{"entities":[{"id":"sam-altman","name":"Sam Altman"},{"id":"deceptive-alignment","name":"Deceptive alignment"},{"id":"google","name":"Google"},{"id":"rsp-v2-2","name":"Anthropic Responsible Scaling Policy — v2.2"},{"id":"rsp-v3-1","name":"Anthropic Responsible Scaling Policy — v3.1"},{"id":"rsp-v3-2","name":"Anthropic Responsible Scaling Policy — v3.2"},{"id":"fsf-v2-0","name":"Google DeepMind Frontier Safety Framework — v2.0"},{"id":"fsf-v3-0","name":"Google DeepMind Frontier Safety Framework — v3.0"},{"id":"faif-2026","name":"xAI Frontier Artificial Intelligence Framework — June 2026 edition"},{"id":"meta","name":"Meta"},{"id":"research-supercluster","name":"AI Research SuperCluster (RSC)"},{"id":"deepseek-r1-report","name":"DeepSeek-R1 research report"},{"id":"qwen25-announcement","name":"Qwen2.5: A Party of Foundation Models!"},{"id":"openai-inc","name":"OpenAI, Inc. (OpenAI Foundation)"},{"id":"dsit","name":"Department for Science, Innovation and Technology"},{"id":"hubinger-2022-how-likely-deceptive-alignment","name":"How likely is deceptive alignment?"},{"id":"switch-off-if-dangerous","name":"Switching it off if it becomes dangerous"},{"id":"time-2023-shut-it-all-down","name":"Pausing AI Developments Isn't Enough. We Need to Shut it All Down"},{"id":"cn-generative-ai-interim-measures-2023","name":"Interim Measures for the Management of Generative Artificial Intelligence Services (生成式人工智能服务管理暂行办法), Order No. 15"},{"id":"no-current-sign-of-explosive-growth","name":"Current data show no approach to explosive growth"},{"id":"automation-lowers-labour-income","name":"Automation lowers the labour share, wages or employment"},{"id":"compensating-policy-for-ai","name":"Compensating policy for AI’s distributional effects"},{"id":"nordhaus-2015-economic-singularity","name":"Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth"},{"id":"aghion-jones-jones-2017-ai-and-economic-growth","name":"Artificial Intelligence and Economic Growth"},{"id":"cowen-2019-neglected-open-questions","name":"Neglected Open Questions in the Economics of Artificial Intelligence"},{"id":"h-company","name":"H Company"},{"id":"kai-fu-lee","name":"Kai-Fu Lee"},{"id":"karl-tuyls","name":"Karl Tuyls"},{"id":"yang-zhilin","name":"Yang Zhilin"},{"id":"aiia","name":"Artificial Intelligence Industry Alliance (AIIA) / 中国人工智能产业发展联盟"},{"id":"aiia-ai-safety-commitments-2024","name":"Artificial Intelligence Safety Commitments (人工智能安全承诺), December 2024"},{"id":"kimi-k2-report","name":"Kimi K2: Open Agentic Intelligence"},{"id":"cn-ai-content-labelling-measures-2025","name":"Measures for Labelling Artificial Intelligence-Generated Synthetic Content (人工智能生成合成内容标识办法)"},{"id":"seoul-declaration-2024","name":"Seoul Declaration for safe, innovative and inclusive AI (AI Seoul Summit, 21 May 2024)"}],"claims":[{"id":"xm-rsc-initial","statement":"Meta reported RSC operating with 6,080 GPUs.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Company-reported first-phase capacity, not the later expansion target or independent utilization measurement.","evidenceIds":["evidence-xm-rsc-initial"],"attributes":{"transactionType":"infrastructure_use","relationshipContext":"internal","transactionStage":"operational","quantity":{"value":6080,"unit":"GPUs","bound":"exact"}}},{"id":"cn-r1-qwen","statement":"R1’s distillation section cites Qwen2.5’s announcement while naming Qwen base models fine-tuned on R1-generated samples.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Specific documented reuse and citation, not inferred influence, partnership or technological independence. The cited work is the announcement, not Qwen2.5’s later technical report.","evidenceIds":["evidence-cn-r1-qwen"]},{"id":"prov-rsp-v2-2-no-train-deploy-without-safeguards","statement":"RSP v2.2 states that its Responsible Scaling Policy, released in September 2023, is a public commitment not to train or deploy models capable of causing catastrophic harm unless Anthropic has implemented safety and security measures that will keep risks below acceptable levels.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"What the edition states, quoted from the cited expression. It is not evidence of implementation, compliance or effectiveness.","evidenceIds":["evidence-prov-rsp-v2-2-no-train-deploy-without-safeguards"],"attributes":{"provision":{"kind":"required_action","key":"rsp-no-train-deploy-without-safeguards"}}},{"id":"prov-rsp-v3-1-marginal-risk-board-ltbt-approval","statement":"RSP v3.1 states that, where marginal risk analysis plays a major role in a decision to move forward, explicit approval of the Risk Report by the Board and LTBT, rather than just the CEO and RSO, will be required.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"What the edition states, quoted from the cited expression. It is not evidence of implementation, compliance or effectiveness.","evidenceIds":["evidence-prov-rsp-v3-1-marginal-risk-board-ltbt-approval"],"attributes":{"provision":{"kind":"decision_authority","key":"rsp-marginal-risk-board-ltbt-approval","parties":[{"role":"decides","actor":"Board of Directors"},{"role":"decides","actor":"Long-Term Benefit Trust"},{"role":"decides","actor":"CEO"},{"role":"decides","actor":"Responsible Scaling Officer"}]}}},{"id":"ts-b3-hubinger-2022","statement":"Hubinger’s post defines deceptive alignment as a model looking aligned on the training data because it is actively trying to look aligned for instrumental reasons, and analyses how likely gradient descent is to find such models under high and low path dependence.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"The fullest treatment located; an editorial judgement.","evidenceIds":["evidence-ts-b3-hubinger-2022"],"attributes":{"thought":{"roles":["most-developed"],"register":"nonfiction"}}},{"id":"yk-y01-time-2023-shut-it-all-down-c","statement":"“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y01. Argues that the dangerous thresholds lie beyond human-level intelligence and cannot be calculated in advance, so a lab could pass them unnoticed. Written in response to the March 2023 open letter calling for a six-month pause.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Premise addressed: We will recognise the moment when it has become dangerous enough to press the switch. Cited from the TIME original rather than the 2023-04-08 LessWrong crosspost. The cached page's HTML title now reads 'The Only Way to Deal With the Threat From AI? Shut It Down'; the on-page headline is the one given in the locator. Hedged (\"may not be obvious\", \"seems imaginable\"); aimed at the open letter's capability threshold rather than the off-switch argument.","evidenceIds":["evidence-yk-t01-d"],"attributes":{"thought":{"roles":["contesting"],"register":"nonfiction","engagement":{"kind":"engages-premise","rationale":"Under the message-level rule, the passage, read with the paragraph, message or dialogue turn it belongs to, gives a reason against the premise that we will recognise the moment when it has become dangerous enough to press the switch."}}}},{"id":"ec-g06-nordhaus-2015-economic-singularity-b","statement":"Nordhaus, “Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth” (2015-09) bears on G06. Six diagnostic tests on US data (productivity growth, capital share, capital prices, capital-output ratio, information capital share, wage growth) plus a demand-side test. For the two positive tests, simple extrapolation puts the crossing point 100 years or more away.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Reclassified from articulates after the cross-family check; the G06 wording question is an open decision in the project document.","evidenceIds":["evidence-ec-g06-a","evidence-ec-g06-b"],"attributes":{"thought":{"roles":["context"],"register":"nonfiction","missing":"(G06-a) Nordhaus reports two of seven tests positive for Singularity; he concludes that it is \"not near\", not that the data show no sign of approach. Under the literal criterion (\"no sign yet\") the record bears on G06. (G06-b) Nordhaus reports two of seven tests positive for Singularity; he concludes that it is \"not near\", not that the data show no sign of approach. Under the literal criterion (\"no sign yet\") the record bears on G06.","attested":{"date":"2016-08-09","note":"Capture of the Cowles PDF."