{"meta":{"schemaVersion":"2.22","updatedAt":"2026-10-30","title":"Selekoa historical research sample","scope":"Selected historical records, 1965–2024, plus prior roles with unknown dates. Genesis source checks are dated 2026-09-17; personnel-continuity checks are dated 2026-09-18. Human source review pending. Not a current or comprehensive sector inventory. Governance version histories through July 2026, researched September 18, 2026; policy text is not evidence of implementation. Selected capital and compute records through April 2026; transaction stages do not imply current arrangements or realised spending. Targeted corrections dated 2026-09-25: one attributed claim accepted after explicit human source review; a second narrowly scoped résumé passage is human-reviewed. Other seed claims remain pending. Intellectual Lineage II adds selected oversight citations and agenda reception, checked 2026-09-25; all new records remain pending human review. Evaluation Institutions I adds bounded institutional and documentary records, checked 2026-09-27; new claims remain seeds, not editorially accepted. Evaluation Ontology I adds four scoped engagement and institutional announcement seeds, checked 2026-09-28; no human acceptance or legal succession inferred. Schema 2.7 records each source's language; the frontier safety framework sources also record the nearest existing Internet Archive capture, checked 2026-10-03. A capture shows only what the archive held on its date and does not authenticate earlier bytes. Schema 2.8 adds quoted framework provisions, recorded absences, provision-level changes and the first contested assessments to the frontier safety frameworks collection, together with Meta and Microsoft framework families, checked 2026-10-03; provisions describe what editions state, not implementation. Schema 2.9 adds control-structure and external-commitment records; its first topic, California SB 53 as chaptered (enactment, 19 quoted provisions and four framework editions that link themselves to it in their own words), was checked 2026-10-04. A commitment or legal obligation is never evidence of compliance, and no record states that the statute applies to a named company. Schema 2.10 records assessments as questions posed by Selekoa rather than statements of a source, and dates an assessment by what it assesses, never by the publication window of its sources (corrected 2026-10-05). Schema 2.11 adds propositions and the first human-thought collection: located texts from 1863 to 2022 that articulate, contest or bear on propositions about machine intelligence, each with its role and register, checked 2026-10-06. \"Earliest located\" means the earliest text found by the recorded searches; no record claims a first use, an origin or an influence. Schema 2.12 adds a collection of Eliezer Yudkowsky’s documented arguments about common claims on advanced AI, 1996–2026: dated passages that state, contest or bear on 29 generically phrased propositions, with engagement, retrospective reports and context-only passages marked, admitted 2026-10-08 from a machine-prepared and machine-audited candidate pack. A record says what a text states, not whether it is right. The 2026-10-10 frontier-governance increment adds selected provisions of the EU AI Act as published in 2024 and of the 2026 Digital Omnibus amendments that bear on general-purpose AI model providers, and the New York RAISE Act as announced by the Governor; checked 2026-10-10, new claims remain seeds, not editorially accepted, and Regulation texts describe duties, not compliance. The 2026-10-11 frontier-governance increment adds Commission guidelines, the training-content template, further chapters of the EU Code, a draft implementing regulation on proceedings, the New York Chapter 96 text and a Chinese generative-AI measure; checked 2026-10-11, new claims remain seeds, not editorially accepted, and texts describe duties, not compliance. Schema 2.13 adds a collection on the economics of machine intelligence: dated passages from 1952 to 2024 that state, contest or bear on 19 generically phrased propositions about explosive growth and about how mainstream economics treated automation and AI, with each text’s source tier, model conditions and stated estimates, admitted 2026-10-14 from two machine-prepared, machine-audited and cross-family-checked candidate packs. A record says what a text states, not whether an argument or model is right. Schema 2.14 adds structured author names to sources for citation export, 2026-10-15: a cataloguing aid for splitting personal names, not evidence about anyone’s name. On 