{"$schema":"https://policywindow.org/wiki/substantiation-ledger.json","name":"Policy Window — substantiation-closure ledger","description":"Every genuine uncited factual prose claim, classified by an offline adversarial panel into its substantiation route: primary_source_backed (a claim about the article's subject, established by its cited primary source), corpus_backed (a specific in-corpus literature item supports it — refId given), needs_external_ref (a checkable claim needing a source not yet present — the actionable sourcing queue), or not_a_claim (framing the factual detector over-flagged). Complements /wiki/claim-substantiation (which measures the inline-citation rate); this CLOSES the residual or honestly queues it. Re-derivable.","docs":"https://policywindow.org/wiki/claim-substantiation","method":"Offline adversarial classifier panel: for each genuine uncited factual prose claim, one classifier assigns a substantiation ROUTE against the article's subject (its cited primary source) and its in-corpus literature — primary_source_backed (a claim about the subject, established by the article's primary source), corpus_backed (a specific literature item supports it; refId recorded), needs_external_ref (a checkable claim needing a source not yet present), or not_a_claim (framing the detector over-flagged). Refute-by-default: default to needs_external_ref; credit a route only when clearly established. Credits (primary/corpus) are independently re-verified for over-crediting. Committed here; the serve-time report reads it deterministically (no serve-time model call).","verificationNote":"Independently over-crediting-audited (8 skeptical auditors, refute-by-default): 11 of 85 credits demoted (10 to needs_external_ref, 1 to not_a_claim) — mostly claims about external EVENTS or OTHER instruments the classifier had over-attributed to the subject’s primary source. Corrections applied here; the closureRate is post-audit. RESOLUTION AUDIT: the 155 fetch-verified resolutions were then independently re-audited (refute-by-default, 10 auditors) for quote-establishes-the-specific-claim; 28 whose quote was merely TOPICAL (did not establish the load-bearing date/threshold/mechanism) were demoted back to needs_external_ref. The recorded resolutions are the survivors. | 2026-07-02 cross-audit: 4 externally_resolved credits demoted to needs_external_ref (recorded source did not support the claim); orphaned routes pruned after prose corrections. | 2026-07-02 cross-audit finding [17]: the not_a_claim route (73 entries, previously never independently re-audited) was re-audited by two independent judges; 58 entries found to contain checkable factual limbs were returned to the UNCLASSIFIED (pending) queue rather than remain excluded — an honest correction of an over-broad exclusion. Remaining not_a_claim entries survived the re-audit.","resolutionMethod":"needs_external_ref claims closed by a two-stage pipeline: (1) research a real supporting source via web search; (2) an INDEPENDENT verifier fetches the source and must quote the passage that establishes the claim (refute-by-default). Only fetch-verified references (with a supporting quote) are recorded as externally_resolved.","resolvedAt":"2026-07-01","classifiedAt":"2026-07-01","version":"12","summary":{"candidates":291,"classified":165,"pending":126,"primarySourceBacked":62,"corpusBacked":6,"externallyResolved":61,"needsExternalRef":21,"notAClaim":15,"closureRate":0.86},"resolvedReferences":[{"kind":"topic","slug":"agentic-systems-governance","sentence":"The G7 Hiroshima Process International Code of Conduct (adopted 30 October 2023) sets eleven voluntary actions — risk identification across the lifecycle, red-teaming, incident reporting — now monitored via an OECD reporting framework (OECD, February 2025).","citation":"G7, \"Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems,\" 30 October 2023; and OECD.AI Hiroshima AI Process (HAIP) Reporting Framework, 2024-2025.","url":"https://www.mofa.go.jp/files/100573473.pdf","sourceType":"primary/official (G7 