},"tier":{"value":1,"basis":"Tier 1 (unchanged). Economist (Yale) in a core economics series: Cowles Foundation Discussion Paper and NBER Working Paper; journal version in American Economic Journal: Macroeconomics."},"conditions":"Singularity defined in the paper as \"a time when the economic growth rate crosses 20% per year\"; tests on US data to about 2014."}}},{"id":"ec-m06a-aghion-jones-jones-2017-ai-and-economic-growth-b","statement":"Aghion, Jones, Jones, “Artificial Intelligence and Economic Growth” (2017-10) bears on M06a. Argues that Baumol's cost disease can keep the labour share substantial even as automation advances.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"NBER Working Paper 23928, October 2017 (NBER Conference on AI paper; later a chapter in The Economics of Artificial Intelligence). Model result under stated conditions.","evidenceIds":["evidence-ec-m06a-11"],"attributes":{"thought":{"roles":["context"],"register":"nonfiction","missing":"A model in which the labour share stays substantial despite nearly complete automation; it does not state a fall for some workers, though the paper also notes a possible prolonged high-capital-share period.","tier":{"value":1,"basis":"NBER working paper (core venue)."},"conditions":"Elasticity of substitution across tasks below one; automation exogenous; sufficient conditions for balanced growth with a constant capital share well below 100%."}}},{"id":"ec-m06b-cowen-2019-neglected-open-questions-a","statement":"Cowen, “Neglected Open Questions in the Economics of Artificial Intelligence” (2019-05) articulates M06b. Cowen suggests that if AI makes outputs very cheap, wealth-based transfers could complement education and training as policy responses to automation.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Speculative policy options, conditional on AI-driven low consumer prices; framed as 'options to consider'.","evidenceIds":["evidence-ec-m06b-24"],"attributes":{"thought":{"roles":["articulation"],"register":"nonfiction","tier":{"value":1,"basis":"Chapter of the NBER Economics of AI volume (core venue); a short comment chapter, but a chapter of a scholarly volume, not short-form commentary."},"conditions":"Very high productivity and very low output prices from AI."}}},{"id":"lg-oai-sam-altman-nonprofit-board-2019","statement":"OpenAI’s OpenAI LP announcement of 11 March 2019 lists Sam Altman as a member of the OpenAI Nonprofit board and as an OpenAI LP employee with the designation CEO.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"A board list as of 11 March 2019 (the source’s footnote); it does not establish when the person joined or left the board. The source gives the nonprofit’s legal name as OpenAI Inc.","evidenceIds":["lg-ev-oai-sam-altman-nonprofit-board-2019","lg-ev-oai-sam-altman-nonprofit-board-2019-inc"],"attributes":{"role":"Board member; listed as CEO"}},{"id":"lg-cn-kai-fu-lee-google-prior","statement":"01.AI’s management-team page states that Kai-Fu Lee was formerly a global vice president of Google and president of Google Greater China.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"A biography on 01.AI’s page, not a Google record; no tenure is established. English rendering of the Chinese passage is a machine translation; no human translation review.","evidenceIds":["lg-ev-cn-kai-fu-lee-google-prior"],"attributes":{"role":"全球副总裁兼大中华区总裁 (global vice president and president, Greater China)"}},{"id":"lg-gw-tuyls-cofounded-h-company","statement":"TechCrunch reported on 21 May 2024 that Karl Tuyls was one of H’s co-founders.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"The record rests on reporting (TechCrunch); no H Company page naming its founders was retrieved. It does not establish the founding date or his title at H.","evidenceIds":["lg-ev-gw-tuyls-cofounded-h-company"],"attributes":{"role":"co-founder"}},{"id":"cmp-fsf-v2-0-d4-determination","statement":"FSF · v2.0 states a provision on the comparison question “Determination of capability or safeguard adequacy”.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"What the edition states, quoted from the cited expression. It is not evidence of implementation, compliance or effectiveness, and a shared question does not make two editions’ provisions equivalent.","evidenceIds":["ev-cmp-fsf-v2-0-d4-determination-1","ev-cmp-fsf-v2-0-d4-determination-2"],"attributes":{"provision":{"kind":"decision_authority","key":"comparison-d4-determination","facet":{"id":"D4.determination","status":"stated"}}}},{"id":"cmp-faif-2026-d4-oversight","statement":"No passage meeting the comparison question “Board or external oversight” was found in FAIF · June 2026 edition within the inspected scope.","derivation":"inferred","resolution":"supported","inferenceRationale":"Premises: the comparison question asks for the following. A role the edition gives to a board, board committee, trust, or external party in overseeing or reviewing those decisions. The inspected scope of FAIF · June 2026 edition was: Whole text layer, PDF pp. 1–9 (p. 10 blank): all sections read. Method: the edition was read in full and its text layer searched case-insensitively for board, director, external/third, independ, audit, oversight, authorit. Hits examined and not counted: No board or director is mentioned. PDF p. 2 'relevant external research' is an input to risk identification; PDF p. 4 'scalable model oversight' concerns oversight of models; PDF p. 8 feedback from 'other third parties' is an incident channel; PDF p. 9 incident reports to 'relevant authorities' when legally reportable (recorded in “Public or external reporting”) and 'ongoing legal and compliance reviews' (internal) do not give an external party or board a role in reviewing decisions. Confirmed not found on 2026-10-26 by a cross-family archive check of the Internet Archive capture of 2026-07-16 (https://web.archive.org/web/20260716003429/https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf), whose SHA-256 matches the recorded publisher copy. The conclusion is bounded by that scope and method.","editorialState":"seed","qualification":"A recorded absence within the stated scope: not a claim that the provision was deleted, and not a statement about practice or intent.","evidenceIds":["ev-cmp-faif-2026-d4-oversight"],"attributes":{"provision":{"kind":"decision_authority","key":"comparison-d4-oversight","facet":{"id":"D4.oversight"}}}},{"id":"cmp-rsp-v3-2-d7-internal","statement":"RSP · v3.2 states a provision on the comparison question “Internal reporting”.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Also: the Board receives quarterly updates on noncompliance reports (§4); the Frontier Safety Roadmap is shared with every full-time employee, the Board and the LTBT (§2); the CEO and RSO, once they have approved a Risk Report, share their decisions, the report and the internal feedback with the Board and the LTBT (§3.4). What the edition states, paraphrased from the cited expression with its page. It is not evidence of implementation, compliance or effectiveness, and a shared question does not make two editions’ provisions equivalent.","evidenceIds":["ev-cmp-rsp-v3-2-d7-internal-1","ev-cmp-rsp-v3-2-d7-internal-2"],"attributes":{"provision":{"kind":"reporting","key":"comparison-d7-internal","facet":{"id":"D7.internal","status":"stated"}}}},{"id":"c2-governance-aiia-aisc-published","statement":"The Artificial Intelligence Industry Alliance (AIIA) drafted and issued the Artificial Intelligence Safety Commitments (人工智能安全承诺), a text dated December 2024 whose full text CAICT posted on 24 December 2024.