, severity, stakeholders and country from the top three articles), new items matched to existing events by embedding similarity, its census intake "automated" and its cadence daily. Each record's title, summary and metadata are written by a model: the methodology page says the title, summary, harm type, severity, affected stakeholders and country are "LLM-generated (OpenAI’s o3-mini) from the top three articles of each event", the articles chosen by their outlets' ranking in their country and the metadata generated once, when the item is added, and on the monitor's own page its summaries and the explanations of its labels are marked "[AI generated]", some explanations ruling out further classes by name ("It is not Complementary Information … nor is it Beneficial Use or Unrelated"). The MMRS record draws the consequence: the record does not preserve the original outputs, its titles, summaries and metadata LLM-written from the top articles, so "the monitor is itself a machine reception of the news: its summaries are compositions, and the articles are their sources"; the Capture Registry (2026-10-29) sets this reading against the composition layer's, which reads the monitor "as an aggregator of the news". Its page listed "About 18579 incidents & hazards" on 29 October 2026, the count the MMRS record observed on 28 October; the methodology page gives two daily rates, about 1,000 events analysed a day "resulting in around 10 AI incidents or hazards" and, in its table's note, "on average 30 incidents and hazards per day". Its page says incidents "seem to be getting more media attention lately" but have gone down as a share of all AI news, and the trends paper counts 92 a month in 2022 rising to 324 in 2025 while their share of AI news fell from 3.2% to around 2.5%. Its limits are stated by its sources: the methodology page recognises that these publicly reported incidents and hazards likely "only represent a subset of all AI incidents and hazards worldwide", a starting point for an evidence base, and the trends paper that its findings "reflect media and public interest rather than comprehensive or systematic tracking", media capturing "only what journalists are aware of and deemed newsworthy", that reliance on media may admit incidents mischaracterised, exaggerated or inaccurately attributed to AI, and that its trends do not measure severity, harm type or the number of stakeholders affected. Its disclosures say that its information should not be reported as the official views of the OECD or its members, that it is populated with articles from third-party outlets and news aggregators with which the OECD has no affiliation, that the OECD "does not independently verify" their accuracy, completeness or validity, and that collection and analysis seek reliability, objectivity and quality "to the best extent possible". What its selection leaves out, in the MMRS record's reading: admission requires harm, an incident or a hazard that could plausibly lead to one, and "a harm no outlet reports has no event to cluster". The MMRS record keeps itself apart from the entity, each assertion carrying a source and a basis and the lifecycle kept apart from the observation that saw it: the monitor active as read on 28 October 2026, its launch date and its licence unchecked; the trends paper carries its own licence on its rights page, Creative Commons Attribution 4.0. The fourteen themes are the trends paper's (OECD Artificial Intelligence Papers No. 53, February 2026): it classifies the monitor's media reports into 14 themes, among them synthetic media, child safety, cyberattacks, privacy and health, with three patterns of coverage, increasing, intermittent and declining, by collecting the monitor's "LLM-generated summaries", clustering them by k-means into 25 clusters, a number chosen "based on a qualitative keywords review" when the statistical tests gave inconsistent results, consolidating them into 14 and assigning each item a single theme. The Capture Registry (2026-10-29) records that the composition's count of fourteen themes comes from that card, that the methodology page gives no count of themes, and that it did not verify the count. The trends paper places the monitor where there are "few public and comparable sources of information on AI incidents and hazards", media reports among the proxy indicators that give partial insight, beside the common framework for reporting AI incidents that the Global Partnership on AI approved in early 2025; the methodology page names as next steps an open submission process under that framework and court rulings and supervisory decisions, and the trends paper open submissions and a record reorganised by theme. In the MMRS record it is the automated counterpart of the AI Incident Database, "whose editors resolve each submission", and with the AIAAIC Repository one of the census's three harm repositories, its neighbours in the harm basin, "automated and news-derived" in the MMRS record of the AI Incident Database; the definitions paper thanks that database's founder, Sean McGregor (Responsible AI Collaborative), among the experts who gave feedback. A research study can enter it as an event (MMRS record): the Tow Center's study of AI search citation is filed as an incident, "a research study entering a harm monitor as an event", and on its page the item, titled "AI Search Engines Fabricate News Citations, Undermining Publishers", carries a machine-written summary in which eight search engines and chatbots "frequently misquote or fabricate news citations, remixing publisher content without credit", is labelled an AI incident with harm types economic/property, reputational and public interest, and is built from news articles about the study, the earliest dated 11 March 2025. The archive sets one of its own instruments apart from the class: "Media monitors track coverage of events", while the Encyclotron (#637) tracks the representation of entire fields of scholarship, "not what was said yesterday but what the retrieval layer thinks the field has been saying for decades". Its card rail at its name also carries a conference talk admitted by its title, "Establishing Effective Serious Incident Reporting Frameworks – Jimmy Farrell | IASEAI 2025". At its name the Overview gave the monitor's input as its methodology page does, Event Registry and 150,000 articles a day, and a maker close to the page's, and did not say that large language models have classified its events since November 2024 and written their titles and summaries: it composed the monitor "with its inputs and definitions right and its classifier left out, the part of its method that makes each record a machine composition" (Capture Registry, 2026-10-29, checked against the MMRS record and the entity's own pages), the record holding "what the composition drops", and the MMRS record lists the capture as the composition layer's account of the entity.