code of conduct, MOFA Japan; OECD reporting framework)","supportingQuote":"Adoption Date: October 30, 2023 ... the document lists 11 actions. Action 1: 'employing diverse internal and independent external testing measures, through a combination of methods for evaluations, such as red-teaming' and 'to identify, evaluate, and mitigate risks across the AI lifecycle.' Action 4"},{"kind":"topic","slug":"ai-worker-displacement","sentence":"The international layer is firming: the Council of Europe Framework Convention on AI (2024) explicitly contemplates \"socio-economic aspects, such as employment and labour\" among AI's impacts, signalling a possible future binding hook (Council of Europe 2024).","citation":"Council of Europe, Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), opened for signature 5 Sept. 2024, preamble","url":"https://rm.coe.int/1680afae3c","sourceType":"primary_official","supportingQuote":"including, but not limited to, human health and the environment, and socio-economic aspects, such as employment and labour"},{"kind":"topic","slug":"ai-worker-displacement","sentence":"\"Stick\" proposals tax automation — Maryland's withdrawn HB 314 would have levied roughly USD 900 per displaced worker to fund placement and retraining, reducible by half for employers offering twelve weeks' severance or in-house redeployment — while \"carrot\" bills (e.g., pending New Jersey measures) reward hiring displaced workers and fund apprenticeships (Bloomberg Law 2025; Potomac Legal Group 2026).","citation":"Potomac Legal Group, \"Maryland Withdraws 'AI Tax' Bill for Worker Displacement\" (2026); Maryland General Assembly, HB 314 (2026 Regular Session)","url":"https://potomaclegalgroup.com/maryland-withdraws-ai-tax-bill-for-worker-displacement/","sourceType":"reputable_secondary","supportingQuote":"The legislation proposed an assessment of $900.00 per displaced employee ... Employers could reduce this assessment by fifty percent if they provided at least twelve weeks of severance pay ... the bill was recently withdrawn by its sponsor as of February 17, 2026"},{"kind":"benchmark","slug":"aime-2024","sentence":"Frontier scores on AIME 2024 climbed from near-floor to near-ceiling within roughly a year, driven by the shift from general-purpose to inference-time-reasoning models.","citation":"OpenAI, \"Learning to reason with LLMs\" (September 12, 2024). On the 2024 AIME exams, GPT-4o solved on average 12% (1.8/15) of problems, while o1 averaged 74% (11.1/15) with a single sample, 83% with consensus among 64 samples, and 93% when re-ranking 1000 samples.","url":"https://openai.com/index/learning-to-reason-with-llms/","sourceType":"official-primary","supportingQuote":"We evaluated math performance on AIME, an exam designed to challenge the brightest high school math students in America. On the 2024 AIME exams, GPT-4o only solved on average 12% (1.8/15) of problems. o1 averaged 74% (11.1/15) with a single sample per problem, 83% (12.5/15) with consensus among 64 s"},{"kind":"concept","slug":"alignment","sentence":"The formal apparatus accumulated through the 2010s: corrigibility — an agent's tolerance of correction and shutdown — was given a decision-theoretic treatment by Soares, Fallenstein, Armstrong and Yudkowsky (\"Corrigibility,\" AAAI-15 workshop, 2015), and preference-learning mechanics matured with Christiano et al.","citation":"Nate Soares, Benja Fallenstein, Stuart Armstrong & Eliezer Yudkowsky, \"Corrigibility,\" Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI-15), AI & Ethics Workshop, 2015.","url":"https://intelligence.org/files/Corrigibility.pdf","sourceType":"peer_reviewed","supportingQuote":"We call an AI system \"corrigible\" if it cooperates with what its creators regard as a corrective intervention, despite default incentives for rational agents to resist attempts to shut them down or modify their preferences."},{"kind":"concept","slug":"chain-of-thought-monitoring","sentence":"The oversight opportunity, and its central hazard, were demonstrated by Baker et al.","citation":"Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madry, A., Zaremba, W., Pachocki, J., & Farhi, D. (2025). Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. arXiv:2503.11926.","url":"https://arxiv.org/abs/2503.11926","sourceType":"peer_reviewed","supportingQuote":"CoT monitoring can be far more effective than monitoring agent actions and outputs alone ... agents learn obfuscated reward hacking, hiding their intent within the CoT while still exhibiting a significant rate of reward hacking"},{"kind":"topic","slug":"compute-export-controls","sentence":"The dominant instrument is the US Export Administration Regulations (EAR): the October 2023 BIS rules control advanced chips via performance-defined classifications — ECCN 3A090/4A090, the .a tier capturing ICs at total-processing-performance ≥ 4800 (BIS, 88 FR, doc. 2023-23055).","citation":"Bureau of Industry and Security (BIS), \"Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections,\" 88 Fed. Reg. (Oct. 25, 2023), doc. 2023-23055.","url":"https://www.federalregister.gov/documents/2023/10/25/2023-23055/implementation-of-additional-export-controls-certain-advanced-computing-items-supercomputer-and","sourceType":"primary_official","supportingQuote":"BIS expanded ECCN 3A090.a to now broadly control ICs with one or more digital processing units having either (a) a \"total processing performance\" of 4800 or more, or (2) a \"total processing performance\" of 1600 or more and a \"performance density\" of 5.92 or more."},{"kind":"topic","slug":"compute-export-controls","sentence":"January 2025's \"Framework for AI Diffusion\" extended this to intangibles, creating ECCN 4E091 for closed-weight models trained on >10^26 operations (BIS, 90 FR, doc. 2025-00636).","citation":"U.S. Bureau of Industry and Security, \"Framework for Artificial Intelligence Diffusion,\" 90 FR 4544, Federal Register doc. 2025-00636 (Jan. 15, 2025).","url":"https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion","sourceType":"primary_official","supportingQuote":"The BIS Framework for Artificial Intelligence Diffusion added a new ECCN 4E091, which specifies model weights of any closed weight model that has been trained on more than 10^26 computation operations."},{"kind":"topic","slug":"compute-reporting","sentence":"A second fault line is the threshold's numeric level and durability: the EU set 10^25 FLOP, the US BIS proposal 10^26 FLOP (89 Fed.","citation":"Bureau of Industry and Security (BIS), U.S. Dept. of Commerce, \"Establishment of Reporting Requirements for the Development of Advanced Artificial Intelligence Models and Computing Clusters,\" Proposed Rule, 89 Fed. Reg. 73021 (Sept. 11, 2024); EU AI Act (Regulation 2024/1689) Article 51(2).","url":"https://www.federalregister.gov/documents/2024/09/11/2024-20529/establishment-of-reporting-requirements-for-the-development-of-advanced-artificial-intelligence","sourceType":"primary / official (Federal Register proposed rule + EU legislative text)","supportingQuote":"A dual-use foundation model training run triggers reporting requirements if it utilizes more than 10^26 computational operations (e.g., integer or floating-point operations)."},{"kind":"topic","slug":"criminal-justice","sentence":"Commentators dispute whether the EU ban bites at all, since the \"solely\"-profiling threshold and the carve-out for tools supporting fact-based human assessment may leave most deployed predictive-policing systems untouched (Free, European Law Blog, 2024; Future of Privacy Forum analysis, 2026).","citation":"Future of Privacy Forum, \"Red Lines under the EU AI Act: Unpacking the Prohibition of Individual Risk Assessment for the Prediction of Criminal Offences\" (March 11, 2026); and D. Free, \"Predictive Policing in the AI Act: meaningful ban or paper tiger?