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"CAICT’s post says that AIIA drafted and issued the text, and the cover names AIIA. The text and the signatory list are published only as an image in the post; the records quote or describe a transcription of that image. People’s Daily Online’s text reproduction (29 December 2024) omits 红队 (red-team) in Commitment II; the image and the itcloudbd reproduction include it. The text describes itself as industry self-discipline; it is not a law or regulation, and this record does not establish its effect.","evidenceIds":["c2-ev-governance-aiia-aisc-published-1","c2-ev-governance-aiia-aisc-published-2"]},{"id":"c2-moonshot-k2-author-yang","statement":"Zhilin Yang is listed among the authors of the Kimi K2 technical report (arXiv version 1).","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Authorship credit in a team author list does not establish responsibility for particular methods or a management role. The identification rests on the name as printed in the developer’s own author list, which gives no per-author affiliation.","evidenceIds":["c2-ev-moonshot-k2-author-yang"]},{"id":"fl-cn-genai-art7-lawful-sources","statement":"Article 7, point (1), of the Interim Measures requires providers to use data and foundation models that have lawful sources.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Whether any named developer or service falls within the category is set aside. The text does not define \"lawful source\".","evidenceIds":["evidence-fl-cn-genai-art7-lawful-sources-1"],"attributes":{"provision":{"kind":"required_action","key":"cn-genai-art7-1"}}},{"id":"fl-cn-labelling-implicit-metadata","statement":"Article 5 of the Labelling Measures requires service providers to add implicit labels to the file metadata of generated synthetic content, containing production information including the content's attributes, the provider's name or code and a content number, and encourages digital watermarks.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Whether any named developer or service falls within the category is set aside.","evidenceIds":["evidence-fl-cn-labelling-implicit-metadata-1","evidence-fl-cn-labelling-implicit-metadata-2"],"attributes":{"provision":{"kind":"required_action","key":"cn-labelling-art5"}}},{"id":"fl-intl-seoul-declaration-published","statement":"The Department for Science, Innovation and Technology published the Seoul Declaration for safe, innovative and inclusive AI, with the annexed Statement of Intent on AI safety science, on GOV.UK on 21 May 2024.","derivation":"explicit","resolution":"supported","editorialState":"seed","qualification":"Records the GOV.UK publication; the Korean government's publication was not inspected.","evidenceIds":["evidence-fl-intl-seoul-declaration-published-1"]},{"id":"ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat","statement":"The Frontier Safety Framework v3.0 defines its CBRN risk domain as risks of models assisting in the development, preparation and/or execution of a CBRN threat, where v2.0 defined the same domain as risks of assisting in a CBRN attack.","derivation":"inferred","resolution":"supported","inferenceRationale":"Premises: v3.0’s misuse risk domain bullet on PDF p. 4 ends “a chemical, biological, radiological, or nuclear (“CBRN”) threat”; v2.0’s corresponding bullet on PDF p. 2 ends “attack” (quoted in recorded separately). The two bullets are otherwise identically worded. No change note in v3.0 records this difference, so the comparison is Selekoa’s.","editorialState":"seed","qualification":"A one-word difference between the inspected editions. The record does not say that the scope of the domain changed in practice, and no publisher statement interprets the wording.","evidenceIds":["evidence-ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat-1","evidence-ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat-2"],"attributes":{"changeBasis":"document_comparison"}}],"sources":[{"id":"source-rsp-2-2","title":"Responsible Scaling Policy — v2.2","authors":["Anthropic"],"publisher":"Anthropic","publicationDate":null,"version":"Responsible Scaling Policy v2.2; effective date 2025-05-14.","versionDate":"2025-05-14","canonicalUrl":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","evidenceUrl":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20250514223759/https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","archivedAt":"2025-05-14","accessedAt":"2026-09-18","reuse":{"attribution":"Anthropic, Responsible Scaling Policy — v2.2.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-rsp-3-1","title":"Responsible Scaling Policy — v3.1","authors":["Anthropic"],"publisher":"Anthropic","publicationDate":null,"version":"Responsible Scaling Policy v3.1; effective date 2026-04-02.","versionDate":"2026-04-02","canonicalUrl":"https://www-cdn.anthropic.com/files/4zrzovbb/website/bf04581e4f329735fd90634f6a1962c13c0bd351.pdf","evidenceUrl":"https://www-cdn.anthropic.com/files/4zrzovbb/website/bf04581e4f329735fd90634f6a1962c13c0bd351.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20260407193305/https://www-cdn.anthropic.com/files/4zrzovbb/website/bf04581e4f329735fd90634f6a1962c13c0bd351.pdf","archivedAt":"2026-04-07","accessedAt":"2026-09-18","reuse":{"attribution":"Anthropic, Responsible Scaling Policy — v3.1.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-rsp-3-2","title":"Responsible Scaling Policy — v3.2","authors":["Anthropic"],"publisher":"Anthropic","publicationDate":null,"version":"Responsible Scaling Policy v3.2; effective date 2026-04-29.","versionDate":"2026-04-29","canonicalUrl":"https://cdn.sanity.io/files/4zrzovbb/website/28c6241900d90410628a8a2003a5572faae4365a.pdf","evidenceUrl":"https://cdn.sanity.io/files/4zrzovbb/website/28c6241900d90410628a8a2003a5572faae4365a.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20260505141510/https://cdn.sanity.io/files/4zrzovbb/website/28c6241900d90410628a8a2003a5572faae4365a.pdf","archivedAt":"2026-05-05","accessedAt":"2026-09-18","reuse":{"attribution":"Anthropic, Responsible Scaling Policy — v3.2.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-fsf-2-0","title":"Frontier Safety Framework — v2.0","authors":["Google DeepMind"],"publisher":"Google DeepMind","publicationDate":"2025-02-04","version":"Frontier Safety Framework v2.0; publication date 2025-02-04.","versionDate":"2025-02-04","canonicalUrl":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf","evidenceUrl":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20250205011727/https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/updating-the-frontier-safety-framework/Frontier%20Safety%20Framework%202.0.pdf","archivedAt":"2025-02-05","accessedAt":"2026-09-18","reuse":{"attribution":"Google DeepMind, Frontier Safety Framework — v2.0.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-fsf-3-0","title":"Frontier Safety Framework — v3.0","authors":["Google DeepMind"],"publisher":"Google DeepMind","publicationDate":"2025-09-22","version":"Frontier Safety Framework v3.0; publication date 2025-09-22.","versionDate":"2025-09-22","canonicalUrl":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf","evidenceUrl":"https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20250922134959/https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf","archivedAt":"2025-09-22","accessedAt":"2026-09-18","reuse":{"attribution":"Google DeepMind, Frontier Safety Framework — v3.0.