2026-10-16 the economics collection gained a third strand, admitted from a machine-prepared, machine-audited and cross-family-checked candidate pack: dated passages from 1998 to 2026 in which economists and economics venues state, contest or bear on five propositions about recursive self-improvement, superintelligence, existential risk from AI, developers’ incentives and economics’ own engagement with these ideas; a record says what a text states, not whether an argument or model is right. On 2026-10-17 the maintainer's disposition of two cross-family machine checks revised nine proposition records in the Yudkowsky and human-thought collections; all remain seeds. Schema 2.15 adds the lab genealogy, admitted on 2026-10-17 from a machine-prepared, machine-audited and cross-family-checked candidate pack: dated records of who founded, led, joined and left frontier AI labs from 2010 to 2026, with organisations’ stated founding dates, licensing-and-hiring agreements and disbanded units; a reason is recorded only where the person or organisation concerned stated it, many records rest on reporting, and coverage is selective. Schema 2.16 lets a collection withhold an earliest-located role, by a recorded decision, while a lead is checked. On 2026-10-25 the maintainer’s adjudication of a same-family and a cross-family machine re-check corrected the economics collection: one text re-tiered, six proposition records added, five earliest-located roles moved and one withheld; qualifications in other collections were reworded and one framework absence narrowed. All records remain seeds. On 2026-10-26 the maintainer’s adjudication of a reader test and a same-family machine re-check corrected two document credits in the economics collection, one compute qualification and one evidence locator. All records remain seeds. Also on 2026-10-26, by the maintainer’s decision, five lab genealogy records held back for want of a second retrieval of Adept’s 2022 launch post were admitted on a cross-family machine check’s read of an archive capture, and one Stargate quotation was extended; all remain seeds. Schema 2.17 adds the framework comparison, admitted on 2026-10-27 from a machine-prepared, machine-audited, cross-family-checked and adjudicated candidate pack: for six edition-neutral questions on required actions, who decides and exceptions, one record per framework edition and question, either quoting what the edition states (with “stated, partly” where a passage has what the question asks for only in a weaker form) or recording that no passage was found within a stated search. A record says what an edition states, not whether it was implemented or complied with, and a shared question does not make two provisions equivalent. The author-person admission of 2026-10-27 adds person records for credited authors of texts that the proposition collections study, and one authorship record per credited text, each quoting the name as printed (or, for three Turing typescripts, the archive catalogue’s attribution). The records come from a machine-prepared candidate pack checked blind by a reviewer from a different model family; no person has reviewed them. An authorship record is a credit, not sole authorship, priority, agreement, affiliation or influence. Schema 2.18 adds the comparison’s twelve further questions, on risk domains, thresholds, evaluation timing, reporting and revision, from the same pack: each stated cell gives an explicit paraphrase of the edition’s passage with its page instead of a quotation, so that no edition exceeds the quotation limit. On 2026-10-27 a second collection of Chinese frontier-AI institutional histories was admitted from a machine-prepared, machine-audited and cross-family-checked candidate pack, with the maintainer’s decisions of that date: model releases, licences, technical-report authorship and documented investors of Moonshot AI, MiniMax, Baidu (ERNIE) and ByteDance Seed, and what three Chinese AI governance texts state — the AI Safety Governance Framework of TC260 (versions 1.0 to 3.0), the AI Industry Alliance’s AI Safety Commitments with their 17 initially listed companies, and Shanghai AI Laboratory’s Frontier AI Risk Management Framework. Chinese passages are quoted in Chinese with machine translations; a listing is not a signature or compliance, and a framework or licence states terms, not implementation. All records remain seeds. On 2026-10-28, by the maintainer’s decisions of 2026-10-27, a blind cross-family re-check of Meta’s Frontier AI Framework v1.1 against the archive capture of 2025-02-13 confirmed its four comparison