\" European Law Blog (2024).","url":"https://fpf.org/blog/red-lines-under-the-eu-ai-act-unpacking-the-prohibition-of-individual-risk-assessment-for-the-prediction-of-criminal-offences/","sourceType":"institutional analysis (think tank / policy)","supportingQuote":"The risk assessments covered by the analyzed provision are only prohibited when they are based solely on the profiling of a person or the assessment of their personality traits and characteristics. ... Importantly, the prohibition does not apply when AI systems support human assessment regarding a p"},{"kind":"topic","slug":"criminal-justice","sentence":"The EU AI Act's prohibition on solely-profiling individual crime prediction became applicable on 2 February 2025, the first hard legal constraint specific to predictive policing anywhere (Reg.","citation":"European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5(1)(d); prohibitions applicable 2 February 2025 (six months after entry into force on 1 August 2024)","url":"https://artificialintelligenceact.eu/article/5/","sourceType":"primary_official","supportingQuote":"the placing on the market, the putting into service for this specific purpose, or the use of an AI system for making risk assessments of natural persons in order to assess or predict the risk of a natural person committing a criminal offence, based solely on the profiling of a natural person or on a"},{"kind":"topic","slug":"criminal-justice","sentence":"The EU treats one application — purely profiling-based individual crime prediction — as an unacceptable risk warranting an outright ban (Reg.","citation":"Future of Privacy Forum, 'Red Lines under the EU AI Act: Unpacking the Prohibition of Individual Risk Assessment for the Prediction of Criminal Offences' (2025); Regulation (EU) 2024/1689 Art. 5(1)(d)","url":"https://fpf.org/blog/red-lines-under-the-eu-ai-act-unpacking-the-prohibition-of-individual-risk-assessment-for-the-prediction-of-criminal-offences/","sourceType":"institutional","supportingQuote":"Article 5(1)(d) AI Act establishes a crucial prohibition on AI systems that assess or predict the likelihood of natural persons committing criminal offenses based solely on profiling or personality assessment."},{"kind":"topic","slug":"criminal-justice","sentence":"The Council of Europe Convention, the only binding treaty here, exempts national-security and defence activities and lets each Party choose whether to apply its rules to private actors at all — an opt-in for the private sector that civil-society groups and the European Data Protection Supervisor warned could hollow out protection and shelter state surveillance (CETaS/Alan Turing Institute, 2024; CAIDP treaty brief, 2025).","citation":"Centre for Emerging Technology and Security (CETaS), The Alan Turing Institute, \"The Council of Europe Convention on AI: National Security Implications\" (September 2024); and Center for AI and Digital Policy (CAIDP), \"International AI Treaty\" resource brief.","url":"https://cetas.turing.ac.uk/publications/council-europe-convention-ai-national-security-implications","sourceType":"institutional research (national research institute) + civil-society treaty brief","supportingQuote":"Parties to the convention will not be required to apply the treaty's provisions to activities related to the protection of national security interests ... The convention will not apply to national defence matters... [and] The convention offers parties two ways of complying with its principles and ob"},{"kind":"topic","slug":"deepfakes","sentence":"(4) Notice-and-takedown: the US TAKE IT DOWN Act requires covered platforms to remove notified non-consensual intimate imagery, including digital forgeries, within 48 hours, enforced by the FTC (TAKE IT DOWN Act, Pub.","citation":"TAKE IT DOWN Act, Pub. L. No. 119-12 (S.146), 119th Congress, signed May 19, 2025; enforced by the FTC (Section 3).","url":"https://www.congress.gov/bill/119th-congress/senate-bill/146/text","sourceType":"primary_official","supportingQuote":"a covered platform shall, as soon as possible, but not later than 48 hours after receiving such request-- (A) remove the intimate visual depiction ... 