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-faif-2026","title":"Frontier Artificial Intelligence Framework — June 2026 edition","authors":["xAI"],"publisher":"xAI","publicationDate":null,"version":"Frontier Artificial Intelligence Framework June 2026 edition; effective date 2026-06-30. Visible public cover identifies the effective edition; stale embedded PDF draft metadata is not used to infer publication status.","versionDate":"2026-06-30","canonicalUrl":"https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf","evidenceUrl":"https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20260716003429/https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf","archivedAt":"2026-07-16","accessedAt":"2026-09-18","reuse":{"attribution":"xAI, Frontier Artificial Intelligence Framework — June 2026 edition.","licence":"No open redistribution licence asserted","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-atlas-openai-lp","title":"OpenAI LP","authors":["Greg Brockman","Ilya Sutskever","OpenAI"],"publisher":"OpenAI","publicationDate":"2019-03-11","version":"Dated publisher page inspected 2026-09-28 and again 2026-10-04 at its current address; the quotes used from 2026-10-04 also match the Internet Archive capture of 2019-03-11 (16:15 UTC) at the original address, which uses straight quotation marks and an additional standfirst.","canonicalUrl":"https://openai.com/index/openai-lp/","evidenceUrl":"https://openai.com/index/openai-lp/","language":"en","archivedUrl":"https://web.archive.org/web/20190311161512/https://openai.com/blog/openai-lp/","archivedAt":"2019-03-11","accessedAt":"2026-10-04","reuse":{"attribution":"Greg Brockman; Ilya Sutskever; OpenAI. OpenAI LP.","licence":"No redistribution licence asserted.","note":"Selekoa practice: metadata, attributed paraphrases and short attributed quotations under the editorial policy (at most 50 words each; at most 400 quoted words per document across the dataset, and about 10% of a document under about 2,000 words). Documents are linked externally, not redistributed. Publisher terms that forbid quotation override this practice."}},{"id":"source-xm-rsc-2022","title":"Introducing the AI Research SuperCluster — Meta’s cutting-edge AI supercomputer for AI research","authors":["Kevin Lee","Shubho Sengupta"],"publisher":"Meta","publicationDate":"2022-01-24","version":"January 24, 2022 announcement; January 28 privacy update explicitly distinguished.","canonicalUrl":"https://ai.meta.com/blog/ai-rsc/","evidenceUrl":"https://ai.meta.com/blog/ai-rsc/","language":"en","archivedUrl":"https://web.archive.org/web/20260921080617/https://ai.meta.com/blog/ai-rsc/","archivedAt":"2026-09-21","accessedAt":"2026-09-29","reuse":{"attribution":"Introducing the AI Research SuperCluster — Meta’s cutting-edge AI supercomputer for AI research — Meta","licence":"No redistribution licence asserted.","note":"Source-linked metadata and brief evidence paraphrases; original text remains with its publisher."}},{"id":"source-cn-r1","title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","authors":["DeepSeek-AI"],"publisher":"DeepSeek-AI","publicationDate":"2025-01-22","version":"Source language: English (en). arXiv:2501.12948v1; fixed first version.","canonicalUrl":"https://arxiv.org/html/2501.12948v1","evidenceUrl":"https://arxiv.org/html/2501.12948v1","language":"en","accessedAt":"2026-09-29","reuse":{"attribution":"DeepSeek-AI, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (2025-01-22).","licence":"No redistribution licence asserted for this source expression","note":"Bibliographic metadata and brief attributed paraphrases only; model licensing is distinct from source-document reuse."}},{"id":"source-hubinger-2022-how-likely-deceptive-alignment","title":"How likely is deceptive alignment?","authors":["Evan Hubinger"],"publisher":"Alignment Forum and LessWrong","publicationDate":"2022-08-30","version":"Text checked in the post’s first stored version through the LessWrong GraphQL API on 2026-10-06; content version 1.0.0.","canonicalUrl":"https://www.alignmentforum.org/posts/A9NxPTwbw6r6Awuwt/how-likely-is-deceptive-alignment","evidenceUrl":"https://www.alignmentforum.org/posts/A9NxPTwbw6r6Awuwt/how-likely-is-deceptive-alignment","language":"en","archivedUrl":"https://web.archive.org/web/20220830201821/https://www.alignmentforum.org/posts/A9NxPTwbw6r6Awuwt/how-likely-is-deceptive-alignment","archivedAt":"2022-08-30","accessedAt":"2026-10-06","reuse":{"attribution":"Evan Hubinger, \"How likely is deceptive alignment?\", Alignment Forum and LessWrong, 2022-08-30.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-time-2023-shut-it-all-down","title":"Pausing AI Developments Isn't Enough. We Need to Shut it All Down","authors":["Eliezer Yudkowsky"],"publisher":"TIME","publicationDate":"2023-03-29","version":"TIME, retrieved 2026-10-06 (SHA-256 4a8c77e15456ca68b530df8bd825296e3923f621b0c0b09dacce312ae94182d6). Page metadata datePublished 2023-03-29T22:01:58Z. The page title currently reads \"The Only Way to Deal With the Threat From AI? Shut It Down\"; the headline is the one crossposted to LessWrong on 2023-04-08.","canonicalUrl":"https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/","evidenceUrl":"https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/","language":"en","archivedUrl":"https://web.archive.org/web/20261003171949/https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/","archivedAt":"2026-10-03","accessedAt":"2026-10-06","reuse":{"attribution":"Eliezer Yudkowsky, “Pausing AI Developments Isn't Enough. We Need to Shut it All Down”, TIME, 2023-03-29.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-nordhaus-2015-cowles","title":"Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth","authors":["William D. Nordhaus"],"publisher":"Cowles Foundation Discussion Paper No. 2021, Yale University","publicationDate":"2015-09","version":"Wayback capture of the Cowles Foundation PDF (cowles.yale.edu/.../d2021.pdf); compared with a capture of NBER Working Paper 21547. Capture retrieved 2026-10-08 (SHA-256 d9348fe608486784675d6f95dfb78367801c0ca33217b500a0d31bee1dea5f97). Versions: Cowles Foundation Discussion Paper No. 2021 (2015-09): checked; NBER Working Paper 21547 (2015-09): compared; American Economic Journal: Macroeconomics 13(1) (2021-01): not compared. Published version: Nordhaus, W. D. (2021). Are we approaching an economic singularity? Information technology and the future of economic growth. American Economic Journal: Macroeconomics, 13(1), 299–332. https://doi.org/10.1257/mac.20170105 Working paper: Nordhaus, W. D. (2015). Cowles Foundation Discussion Paper No. 2021; NBER Working Paper No. 21547. https://doi.org/10.3386/w21547","canonicalUrl":"http://cowles.yale.edu/sites/default/files/files/pub/d20/d2021.pdf","evidenceUrl":"https://web.archive.org/web/20160809230752/http://cowles.yale.edu/sites/default/files/files/pub/d20/d2021.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20160809230752/http://cowles.yale.edu/sites/default/files/files/pub/d20/d2021.pdf","archivedAt":"2016-08-09","accessedAt":"2026-10-08","reuse":{"attribution":"William D. Nordhaus, “Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth”, Cowles Foundation Discussion Paper No. 2021, Yale University, 2015-09.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-aghion-jones-jones-2017-nber-capture","title":"Artificial Intelligence and Economic Growth","authors":["Philippe Aghion","Benjamin F. Jones","Charles I. Jones"],"publisher":"NBER Working Paper 23928","publicationDate":"2017-10","version":"Wayback capture of the NBER PDF (w23928); compared with the author-hosted Version 1.0 (web.stanford.edu/~chadj/AI.pdf, 2018 capture). Capture retrieved 2026-10-08 (SHA-256 a8e68273c738c80fe565afa60bc27ec5e8aaf3135011a0f760233896f1f2100c). Versions: NBER Working Paper 23928 (2017-10): checked; author copy (Stanford), version 1.0 (2017-10-10): compared; book chapter, University of Chicago Press (2019): not retrieved. Published version: Aghion, P., Jones, B. F., & Jones, C. I. (2019). Artificial intelligence and economic growth. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The economics of artificial intelligence: An agenda (pp. 237–290). University of Chicago Press. https://doi.org/10.7208/chicago/9780226613475.003.0009 Working paper: NBER Working Paper No. 23928 (2017). https://doi.org/10.3386/w23928. Re-inspected: Full text. Current copy retrieved directly on 2026-10-09; no archive capture (Wayback unreachable), so earlier bytes are not authenticated. SHA-256 a8e68273c738c80fe565afa60bc27ec5e8aaf3135011a0f760233896f1f2100c. (Merged on 2026-10-29 with the record of the re-inspection, source-aghion-jones-jones-2017-nber, by consistency audit fix B4.)","canonicalUrl":"https://www.nber.org/system/files/working_papers/w23928/w23928.pdf","evidenceUrl":"https://web.archive.org/web/20210303021500/https://www.nber.org/system/files/working_papers/w23928/w23928.pdf","language":"en","archivedUrl":"https://web.archive.org/web/20210303021500/https://www.nber.org/system/files/working_papers/w23928/w23928.pdf","archivedAt":"2021-03-03","accessedAt":"2026-10-09","reuse":{"attribution":"Philippe Aghion, Benjamin F. Jones, Charles I. Jones, “Artificial Intelligence and Economic Growth”, NBER Working Paper 23928, 2017-10.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-cowen-2019-neglected-open-questions","title":"Neglected Open Questions in the Economics of Artificial Intelligence","authors":["Tyler Cowen"],"publisher":"NBER volume chapter","publicationDate":"2019-05","version":"Full text. Current copy retrieved directly on 2026-10-09; no archive capture (Wayback unreachable), so earlier bytes are not authenticated. SHA-256 4501069cee0a059211427ad6e7d6ca65cb809f37623717d15b881350d677377d.","canonicalUrl":"https://www.nber.org/system/files/chapters/c14032/c14032.pdf","evidenceUrl":"https://www.nber.org/system/files/chapters/c14032/c14032.pdf","language":"en","accessedAt":"2026-10-09","reuse":{"attribution":"Tyler Cowen, “Neglected Open Questions in the Economics of Artificial Intelligence”, NBER volume chapter, 2019-05.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"lg-src-cn-01ai-about","title":"关于我们 — 01.AI 零一万物","authors":["01.AI"],"publisher":"01.AI (零一万物)","publicationDate":null,"version":"Undated page, retrieved on 2026-10-15 (SHA-256 c6d11a3575e252ae47dc7548f4c1302dbd289897ab1aaf6c4d2f9b9f7829410d) directly from the publisher. Its wording can change; earlier wording is not authenticated.","canonicalUrl":"https://www.lingyiwanwu.com/about.html","evidenceUrl":"https://www.lingyiwanwu.com/about.html","language":"zh","archivedUrl":"https://web.archive.org/web/20261007110726/https://www.lingyiwanwu.com/about.html","archivedAt":"2026-10-07","accessedAt":"2026-10-15","reuse":{"attribution":"01.AI, “关于我们 — 01.AI 零一万物”.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"lg-src-gw-tc-h-2024","title":"French AI startup H raises $220M seed round","authors":["Romain Dillet"],"publisher":"TechCrunch","publicationDate":"2024-05-21","version":"Dated page, retrieved on 2026-10-15 (SHA-256 3c4d79349f2d9a33d6127ac538312938c246b256978de04fc574e474ccdc5102) directly from the publisher.","canonicalUrl":"https://techcrunch.com/2024/05/21/french-ai-startup-h-raises-220-million-seed-round/","evidenceUrl":"https://techcrunch.com/2024/05/21/french-ai-startup-h-raises-220-million-seed-round/","language":"en","archivedUrl":"https://web.archive.org/web/20260917114204/https://techcrunch.com/2024/05/21/french-ai-startup-h-raises-220-million-seed-round/","archivedAt":"2026-09-17","accessedAt":"2026-10-15","reuse":{"attribution":"Romain Dillet, “French AI startup H raises $220M seed round”, 2024-05-21.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"c2-src-caict-wx-aisc-2024","title":"守护AI安全，共建行业自律典范——首批17家企业签署《人工智能安全承诺》","authors":["中国信通院CAICT"],"publisher":"中国信息通信研究院 (CAICT), WeChat official account 中国信通院CAICT","publicationDate":"2024-12-24","version":"Dated post, retrieved on 2026-10-27 (SHA-256 4a3b099c801423b12830598c841921a74d01d4d9d3f69379669aaf17e6c38d78) directly from the publisher. The commitment text (Chinese and English) and the initial signatory list are a single embedded image (1000 × 9898 pixels, SHA-256 7e9145350d1b882d22ab3f7fc4d4c0226f06194f94264d92291ee065fbf4aa27), not in the page’s text layer; quotations from it are transcriptions. The cross-family check found the same image in an Internet Archive capture of 4 March 2026 with the same SHA-256. Its wording can change; earlier wording is not authenticated. The cross-family check (2026-10-27) located the Internet Archive capture linked here; a capture shows only what the archive retrieved that day. English title (machine translation): “Safeguarding AI safety, building a model of industry self-discipline: initial 17 companies sign the AI Safety Commitments”.","canonicalUrl":"https://mp.weixin.qq.com/s/s-XFKQCWhu0uye4opgb3Ng","evidenceUrl":"https://mp.weixin.qq.com/s/s-XFKQCWhu0uye4opgb3Ng","language":"zh-Hans","archivedUrl":"https://web.archive.org/web/20260724145912/https://mp.weixin.qq.com/s/s-XFKQCWhu0uye4opgb3Ng","archivedAt":"2026-07-24","accessedAt":"2026-10-27","reuse":{"attribution":"中国信通院CAICT, “守护AI安全，共建行业自律典范——首批17家企业签署《人工智能安全承诺》”, 2024-12-24.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"c2-src-kimi-k2-abs-v1","title":"Kimi K2: Open Agentic Intelligence (arXiv abstract page, v1)","authors":["Kimi Team"],"publisher":"arXiv","publicationDate":"2025-07-28","version":"Dated document, retrieved on 2026-10-27 (SHA-256 6a386524d89cd7292c48dbf87bd1eddf37aeef7543cafc1c778738106d4401bf) directly from the publisher. Version: 2507.20534v1. Abstract page of version 1 (submitted 28 July 2025); the page notes a later v2 of 3 February 2026, not inspected.","canonicalUrl":"https://arxiv.org/abs/2507.20534v1","evidenceUrl":"https://arxiv.org/abs/2507.20534v1","language":"en","accessedAt":"2026-10-27","reuse":{"attribution":"Kimi Team, “Kimi K2: Open Agentic Intelligence (arXiv abstract page, v1)”, 2025-07-28.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-cn-generative-ai-measures-2023-govcn","title":"生成式人工智能服务管理暂行办法 (Interim Measures for the Management of Generative Artificial Intelligence Services), gov.cn State Council policy library copy","authors":["国家互联网信息办公室 (Cyberspace Administration of China) and six other departments"],"publisher":"中国政府网 (www.gov.cn), 国务院政策文件库","publicationDate":"2023-07-10","version":"Retrieved on 2026-10-27 (SHA-256 00bd290a9d14e7354b7131d30255daae26ce7ff73588483cbeb2030f56878277). A second posting of the text of the dataset's source for the Measures (the CAC's posting of 13 July 2023): the gov.cn State Council policy library page, which gives its source as 国家网信办网站 and its document date (成文日期) as 2023-07-10. The page body repeats the text twice (a print copy). The two postings were not compared word for word; every passage cited from this source was matched in this copy. The quotation limit is applied to this source and the CAC posting together. The English in parentheses in the title is a machine translation.","canonicalUrl":"https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm","evidenceUrl":"https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm","language":"zh-Hans","accessedAt":"2026-10-27","reuse":{"attribution":"国家互联网信息办公室 (Cyberspace Administration of China) and six other departments, “生成式人工智能服务管理暂行办法 (Interim Measures for the Management of Generative Artificial Intelligence Services), gov.cn State Council policy library copy”, 2023-07-10.