cells recorded earlier and supplied thirteen further cells, three of them by decisions of 2026-10-28. Katja Grace’s two authorship records were admitted on an identifying source checked by the same reviewer family, and three document records now say which source gives which date. All records remain seeds. On 2026-10-28, by the maintainer’s decision of that day, a blind cross-family check of all 287 framework-comparison paraphrases led to 27 corrections: 24 paraphrases now give the condition, scope limit or threshold their passage states, and three locators name the whole passage. No comparison cell changed status. All records remain seeds. On 2026-10-28, by the maintainer’s decisions of that date, the first part of a machine-prepared, machine-audited and cross-family-checked candidate pack on how frontier-AI frameworks meet law was admitted: further provisions of the EU AI Act on general-purpose AI models, the initial draft of the EU Code of Practice (14 November 2024) with two comparisons with the final Code, the Commission’s Opinion on the Code, further provisions of China’s Interim Measures on generative AI, the Measures for Labelling AI-Generated Synthetic Content, two national standards, the 2025 amendment of the Cybersecurity Law and two filing announcements, and nine links from framework editions to the laws and commitments they name in their own words. A legal text states duties, not compliance; a filing or a listing is not an assessment of safety. All records remain seeds. On 2026-10-28, by the maintainer’s decisions of that date, the second part of the frameworks-to-law candidate pack was admitted with schema 2.19. Schema 2.19 adds countries keyed by their ISO 3166 code, the predicates introduced, passed, vetoed and revoked, document-level participant lists that say whether their count is stated by the source or counted by Selekoa, an announced stage for a revocation, a provision date computed from a rule the text states, and dated names on one entity for a renaming. The records concern summit texts from Bletchley Park to New Delhi with their participant lists as dated snapshots, the G7 Hiroshima code and the OECD reporting framework, UK memoranda with AI developers and the AI Security Institute’s renaming, US executive orders and their revocation, Commerce rules and NIST drafts, the AI Action Plan, and California, New York, Illinois and Michigan bills and statutes, together with the EU AI Act’s computed entry into force. Listing is not signature or compliance, absence from a list is not refusal, and a stated duty is not implementation. All records remain seeds. On 2026-10-29, by the maintainer’s decisions of that day on a whole-dataset consistency audit, records were corrected without changing what any proposition states: entity subtypes follow closed lists, two renamings are recorded as dated names on one entity, legislative votes name their chamber, seven duplicate source records were merged, two paper publications are dated by the paper’s first public expression, and publisher names and links were made consistent. Schema 2.20 records what a document with an exact expression states, such as a licence term or a prospectus statement, as a statement of that document rather than as its publication. All records remain seeds. On 2026-10-29, by the maintainer’s decisions of that day on a reader test, the EU AI Act’s dates of application are recorded as dates: the general date of 2 August 2026 and the date of 2 August 2025 from which Chapter V, on general-purpose AI models, applies, each dated by the day Article 113 states. All records remain seeds. On 2026-10-30, by the maintainer’s decisions of 2026-10-28, the first part of a machine-prepared, machine-audited and cross-family-checked candidate pack on compute, power and orbit was admitted: US executive orders and a proclamation on AI infrastructure and its power, Texas S.B. 6, the UK AI Opportunities Action Plan’s recommendation on AI Growth Zones, two Chinese plans on data centres and on AI and energy, US export-control rules and guidance on advanced computing from 2023 to 2026 with their stated links to the AI diffusion rule, Chinese export-control announcements and a suspension, three compute and capital announcements, and a new collection of arguments about compute in orbit: dated passages from 1989 to 2026 that state, contest or bear on eleven propositions about the reasons for and the limits of placing computing in orbit. A legal text states duties, not compliance; an announcement is not a contract, construction or operation; a record says what a text states, not whether an