T against L_F(B) (AIO's ontology at work on its own cards): At 'oecd ai incidents and hazards monitor' (T1, six cards: B1–B6) the composition gives the monitor by its own pages and the OECD's papers. Composed from its cards: the maker and supporter (B3 F11, at a near grain: 'developed by the OECD.AI Policy Observatory in collaboration with' for 'initiated and is being developed by the OECD.AI expert group on AI incidents with the support of', CAP-1-05); real-time tracking and the evidence base (B3 F12; B4 F38; B1 F2); Event Registry and the 150,000 articles a day (B3 F13), with the AI-tagged events (F14, 'to detect AI-related events'); the definitions of incident and hazard (B5 F51–F53; B3 F19–F20), near misses as 'a near-harm' (F53). Assigned to another: the processing of 'over 150,000 news articles daily' is Event Registry's (F13: 'a news intelligence platform that monitors and aggregates world news by processing over 150 000 news articles daily'); T1-2 composes it as the monitor's ('It continuously aggregates and analyzes world news … processing over 150,000 news articles daily'). Substituted reading: the fourteen themes (T1-4, 'Thematic Classification: Media reports are categorized into 14 distinct risk themes') are the trends paper's analysis of the monitor's data (B4 F35), made by k-means clustering of the monitor's LLM-generated summaries into 25 clusters consolidated to 14 (F41, F42), each item assigned one theme (F43); T composes them as a classification the monitor performs; the methodology page gives no count (CAP-1-06, now read against B4); the paper's next steps include reorganising the record by theme (F46), not yet done. The theme names T lists beyond F35's span (cyberattacks and fraud, autonomous vehicles, LLM hallucinations, labour market disruptions) are in B4's executive summary outside the ledgered spans. Limit dropped: T1-2's 'Real-Time Tracking … continuously' against the daily pipeline on events one to four days old (F31); T1-5 drops harm (d), 'harm to property, communities or the environment', and 'directly or indirectly' and 'series of events' from the incident (F51); T1-6 drops 'malfunction' from the hazard (F52). Force raised: T1-3, 'empirical data on how AI risks materialize globally', against the field's own limits: 'only represent a subset of all AI incidents and hazards worldwide … a useful starting point' (F15), 'reflect media and public interest rather than comprehensive or systematic tracking' (F36), errors of inclusion (F37), trends that do not measure severity (F44), and information the OECD 'does not independently verify' (F24). Composed without attribution in the body, at the source's grain: the definitions (T1-5, T1-6) carry markers but not the paper (F48, F18), which is card 5. Available, not composed: the self-description 'Automated media discourse monitor of AI incidents and hazards (Beta)' (F1); the LLM classification since November 2024 (F25–F27), the o3-mini metadata from the top three articles (F28–F29), grouping by summary similarity (F30) and the earlier 'traditional machine learning techniques' (F34), which the Capture Registry names as the classifier left out (CAP-1-02, CAP-1-07); the '[AI generated]' marks and the further classes on the monitor's own page (F8, F9); the unit as an event reported by one or more articles, on Microsoft Azure (F6); the count (F7) and the two daily rates (F32, F33); the trend, more attention and a falling share (F4, F40); the top 50% of outlets by Alexa rank (F39); the disclaimers (F5, F17, F23, F24); the graded terms (F54, F55, F58), the place of reputational harm, misinformation and authors' rights under (c) (F56), the jurisdictional limit (F57), the AI system definition (F21), the paper's authors, approval and stocktaking (F49, F59) and the expert group's formalisation in January 2023 (F50); the former name (F22); next steps (F16, F46); the common reporting framework and the scarcity of public sources (F45); the AI Incident Database's founder among the definitions' reviewers (F60); the OECD topic page's snippet (F10, card 2, unreachable); the video (F61, title only). Content replaced by use: the access line (T1-7) and the closer (T1-8), which offers the common reporting framework guidelines (F45, F16) as a next step. Not surfaced by T: the holding B7 (the Tow Center item, F62–F64), an own page the MMRS record cites, admitted beside the cards by the operator's ruling. Wikipedia has no article (W1) and the OECD article does not mention the monitor (W2).