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-cn-labelling-measures-2025","title":"关于印发《人工智能生成合成内容标识办法》的通知 (Notice on issuing the Measures for Labelling AI-Generated Synthetic Content), 国信办通字〔2025〕2号","authors":["国家互联网信息办公室 (Cyberspace Administration of China)","工业和信息化部 (Ministry of Industry and Information Technology)","公安部 (Ministry of Public Security)","国家广播电视总局 (National Radio and Television Administration)"],"publisher":"中国政府网 (www.gov.cn), 国务院政策文件库","publicationDate":"2025-03-07","version":"Retrieved on 2026-10-27 (SHA-256 0c0d1d32be9f69843088b0ad2664734b401485166bd04bdfdf35d42d9b77b84a). gov.cn policy-library page (source given as 国家网信办网站; 成文日期 2025-03-07), retrieved 2026-10-27. The English in parentheses in the title is a machine translation.","canonicalUrl":"https://www.gov.cn/zhengce/zhengceku/202503/content_7014286.htm","evidenceUrl":"https://www.gov.cn/zhengce/zhengceku/202503/content_7014286.htm","language":"zh-Hans","accessedAt":"2026-10-27","reuse":{"attribution":"国家互联网信息办公室 (Cyberspace Administration of China), 工业和信息化部 (Ministry of Industry and Information Technology), 公安部 (Ministry of Public Security), 国家广播电视总局 (National Radio and Television Administration), “关于印发《人工智能生成合成内容标识办法》的通知 (Notice on issuing the Measures for Labelling AI-Generated Synthetic Content), 国信办通字〔2025〕2号”, 2025-03-07.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}},{"id":"source-seoul-declaration-page","title":"Seoul Declaration for safe, innovative and inclusive AI: AI Seoul Summit 2024 (publication page)","authors":["UK government"],"publisher":"Department for Science, Innovation and Technology","publicationDate":"2024-05-21","version":"Retrieved on 2026-10-27 (SHA-256 ded2b99c4c78aae39d36bbffb1e8ade04c2dc5b767de9f8a5fcbbfc02a4eb13c). GOV.UK publication page as served on 2026-10-27 (Published 21 May 2024; no update listed) Its wording can change; earlier wording is not authenticated. Publisher as printed on the inspected expression: GOV.UK.","canonicalUrl":"https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024","evidenceUrl":"https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024","language":"en","accessedAt":"2026-10-27","reuse":{"attribution":"UK government, “Seoul Declaration for safe, innovative and inclusive AI: AI Seoul Summit 2024 (publication page)”, 2024-05-21.","licence":"No open redistribution licence asserted","note":"Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution."}}],"evidence":[{"id":"evidence-xm-rsc-initial","sourceRole":"primary","support":"supports","locator":"“Building RSC” opening paragraph and “Phase two and beyond”.","passage":"Meta reported RSC operating with 6,080 GPUs.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-09-29","note":"Passage checked by machine against the identified source expression. No human editorial acceptance or independent verification of implementation is implied."},{"id":"evidence-cn-r1-qwen","sourceRole":"primary","support":"supports","locator":"§2.4 Distillation, first two paragraphs; References, Qwen (2024b)","passage":"The section names Qwen2.5 models and cites Qwen2.5: A Party of Foundation Models.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-09-29","note":"English source paraphrased and checked by machine. Exploratory seed, not editorially accepted. Source reporting is not independent verification of implementation."},{"id":"evidence-prov-rsp-v2-2-no-train-deploy-without-safeguards","sourceRole":"primary","support":"supports","locator":"PDF p. 2 (unnumbered), Executive Summary, first paragraph, first sentence","passage":"In September 2023, we released our Responsible Scaling Policy (RSP), a public commitment not to train or deploy models capable of causing catastrophic harm unless we have implemented safety and security measures that will keep risks below acceptable levels.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-03","note":"Verbatim excerpt checked by an automated research assistant on 2026-10-03 against the inspected PDF, byte-identical to the Internet Archive capture of 2025-05-14. Extraction note: Repeated in the Introduction (PDF p. 5, printed p. 1). Operative forms: §4.3 final sentence (PDF p. 14) and §6.2 'Monitoring pretraining' (PDF p. 15). This is the provision checked for in v3.0–v3.4 (see absences). Human editorial review is pending."},{"id":"evidence-prov-rsp-v3-1-marginal-risk-board-ltbt-approval","sourceRole":"primary","support":"supports","locator":"PDF p. 13 (printed p. 12), §3.4 Procedures, item 5 ('Modified process when marginal risk analysis is important to our case')","passage":"In the event that marginal risk analysis (see previous section) plays a major role in a decision to move forward, explicit approval of the Risk Report by the Board and LTBT (rather than just the CEO and RSO) will be required.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-03","note":"Verbatim excerpt checked by an automated research assistant on 2026-10-03 against the inspected PDF, byte-identical to the Internet Archive capture of 2026-04-07. Extraction note: Same wording in v3.0 (PDF p. 12; not proposed there: quotation limit). A footnote in §3.3 defines marginal risk analysis as arguing that Anthropic's systems impose relatively lower risk given the risks unavoidably posed by other AI systems. Human editorial review is pending."},{"id":"evidence-ts-b3-hubinger-2022","sourceRole":"primary","support":"supports","locator":"Opening section of the transcript","passage":"deceptive alignment is a situation where the reason that your model looks aligned on the training data is because it is actively trying to look aligned for instrumental reasons [...] we're concerned primarily with the schemers and that is the deceptive alignment category.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-06","note":"Checked against the first stored version on 2026-10-06."},{"id":"evidence-yk-t01-d","sourceRole":"primary","support":"supports","locator":"TIME, 'Pausing AI Developments Isn't Enough. We Need to Shut it All Down' (29 March 2023), paragraph beginning 'The key issue is not “human-competitive” intelligence'","passage":"we definitely can’t calculate in advance what happens when, and it currently seems imaginable that a research lab would cross critical lines without noticing.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-06","note":"Candidate pack passage T01-d. The quotation (24 words; trimmed 2026-10-13 from 31 words) matched the retrieved expression (SHA-256 4a8c77e15456ca68b530df8bd825296e3923f621b0c0b09dacce312ae94182d6) when checked on 2026-10-06."},{"id":"evidence-ec-g06-a","sourceRole":"primary","support":"supports","locator":"Abstract","passage":"The tests suggest that the Singularity is not near.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-14","note":"Candidate pack passage G06-a (research/economics-growth). The quotation (9 words) matched the checked expression (nordhaus-cowles) in the pack's verification. Re-checked at admission on 2026-10-14 against a copy retrieved that day."},{"id":"evidence-ec-g06-b","sourceRole":"primary","support":"supports","locator":"Section IX \"Summary of Tests for Singularity\"","passage":"Five of the seven tests are negative for Singularity while two are positive.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-14","note":"Candidate pack passage G06-b (research/economics-growth). The quotation (13 words) matched the checked expression (nordhaus-cowles) in the pack's verification. Re-checked at admission on 2026-10-14 against a copy retrieved that day."},{"id":"evidence-ec-m06a-11","sourceRole":"primary","support":"supports","locator":"pdf page 47; section 7 Conclusion (printed p. 46)","passage":"The labor share remains substantial because of Baumol's insight: growth is determined not by what we are good at but rather by what is essential and yet hard to improve.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-14","note":"Candidate pack passage M06a-11 (research/economics-mainstream). The quotation (30 words) matched the retrieved current copy (SHA-256 a8e68273c738c80fe565afa60bc27ec5e8aaf3135011a0f760233896f1f2100c) in the pack's verification. Re-checked at admission on 2026-10-14 against a copy retrieved that day."