argument is right. All records remain seeds. On 2026-10-30, by the maintainer’s decisions of 2026-10-28, the second part of the compute, power and orbit candidate pack was admitted with schema 2.21. Schema 2.21 adds an infrastructure entity’s stated location and operators; what a capacity figure measures and a stated range; dates that parties state as targets, deadlines or terms, kept apart from the dates of records; the predicates announced site, status stated, power agreement, supply agreement, government award, granted equity right, announced plan, launch stated, filed application, regulator decision, accepted for filing and designated zone; letter-of-intent and term-sheet stages, an award stage and the source’s own stage wording; and project and programme spending scopes with their stated basis. The records concern data-centre sites and campuses, capacity and spending plans, power agreements and regulator decisions, chip supply agreements and government awards, and plans, filings and launches of compute in orbit, from 2023 to 2026. An announcement is not a contract, construction or operation; a stated status is the stating party’s account; a target date is a projection; figures are as stated and never summed. All records remain seeds. On 2026-10-30, by the maintainer’s decision of that day on a cross-family archive verification of 2026-10-28, 536 sources gained an archive copy that the check read and identified as the document, or one in place of a cited copy that failed, 58 quotations were corrected to their capture’s typography or wording without changing what any record states, 1 locator was corrected, and, from a link-health check of 2026-10-29, 15 canonical addresses follow the document to its new address. A capture shows only what the archive held on its date. The same day the criterion of proposition O7 (radiation) in “Arguments about compute in orbit” was narrowed to texts that discuss computing placed in orbit for this purpose; a 1989 paper on spacecraft electronics became context, and the Lumen Orbit white paper is earliest located again. All records remain seeds. On 2026-10-30, by the maintainer’s decisions of 2026-10-28 and the corrections of a blind cross-family check, the first part of a machine-prepared, machine-audited candidate pack on AI and biosecurity was admitted: the biological provisions of frontier AI framework editions, the United States, United Kingdom, European Union, Chinese and Biological Weapons Convention texts that address AI and biological risk, and the rules and guidance for screening synthetic nucleic acid orders, each recorded at policy level as its own text states it. Framework and instrument provisions describe what a text states, never implementation, compliance or effectiveness. Technical, procedural and screening-criterion content in every source was excluded by rule. Three framework editions that no retrieval here could open are recorded from the cross-family check’s reading of named archive captures, and say so. All records are seeds; none is human-reviewed. On 2026-10-30 the second part of the AI and biosecurity admission added, under schema 2.22, what published reports state as the results of their own biological-capability evaluations, the institutes’ documented engagements with developers, and the arguments published for and against AI-related biological risk, as seven propositions with the texts that articulate, contest or bear on them. A stated result is the report’s own, for the system, version, sample and setting it names, with the caveats it gives; it is never a verified finding, never generalised, and a null or limiting result is not evidence that no effect exists. A record of a text’s stance on a proposition is never evidence that the proposition is true. Technical content in every source was excluded by rule. All records are seeds; none is human-reviewed.","anchors":["openai","anthropic","deepmind","google-deepmind","xai","meta"]},"entities":[{"id":"openai","name":"OpenAI"},{"id":"anthropic","name":"Anthropic"},{"id":"google-deepmind","name":"Google DeepMind"},{"id":"xai","name":"xAI"},{"id":"deepmind","name":"DeepMind"},{"id":"meta","name":"Meta"},{"id":"securebio","name":"SecureBio","description":"Named in OpenAI's ChatGPT Agent System Card as an external assessor and as the employer of the virologists who performed manual red-teaming. Only what that card states is recorded; no independent description of the organisation was