'digital forgery' ... A failure to reasonably comply ... shall be treated as a violation of a rule defining an unfair or a deceptive act or practice "},{"kind":"topic","slug":"development-rights-framing","sentence":"Third is whether the frame is a critique or a tested prescription: the extraction it names is empirically anchored, but cost evidence on its commonest proxy, data localisation, points the other way — Ferracane and van der Marel (2021) and Bauer et al.","citation":"Ferracane, M. F., & van der Marel, E. (2021). \"Do data policy restrictions inhibit trade in services?\" Review of World Economics, 157(4), 727-776; Bauer, M., Lee-Makiyama, H., van der Marel, E., & Verschelde, B. (2014). \"The Costs of Data Localisation: Friendly Fire on Economic Recovery,\" ECIPE Occasional Paper No. 3/2014.","url":"https://link.springer.com/article/10.1007/s10290-021-00417-2","sourceType":"peer_reviewed","supportingQuote":"strict data policies negatively and significantly impact imports of data-intense services. More specifically, restrictions on the cross-border movement of data are found to significantly reduce imports of services"},{"kind":"topic","slug":"environmental-impact-of-training","sentence":"Outside AI-specific law, France regulates the same footprint through general digital-environment rules: the REEN Act (Loi n° 2021-1485) underpins the ARCEP–ADEME digital-footprint observatory created in December 2024 (ARCEP, Dec. 2024).","citation":"ARCEP & ADEME, \"L'Arcep et l'ADEME créent l'observatoire des impacts environnementaux du numérique\" (12 Dec. 2024); Loi n° 2021-1485 (REEN Act), Art. 4.","url":"https://www.arcep.fr/actualites/actualites-et-communiques/detail/n/environnement-121224.html","sourceType":"primary_official","supportingQuote":"l'Arcep et l'ADEME (Agence de la transition écologique) ont décidé de regrouper dans l'observatoire des impacts environnementaux du numérique ... La création de l'observatoire, prévue par la loi visant à réduire l'empreinte environnementale du numérique en France (dite « loi REEN »)"},{"kind":"topic","slug":"environmental-impact-of-training","sentence":"At national level, France's ARCEP–ADEME observatory (created December 2024 under the REEN Act) is extending verified digital-footprint reporting toward AI-specific lifecycle stages (ARCEP, Dec. 2024).","citation":"ARCEP & ADEME, \"L'Arcep et l'ADEME créent l'observatoire des impacts environnementaux du numérique\" (12 Dec. 2024), created under REEN Act (Loi n° 2021-1485), Art. 4.","url":"https://www.arcep.fr/actualites/actualites-et-communiques/detail/n/environnement-121224.html","sourceType":"primary_official","supportingQuote":"quantifier les impacts directs et indirects du numérique sur l'environnement ainsi que la contribution apportée par le numérique, notamment l'intelligence artificielle"},{"kind":"topic","slug":"environmental-impact-of-training","sentence":"Building on this, the European Commission ran a targeted consultation on measuring the energy consumption and emissions of AI models from 7 April to 1 June 2026, explicitly to design a measurement framework for the Act's energy objectives and a possible AI energy-and-emissions label spanning training and inference (European Commission, Apr.–June 2026).","citation":"European Commission, DG CNECT, \"Targeted consultation on measuring energy consumption and emissions of AI models and systems\" (7 Apr.–1 Jun. 2026), Shaping Europe's Digital Future.","url":"https://digital-strategy.ec.europa.eu/en/consultations/targeted-consultation-measuring-energy-consumption-and-emissions-ai-models-and-systems","sourceType":"primary_official","supportingQuote":"Responses to this consultation will help refine the study and contribute to a measurement framework for the energy-related objectives of the AI Act and support the design of a potential AI energy and emission label ... during both the development (training) and operational (inference) stages."