},{"id":"evidence-ec-m06b-24","sourceRole":"primary","support":"supports","locator":"pdf page 3; 15.1 The Distribution of Consumer Surplus (printed p. 392)","passage":"To the extent productivity is very high and prices are very low, it may suffice for workers to own some capital or natural resources. … Giving everyone some land, a birthright grant or shares in a sovereign wealth fund are options to consider","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-14","note":"Candidate pack passage M06b-24 (research/economics-mainstream). The quotation (43 words) matched the retrieved current copy (SHA-256 4501069cee0a059211427ad6e7d6ca65cb809f37623717d15b881350d677377d) in the pack's verification. Re-checked at admission on 2026-10-14 against a copy retrieved that day."},{"id":"lg-ev-oai-sam-altman-nonprofit-board-2019","sourceRole":"primary","support":"supports","locator":"Section “Who’s involved”, board list (footnote A: board as of 11 March 2019)","passage":"The board list names Sam Altman among the OpenAI LP employees on OpenAI Nonprofit’s board, with the designation CEO.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-15","note":"Candidate pack passage, lab genealogy (research/lab-genealogy). The paraphrase’s verbatim anchor “Sam Altman (CEO)” matched the retrieved text on 2026-10-15 and again in a fresh retrieval the same day. Read through the third-party reader relay r.jina.ai, which renders the publisher’s current page (the publisher refused direct retrieval); not an archive copy. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. The paraphrase is Selekoa’s; only the anchor is the source’s wording. No human editorial review.","retrievalRoute":"reader_relay"},{"id":"lg-ev-oai-sam-altman-nonprofit-board-2019-inc","sourceRole":"primary","support":"context","locator":"Paragraph defining the general partner","passage":"OpenAI’s 2019 note gives the legal name of OpenAI Nonprofit as OpenAI Inc.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-15","note":"Candidate pack passage, lab genealogy (research/lab-genealogy). The paraphrase’s verbatim anchor “whose legal name is “OpenAI Inc”” matched the retrieved text on 2026-10-15 and again in a fresh retrieval the same day. Read through the third-party reader relay r.jina.ai, which renders the publisher’s current page (the publisher refused direct retrieval); not an archive copy. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. The paraphrase is Selekoa’s; only the anchor is the source’s wording. No human editorial review.","retrievalRoute":"reader_relay"},{"id":"lg-ev-cn-kai-fu-lee-google-prior","sourceRole":"primary","support":"supports","locator":"关于我们, 管理团队 (Management team), profile of 李开复","passage":"此前曾任谷歌全球副总裁兼大中华区总裁","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-15","note":"Candidate pack passage, lab genealogy (research/lab-genealogy). The quotation (1 words) matched the retrieved text on 2026-10-15 and again in a fresh retrieval the same day. Read directly from the publisher. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. No human editorial review.","retrievalRoute":"direct"},{"id":"lg-ev-gw-tuyls-cofounded-h-company","sourceRole":"secondary","support":"supports","locator":"Paragraph 3","passage":"The four other co-founders all previously worked for DeepMind, the AI company owned by Google: Karl Tuyls was a research director at DeepMind","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-15","note":"Candidate pack passage, lab genealogy (research/lab-genealogy). The quotation (23 words) matched the retrieved text on 2026-10-15 and again in a fresh retrieval the same day. Read directly from the publisher. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. A cross-family machine check (2026-10-16/17) re-read the record’s sources and agreed with the record. No human editorial review.","retrievalRoute":"direct"},{"id":"ev-cmp-fsf-v2-0-d4-determination-1","sourceRole":"primary","support":"supports","locator":"PDF p. 7 (printed p. 7), Governance and Accountability, “Governance” paragraph, first sentence","passage":"For Google models, when alert thresholds are reached, the response plan will be reviewed and approved by appropriate corporate governance bodies such as the Google DeepMind AGI Safety Council, Google DeepMind Responsibility and Safety Council, and/or Google Trust & Compliance Council.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-27","note":"Existing dataset passage for this edition, which contains the comparison’s shorter quotation, first checked on 2026-10-03. The comparison’s quotation was checked by an automated research assistant on 2026-10-27 against the publisher copy, whose SHA-256 matches the copy retrieved on 2026-10-15. Applies 'when alert thresholds are reached'. The sentence also names the Google DeepMind Responsibility and Safety Council and/or Google Trust & Compliance Council, and opens 'For Google models'. Human editorial review is pending."},{"id":"ev-cmp-fsf-v2-0-d4-determination-2","sourceRole":"primary","support":"supports","locator":"PDF p. 4 (printed p. 4), Misuse, Mitigation Approach, “Deployment Mitigations”, numbered step 2","passage":"Pre-deployment review of safety case: general availability deployment8 of a model takes place only after the appropriate corporate governance body determines the safety case regarding each CCL the model has reached to be adequate.","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-27","note":"Existing dataset passage for this edition, which contains the comparison’s shorter quotation, first checked on 2026-10-03. The comparison’s quotation was checked by an automated research assistant on 2026-10-27 against the publisher copy, whose SHA-256 matches the copy retrieved on 2026-10-15. Footnote marker 8 omitted at the ellipsis. Human editorial review is pending."},{"id":"ev-cmp-faif-2026-d4-oversight","sourceRole":"primary","support":"supports","locator":"FAIF · June 2026 edition: Whole text layer, PDF pp. 1–9 (p. 10 blank): all sections read","passage":"Within the inspected scope, no passage meeting the comparison question “Board or external oversight” was found.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-26","note":"Absence recorded by an automated research assistant on 2026-10-15 for the framework comparison, audited in a fresh context, checked by a reviewer from a different model family on 2026-10-15 and re-checked on 2026-10-26 against an Internet Archive capture, because the publisher copy was no longer served (research/archive-checks-2026-10-26/codex-report.md). Human editorial review is pending."},{"id":"ev-cmp-rsp-v3-2-d7-internal-1","sourceRole":"primary","support":"supports","locator":"PDF p. 16 (printed p. 15), §4 Governance, item 3 ('Internal transparency')","passage":"The edition says final Risk Reports, without redactions, will be shared with Anthropic’s regular-clearance staff.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-27","note":"Explicit paraphrase, written on 2026-10-27 by an automated research assistant after re-reading the passage in the publisher copy, whose SHA-256 matches the copy retrieved on 2026-10-15. The comparison quoted the passage; it is given here as a paraphrase to keep the edition within the quotation limit. Human editorial review is pending."},{"id":"ev-cmp-rsp-v3-2-d7-internal-2","sourceRole":"primary","support":"supports","locator":"PDF p. 16 (printed p. 15), §4 Governance, item 2 ('Long Term Benefit Trust')","passage":"The edition says Anthropic will brief the LTBT regularly on plans and developments relating to its Risk Reports, including model training.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-27","note":"Explicit paraphrase, written on 2026-10-27 by an automated research assistant after re-reading the passage in the publisher copy, whose SHA-256 matches the copy retrieved on 2026-10-15. The comparison quoted the passage; it is given here as a paraphrase to keep the edition within the quotation limit. Human editorial review is pending."},{"id":"c2-ev-governance-aiia-aisc-published-1","sourceRole":"primary","support":"supports","locator":"Post, first paragraph (text layer)","passage":"中国人工智能产业发展联盟 ... 