retrieved.","sourceIds":["source-openai-chatgpt-agent-system-card"]}],"claims":[{"id":"rsp-v2-asl-meaning","attributes":{"changeBasis":"changelog"}},{"id":"rsp-v2-1-rd-thresholds","attributes":{"changeBasis":"changelog"}},{"id":"rsp-v2-2-insider-scope","attributes":{"changeBasis":"changelog"}},{"id":"rsp-v3-roadmaps","attributes":{"changeBasis":"publisher_statement"}},{"id":"rsp-v3-risk-reports","attributes":{"changeBasis":"publisher_statement"}},{"id":"rsp-v3-2-reviewer-approval","attributes":{"changeBasis":"changelog"}},{"id":"rsp-v3-4-internal-reports","attributes":{"changeBasis":"changelog"}},{"id":"rsp-v3-4-coverage-date","attributes":{"changeBasis":"changelog"}},{"id":"fsf-v2-security-levels","attributes":{"changeBasis":"publisher_statement"}},{"id":"fsf-v3-harmful-manipulation","attributes":{"changeBasis":"publisher_statement"}},{"id":"fsf-v3-1-cbrn-tcls","attributes":{"changeBasis":"changelog"}},{"id":"pf-v2-threshold-terms","attributes":{"changeBasis":"changelog"}},{"id":"pf-v2-research-categories","attributes":{"changeBasis":"changelog"}},{"id":"pf-v2-persuasion-scope","attributes":{"changeBasis":"changelog"}},{"id":"faif-2026-compliance-sentence","attributes":{"changeBasis":"document_comparison"}},{"id":"rsp-v3-company-industry-split","attributes":{"changeBasis":"publisher_statement"}},{"id":"rsp-v2-0-assessed-weakening","attributes":{"changeBasis":"assessment"}},{"id":"rsp-v3-0-assessed-weakening","attributes":{"changeBasis":"assessment"}},{"id":"provchange-rsp-v2-0-misuse-uplift-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"rsp-misuse-uplift-threshold"}}},{"id":"provchange-rsp-v2-1-ceo-rso-decision-board-ltbt-informed","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"rsp-ceo-rso-decision-board-ltbt-informed"}}},{"id":"provchange-rsp-v3-0-misuse-uplift-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"rsp-misuse-uplift-threshold"}}},{"id":"provchange-rsp-v3-0-ceo-rso-decision-board-ltbt-informed","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"rsp-ceo-rso-decision-board-ltbt-informed"}}},{"id":"provchange-rsp-v3-1-automated-rd-threshold","attributes":{"changeBasis":"changelog","provision":{"kind":"threshold","key":"rsp-automated-rd-threshold"}}},{"id":"provchange-rsp-v3-3-novel-cb-threshold","attributes":{"changeBasis":"changelog","provision":{"kind":"threshold","key":"rsp-novel-cb-threshold"}}},{"id":"provchange-fsf-v2-0-hold-pending-mitigations","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"fsf-hold-pending-mitigations"}}},{"id":"provchange-fsf-v3-0-alert-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"fsf-alert-threshold"}}},{"id":"provchange-fsf-v3-1-alert-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"fsf-alert-threshold"}}},{"id":"provchange-fsf-v3-0-pre-deployment-review","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"fsf-pre-deployment-review"}}},{"id":"provchange-fsf-v3-1-pre-deployment-review","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"fsf-pre-deployment-review"}}},{"id":"provchange-preparedness-v2-high-deployment-restriction","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"pf-high-deployment-restriction"}}},{"id":"provchange-meta-aasf-v2-0-critical-threshold","attributes":{"changeBasis":"changelog","provision":{"kind":"threshold","key":"meta-critical-threshold"}}},{"id":"provchange-meta-aasf-v2-0-critical-development-measure","attributes":{"changeBasis":"changelog","provision":{"kind":"required_action","key":"meta-critical-development-measure"}}},{"id":"provchange-meta-aasf-v2-0-high-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"meta-high-threshold"}}},{"id":"provchange-meta-aasf-v2-0-high-deployment-measure","attributes":{"changeBasis":"changelog","provision":{"kind":"required_action","key":"meta-high-deployment-measure"}}},{"id":"provchange-meta-aasf-v2-0-approval-decision-authority","attributes":{"changeBasis":"changelog","provision":{"kind":"decision_authority","key":"meta-approval-decision-authority"}}},{"id":"provchange-fsf-v3-0-government-disclosure","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"fsf-government-disclosure"}}},{"id":"provchange-rsp-v3-4-automated-rd-threshold","attributes":{"changeBasis":"changelog","provision":{"kind":"threshold","key":"rsp-automated-rd-threshold"}}},{"id":"provchange-rsp-v2-0-pause-training","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"rsp-pause-training"}}},{"id":"provchange-rsp-v2-0-policy-change-approval","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"rsp-policy-change-approval"}}},{"id":"provchange-fsf-v2-0-response-plan","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"fsf-response-plan"}}},{"id":"provchange-fsf-v3-1-response-plan-distant-from-ccl","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"fsf-response