},{"kind":"topic","slug":"environmental-impact-of-training","sentence":"In the United States, the Artificial Intelligence Environmental Impacts Act (S. 3732, 118th Congress, introduced 1 February 2024) would direct an EPA study, a NIST stakeholder consortium, and a voluntary reporting system — but it remains a measurement-and-study bill, not yet enacted, and imposes no caps (Congress.gov, S. 3732).","citation":"Artificial Intelligence Environmental Impacts Act of 2024, S. 3732, 118th Cong. (introduced Feb. 1, 2024, Sen. Markey), Congress.gov.","url":"https://www.congress.gov/bill/118th-congress/senate-bill/3732","sourceType":"primary_official","supportingQuote":"The bill was introduced on February 1, 2024 ... it did not receive a vote ... EPA Study ... NIST Consortium ... Voluntary Reporting System ... Neither the bill nor its House companion bill made it to the floor for a vote before the end of the 118th Congress"},{"kind":"concept","slug":"frontier-tier","sentence":"The earliest attestations are neutral: a March 2018 China Daily report quotes Minister Wan Gang on 'frontier AI-related science issues', and Scopus records a first academic use in 2019 (etymology traced in Nottingham's *Making Science Public*, 'Frontier AI: Tracing the origin of a concept', 2023).","citation":"Brigitte Nerlich et al., \"Frontier AI: Tracing the origin of a concept,\" Making Science Public blog, University of Nottingham, 20 October 2023.","url":"https://blogs.nottingham.ac.uk/makingsciencepublic/2023/10/20/frontier-ai-tracing-the-origin-of-a-concept/","sourceType":"institutional/academic blog (University of Nottingham)","supportingQuote":"Wan [Gang, minister of science and technology] said China will strengthen its AI research and train a new generation of experts to tackle key and frontier AI-related science issues. ... I got only nine hits with the first article published in 2019 by Chinese authors on AI, medicine and healthcare."},{"kind":"concept","slug":"frontier-tier","sentence":"The UK then adopted it officially at the Bletchley Park AI Safety Summit (1–2 November 2023), defining frontier AI as 'highly capable general-purpose AI models' matching or exceeding today's most advanced systems (UK Government, Bletchley Declaration, 2023).","citation":"UK Government, \"The Bletchley Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023\" (GOV.UK, 2023)","url":"https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration/the-bletchley-declaration-by-countries-attending-the-ai-safety-summit-1-2-november-2023","sourceType":"primary_official","supportingQuote":"Particular safety risks arise at the 'frontier' of AI, understood as being those highly capable general-purpose AI models, including foundation models, that could perform a wide variety of tasks ... which match or exceed the capabilities present in today's most advanced models"},{"kind":"concept","slug":"frontier-tier","sentence":"Here membership is not set by a compute number but by a four-part protocol — defined capability thresholds, a commitment to evaluate for them, pre-specified safeguards that engage *if* a threshold is reached, and a pause commitment if those safeguards cannot be implemented (the 'if-then commitment' framing follows Karnofsky 2024, Carnegie Endowment; the four-part protocol is set out in the developer frameworks themselves).","citation":"Holden Karnofsky, \"If-Then Commitments for AI Risk Reduction,\" Carnegie Endowment for International Peace, September 2024","url":"https://carnegieendowment.org/research/2024/09/if-then-commitments-for-ai-risk-reduction","sourceType":"institutional","supportingQuote":"These are commitments of the form: If an AI model has capability X, risk mitigations Y must be in place. And, if needed, we will delay AI deployment and/or development to ensure the mitigations can be present in time."