起草发布《人工智能安全承诺》","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-27","note":"Candidate pack passage, Chinese frontier-AI institutions, second collection (research/china-labs-2). The quotation (24 words, each CJK character counted as one) matched the retrieved text on 2026-10-27 and again in a fresh retrieval the same day. Read directly from the publisher. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. Elided quotation: the ellipses mark omitted words. English machine translation, not verified by a person: “the Artificial Intelligence Industry Alliance of China ... drafted and issued the AI Safety Commitments”. A cross-family machine check (2026-10-27, item K008) re-read the record’s sources and agreed with the record. No human editorial review.","retrievalRoute":"direct"},{"id":"c2-ev-governance-aiia-aisc-published-2","sourceRole":"primary","support":"supports","locator":"Embedded image, cover (transcribed from the image; not in the page’s text layer)","passage":"Transcription of the cover in the image: the AIIA logo, the title 人工智能安全承诺 / ARTIFICIAL INTELLIGENCE SAFETY COMMITMENTS, and the imprint 中国人工智能产业发展联盟 2024年12月.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-27","note":"Candidate pack passage, Chinese frontier-AI institutions, second collection (research/china-labs-2). Transcription of text that appears only in the image embedded in the post (the cover); it is not in the page’s text layer and was not machine-matched. The drafting pass read the image visually, and the cross-family check (2026-10-27) read the image’s cover, Commitments I, II and VI and all 17 listed names and found them in agreement with the pack’s reading. No English rendering is given; no human translation review. A cross-family machine check (2026-10-27, item K008) re-read the record’s sources and agreed with the record. No human editorial review.","retrievalRoute":"direct"},{"id":"c2-ev-moonshot-k2-author-yang","sourceRole":"primary","support":"supports","locator":"arXiv abstract page, author list","passage":"Zhen Yang, Zhilin Yang, Zonghan Yang","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-27","note":"Candidate pack passage, Chinese frontier-AI institutions, second collection (research/china-labs-2). The quotation (6 words) matched the retrieved text on 2026-10-27 and again in a fresh retrieval the same day. Read directly from the publisher. Matching ignored differences of whitespace, case, quotation marks, apostrophes, dashes and Unicode compatibility forms, so the source’s typography can differ. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-fl-cn-genai-art7-lawful-sources-1","sourceRole":"primary","support":"supports","locator":"Article 7 (第七条), point (1)","passage":"使用具有合法来源的数据和基础模型","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-28","note":"Candidate pack passage, frameworks to law (research/frameworks-to-law). The quotation (16 words, each CJK character counted as one) matched the text retrieved for the candidate pack on 2026-10-27 and again at admission on 2026-10-28, in a fresh retrieval whose SHA-256 equals the pack’s. Read directly from www.gov.cn. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. English machine translation, not verified by a person: “use data and foundation models that have lawful sources”. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-fl-cn-labelling-implicit-metadata-1","sourceRole":"primary","support":"supports","locator":"Article 5 (第五条), first paragraph","passage":"隐式标识包含生成合成内容属性信息、服务提供者名称或者编码、内容编号等制作要素信息","passageType":"quote","verification":"machine_checked","checkedAt":"2026-10-28","note":"Candidate pack passage, frameworks to law (research/frameworks-to-law). The quotation (38 words, each CJK character counted as one) matched the text retrieved for the candidate pack on 2026-10-27 and again at admission on 2026-10-28, in a fresh retrieval whose SHA-256 equals the pack’s. Read directly from www.gov.cn. Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. English machine translation, not verified by a person: “the implicit label contains production-element information such as attribute information of the generated synthetic content, the service provider's name or code, and a content number”. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-fl-cn-labelling-implicit-metadata-2","sourceRole":"primary","support":"supports","locator":"Article 5 (第五条), first and second paragraphs","passage":"The implicit label is to be added to the file metadata in accordance with Article 16 of the Deep Synthesis Provisions; providers are encouraged to add implicit labels such as digital watermarks in the content.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-28","note":"Candidate pack passage, frameworks to law (research/frameworks-to-law). The paraphrase is Selekoa’s, written from the text retrieved on 2026-10-27 and read directly from www.gov.cn; it is not a quotation and was not machine-matched. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-fl-intl-seoul-declaration-published-1","sourceRole":"primary","support":"supports","locator":"Publication page metadata","passage":"Paraphrase: the page, from the Department for Science, Innovation and Technology, presents the Seoul Declaration and Statement of Intent toward International Cooperation on AI Safety Science, agreed by world leaders at the AI Seoul Summit, 21 May 2024, with 'Published: 21 May 2024'.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-28","note":"Candidate pack passage, frameworks to law (research/frameworks-to-law). The paraphrase is Selekoa’s, written from the text retrieved on 2026-10-27 and read directly from www.gov.uk; it is not a quotation and was not machine-matched. A cross-family machine check (2026-10-28, item 8) answered the record’s question blind from fresh retrievals and agreed with the record. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat-1","sourceRole":"primary","support":"supports","locator":"PDF p. 4 (printed p. 4), list of misuse risk domains, first bullet","passage":"The bullet defines the CBRN domain as risks of models assisting in the development, preparation, and/or execution of a chemical, biological, radiological, or nuclear threat.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-30","note":"Candidate pack passage, AI and biosecurity (research/ai-bio). The paraphrase is Selekoa’s, written from the text retrieved for the candidate pack on 2026-10-28; it is not a quotation and was not machine-matched. A cross-family machine check (2026-10-30, item 8) re-read the record’s sources and found full support. No human editorial review.","retrievalRoute":"direct"},{"id":"evidence-ab-fw-fsf-v3-0-cbrn-domain-attack-to-threat-2","sourceRole":"primary","support":"supports","locator":"PDF p. 2 (printed p. 2), list of misuse risk domains, first bullet","passage":"The bullet defines the CBRN domain as risks of models assisting in the development, preparation, and/or execution of a chemical, biological, radiological, or nuclear attack.","passageType":"paraphrase","verification":"machine_checked","checkedAt":"2026-10-30","note":"Candidate pack passage, AI and biosecurity (research/ai-bio). The paraphrase is Selekoa’s, written from the text retrieved for the candidate pack on 2026-10-28; it is not a quotation and was not machine-matched. A cross-family machine check (2026-10-30, item 8) re-read the record’s sources and found full support. No human editorial review.","retrievalRoute":"direct"}],"at":{"entities":[13,28,39,44,46,47,50,51,57,157,170,176,184,207,233,336,350,383,503,509,518,519,526,527,553,679,713,718,829,935,936,964,1009,1020],"claims":[349,364,410,419,982,987,1279,1399,1436,1602,1939,1998,2104,2147,2561,2748,2801,2901,2910,2947,3272],"sources":[38,40,41,45,46,52,121,156,163,341,350,477,478,506,577,687,778,790,836,837,852],"evidence":[372,387,442,451,1207,1213,1600,1601,1745,1786,2007,2008,2503,2582,2725,2726,2786,3216,3217,3461,3462,3567,3741,3756,3757,3806,4487,4488]},"version":"1lg47t7"}