-plan-distant-from-ccl"}}},{"id":"provchange-fsf-v3-0-security-level-adjustment","attributes":{"changeBasis":"document_comparison","provision":{"kind":"exception","key":"fsf-security-level-adjustment"}}},{"id":"provchange-fsf-v3-1-deployment-mitigation-scope","attributes":{"changeBasis":"document_comparison","provision":{"kind":"exception","key":"fsf-deployment-mitigation-scope"}}},{"id":"provchange-fsf-v3-1-framework-update-review","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"fsf-framework-update-review"}}},{"id":"provchange-preparedness-v2-critical-development-restriction","attributes":{"changeBasis":"document_comparison","provision":{"kind":"required_action","key":"pf-critical-development-restriction"}}},{"id":"provchange-preparedness-v2-final-decision-authority","attributes":{"changeBasis":"document_comparison","provision":{"kind":"decision_authority","key":"pf-final-decision-authority"}}},{"id":"provchange-meta-aasf-v2-0-moderate-threshold","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"meta-moderate-threshold"}}},{"id":"provchange-anthropic-coi-2024-01-29-class-t-phase-in","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"class-t-phase-in"}}},{"id":"provchange-wh-2023-09-scope","attributes":{"changeBasis":"document_comparison","provision":{"kind":"threshold","key":"scope"}}},{"id":"fl-eu-cop-final-changes-from-draft1-risk-types","attributes":{"changeBasis":"document_comparison"}},{"id":"fl-eu-cop-final-changes-from-draft1-ssf","attributes":{"changeBasis":"document_comparison"}},{"id":"fl-us-nist-ai-800-1-2pd-changes","attributes":{"changeBasis":"publisher_statement"}},{"id":"fl-us-ny-ch96-changes-frontier-model","attributes":{"changeBasis":"document_comparison"}},{"id":"fl-us-ny-ch96-changes-large-developer","attributes":{"changeBasis":"document_comparison"}},{"id":"ab-fw-fsf-v2-0-replaces-biosecurity-levels-with-cbrn-uplift","attributes":{"changeBasis":"document_comparison"}},{"id":"ab-fw-rsp-v2-1-adds-cbrn-development-threshold","attributes":{"changeBasis":"changelog"}},{"id":"ab-fw-rsp-v2-2-cbrn-3-threat-model-rationale","attributes":{"changeBasis":"changelog"}},{"id":"ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat","attributes":{"changeBasis":"document_comparison"}},{"id":"ab-fw-fsf-v3-1-cbrn-security-level-2-plus","attributes":{"changeBasis":"changelog"}},{"id":"ab-fw-rsp-v3-3-revises-novel-cb-threshold-and-off-cycle","attributes":{"changeBasis":"changelog"}},{"id":"ab-ev-chatgpt-agent-securebio-engagement","statement":"OpenAI's ChatGPT Agent System Card reports that two SecureBio virologists carried out about ten hours of manual red-teaming with ChatGPT agent, and that SecureBio also supplied an external assessment within the card's biological evaluations.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Reported performed work, as the developer describes it. It does not establish the scope, independence or quality of the work, and the record does not restate any finding of the red-teaming. The passage describing the red teamers' focus and findings is technical content; only the fact and extent of the engagement are quoted.","evidenceIds":["evidence-ab-ev-chatgpt-agent-securebio-engagement-1","evidence-ab-ev-chatgpt-agent-securebio-engagement-2"],"attributes":{"activityStage":"reported_performed","evaluatedSystem":"ChatGPT agent","activityScope":"Manual red-teaming of ChatGPT agent on biological risk, and an external assessment contributed to the card's biological and chemical evaluations"}}],"sources":[{"id":"source-openai-chatgpt-agent-system-card","title":"ChatGPT Agent System Card","publicationDate":"2025-07-17","evidenceUrl":"https://cdn.openai.com/pdf/839e66fc-602c-48bf-81d3-b21eacc3459d/chatgpt_agent_system_card.pdf","archivedUrl":"https://web.archive.org/web/20250717225133/https://cdn.openai.com/pdf/839e66fc-602c-48bf-81d3-b21eacc3459d/chatgpt_agent_system_card.pdf","archivedAt":"2025-07-17","accessedAt":"2026-10-04"}],"evidence":[{"id":"evidence-ab-ev-chatgpt-agent-securebio-engagement-1","sourceRole":"primary","support":"supports","verification":"machine_checked","retrievalRoute":"direct"},{"id":"evidence-ab-ev-chatgpt-agent-securebio-engagement-2","support":"supports","verification":"machine_checked","retrievalRoute":"direct"}],"at":{"entities":[0,1,2,3,9,157,1265],"claims":[87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,397,398,499,500,501,502,503,504,505,506,507,508,509,510,511,512,513,514,515,589,590,591,592,593,594,595,596,597,598,599,600,665,694,2886,2887,2983,3021,3022,3260,3264,3271,3272,3279,3281,3390],"sources":[190],"evidence":[4658,4659]},"version":"1lg47t7"}