},{"kind":"concept","slug":"frontier-tier","sentence":"Industry institutionalised the term on 26 July 2023 when Anthropic, Google, Microsoft and OpenAI launched the Frontier Model Forum (Microsoft, 'Anthropic, Google, Microsoft, OpenAI launch Frontier Model Forum', 26 July 2023).","citation":"Microsoft On the Issues, \"Anthropic, Google, Microsoft and OpenAI launch Frontier Model Forum,\" 26 July 2023","url":"https://blogs.microsoft.com/on-the-issues/2023/07/26/anthropic-google-microsoft-openai-launch-frontier-model-forum/","sourceType":"primary_official","supportingQuote":"Microsoft, Anthropic, Google, and OpenAI launch Frontier Model Forum"},{"kind":"benchmark","slug":"frontiermath","sentence":"First, Epoch retained a held-out set the funder had not seen, enabling independent re-evaluation; lead mathematician Elliot Glazer noted Epoch \"can't vouch for\" the vendor figure \"until our independent evaluation is complete\" (TechCrunch 2025-01-19) — the subsequent Epoch numbers (17%/10%) came in below the 25.2% headline.","citation":"Kyle Wiggers, 'AI benchmarking organization criticized for waiting to disclose funding from OpenAI', TechCrunch, 19 January 2025","url":"https://techcrunch.com/2025/01/19/ai-benchmarking-organization-criticized-for-waiting-to-disclose-funding-from-openai/","sourceType":"reputable_secondary","supportingQuote":"we can't vouch for them until our independent evaluation is complete ... Epoch AI also has a 'separate holdout set' that serves as an additional safeguard for independent verification"},{"kind":"benchmark","slug":"frontiermath","sentence":"When the original FrontierMath set was unveiled, Terence Tao had described its problems as \"extremely challenging\" and predicted they would \"resist AIs for several years at least\" (VentureBeat, Nov 8 2024).","citation":"Michael Nuñez, 'AI's math problem: FrontierMath benchmark shows how far technology still has to go', VentureBeat, 8 November 2024","url":"https://venturebeat.com/ai/ais-math-problem-frontiermath-benchmark-shows-how-far-technology-still-has-to-go","sourceType":"reputable_secondary","supportingQuote":"These are extremely challenging ... [the benchmark would] resist AIs for several years at least"},{"kind":"benchmark","slug":"frontiermath","sentence":"Epoch AI revealed OpenAI's funding only on 20 December 2024, alongside OpenAI's 25.2% o3 result, and many problem contributors were not told beforehand (TechCrunch 2025-01-19).","citation":"Maxwell Zeff, \"AI benchmarking organization criticized for waiting to disclose funding from OpenAI,\" TechCrunch, 19 January 2025.","url":"https://techcrunch.com/2025/01/19/ai-benchmarking-organization-criticized-for-waiting-to-disclose-funding-from-openai/","sourceType":"reputable_secondary","supportingQuote":"Epoch AI, a nonprofit primarily funded by Open Philanthropy, a research and grantmaking foundation, revealed on December 20 that OpenAI had supported the creation of FrontierMath. FrontierMath, a test with expert-level problems designed to measure an AI's mathematical skills, was one of the benchmar"},{"kind":"benchmark","slug":"gpqa-diamond","sentence":"The benchmark's own creator has since cautioned that when a model scores 85%, it is ambiguous whether it is reasoning through novel problems \"or has it seen enough similar problems in training that it's doing something closer to pattern-matched retrieval\" (Rein, as reported by MindStudio 2025).","citation":"MindStudio, \"GPQA: The Graduate-Level Benchmark Every Major AI Lab Uses — and Why Its Creator Says It Has Limits,\" 2025 (discussing David Rein's cautions)","url":"https://www.mindstudio.ai/blog/gpqa-benchmark-graduate-level-google-proof-qa-creator-limits","sourceType":"reputable_secondary","supportingQuote":"Is the model reasoning through novel scientific problems, or has it seen enough similar problems in training that it's doing something closer to pattern-matched retrieval?"},{"kind":"benchmark","slug":"gpqa-diamond","sentence":"The benchmark's creator concurs, noting models in \"the 80s and 90s\" caused it to \"stop discriminating between good and great,\" and describing GPQA as \"a stepping stone, not a destination\" (Rein, MindStudio 2025).","citation":"MindStudio, \"GPQA: The Graduate-Level Benchmark Every Major AI Lab Uses — and Why Its Creator Says It Has Limits,\" 2025 (discussing David Rein's cautions)","url":"https://www.mindstudio.ai/blog/gpqa-benchmark-graduate-level-google-proof-qa-creator-limits","sourceType":"reputable_secondary","supportingQuote":"Then models started scoring in the 80s and 90s on GPQA Diamond ... it stopped discriminating between good and great. ... GPQA was a stepping stone, not a destination."