v0.9 §0.5: the first diff is already ontological. The field's ontology holds the monitor as an automated media monitor whose classifier and record are machine-made (F1, F26, F28, F8) and whose evidence is bounded by what the media report (F15, F36); T's holds it as a tracking tool that aggregates and analyses the news and categorises it into themes. What T keeps of its cards is the monitor's inputs and definitions; what it drops is the machine in the method and the limits on the evidence, and it composes a paper's analysis of the monitor's machine-written summaries as the monitor's own classification. 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 monitor as an automated monitor of AI harm events in the news, run by OECD.AI (MMRS-01, MMRS-19), and as itself a machine reception of the news, its summaries compositions and the articles their sources (MMRS-12), the record not preserving the original outputs (MMRS-25) — the field states the machine-written record (F28, F8) and the archive reads its consequence; the composition layer's reading against it, 'as an aggregator of the news' (CAP-1-03); what its selection leaves out, 'A harm no outlet reports has no event to cluster' (MMRS-11), with admission requiring harm (MMRS-24) and entering by reputable outlets (MMRS-10), which extends the field's own limits (F15, F36) to the unit; the record kept apart from the entity, with its unchecked launch date and licence (MMRS-31, MMRS-26, MMRS-27); the neighbours, the automated counterpart of the editor-resolved AI Incident Database and one of the census's three harm repositories (MMRS-13, MMRS-15, MMRS-16, MMRS-29); the holding, a research study entering a harm monitor as an event (MMRS-14, MMRS-28), read against its own page (B7, F62–F64); the archive's class-level distinction of media monitors, which 'track coverage of events', from its own instrument of the retrieval layer (#637, O637-01); and the composition layer's observation at the entity's address (CAP-1-01–08, MMRS-30). Coherence (P_coh, composed each with its source): the MMRS record's operator OECD.AI (MMRS-18) beside the methodology page's expert group with McGovern support (F11), not in contradiction; the count about 18,579 observed 2026-10-28 (MMRS-22) and the same count on the page 2026-10-29 (F7), beside the page's two daily rates (F32, F33; internally_inconsistent in B3). Restated lineages (§3.6, the own page keeping the earlier date): MMRS-03, MMRS-19 with F1; MMRS-04, MMRS-05 with F51, F52, F18; MMRS-06 with F14; MMRS-02, MMRS-07, MMRS-21 with F25, F26, F28; MMRS-08 with F30; MMRS-09, MMRS-23 with F31; MMRS-17 with F22; MMRS-20 with F6; MMRS-22 with F7; MMRS-25 with F28; MMRS-14, MMRS-28 with F62. 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', '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', 'F54', 'F55', 'F56', 'F57', 'F58', 'F59', 'F60', 'F61'], 'T1_composed': ['F2', 'F11', 'F12', 'F13', 'F14', 'F19', 'F20', 'F35', 'F38', 'F51', 'F52', 'F53'], '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). T1_cards_claims are the claims of B1–B6 (B7 is an own page T did not surface)"}