},{"kind":"benchmark","slug":"gpqa-diamond","sentence":"The label quality itself holds up — independent review near saturation found ~90-95% of items valid, with only roughly 2-3 of 198 seriously ambiguous (review summarized by IntuitionLabs 2025) — so the residual frontier gap is mostly genuine difficulty rather than flawed keys.","citation":"IntuitionLabs, \"GPQA-Diamond Benchmark: Scores, Leaderboard & How AI Models Compare\" (2025) — reporting Greg Burnham/Epoch AI's manual label audit","url":"https://intuitionlabs.ai/articles/gpqa-diamond-ai-benchmark","sourceType":"reputable_secondary","supportingQuote":"only a small number had major issues or ambiguous answers (roughly 2–3 out of 198) ... about 90–95% of the benchmark appears well-posed ... the remaining ~10% error gap for top AIs likely reflects genuine challenge, not flawed labeling"},{"kind":"benchmark","slug":"gpqa-diamond","sentence":"By 2025-2026 frontier systems cluster in the low-to-mid 90s — e.g., Gemini 3.1 Pro Preview at 94.1% and GPT-5.5 at ~93% on the Artificial Analysis leaderboard (2026).","citation":"Artificial Analysis, \"GPQA Diamond Benchmark Leaderboard\" (2026)","url":"https://artificialanalysis.ai/evaluations/gpqa-diamond","sourceType":"institutional","supportingQuote":"Gemini 3.1 Pro Preview scores the highest on GPQA with a score of 94.1%, followed by GPT-5.5 (xhigh) with a score of 93.5%, and GPT-5.5 (high) with a score of 93.2%"},{"kind":"concept","slug":"hardware-enabled-governance","sentence":"(c) On-chip usage and licensing locks can throttle or gate a chip absent a valid authorization (a 'feature lock' or offline-licensing scheme, per CNAS — Aarne, Fist & Withers 2024, and RAND — Kulp et al.","citation":"Onni Aarne, Tim Fist & Caleb Withers, \"Secure, Governable Chips: Using On-Chip Mechanisms to Manage National Security Risks from AI & Advanced Computing,\" Center for a New American Security (CNAS), January 2024; and Gabriel Kulp et al., \"Hardware-Enabled Governance Mechanisms,\" RAND WR-A3056-1 (2024)","url":"https://www.cnas.org/publications/reports/secure-governable-chips","sourceType":"institutional","supportingQuote":"The core of this proposal is a hardened security module...that can ensure that the chip has valid, up-to-date firmware and software and, where applicable, an up-to-date operating license. If these conditions are not met, it would block the chip from operating."},{"kind":"concept","slug":"hardware-enabled-governance","sentence":"In December 2025 Nvidia disclosed (Reuters, Dec 9 2025) a software-based location-verification option using existing confidential-computing features and server-latency timing — a chip-assisted, read-only mechanism, explicitly not a tamper-resistant on-chip regime and with no kill switch.","citation":"Alexandra Alper & Karen Freifeld, \"Exclusive: Nvidia builds location verification tech that could help fight chip smuggling,\" Reuters, Dec. 9-10, 2025.","url":"https://finance.yahoo.com/news/exclusive-nvidia-builds-location-verification-044501961.html","sourceType":"news","supportingQuote":"The feature, which Nvidia has demonstrated privately in recent ‌months but has not yet released, would be a software option that customers could install. It would tap into what are known as the confidential computing capabilities of its ‌graphics processing units (GPUs). ... would use the time delay"},{"kind":"concept","slug":"hardware-enabled-governance","sentence":"Physical unclonable functions (PUFs) derive a key from uncontrollable manufacturing variation rather than storing it in memory; Herder, Yu, Koushanfar & Devadas (2014, Proc.","citation":"C. Herder, M.-D. Yu, F. Koushanfar, and S. Devadas, \"Physical Unclonable Functions and Applications: A Tutorial,\" Proceedings of the IEEE, vol. 102, no. 8, pp. 1126-1141, 2014, doi:10.1109/JPROC.2014.2320516.","url":"https://www.acsu.buffalo.edu/~mblanton/cse708/pufs-tutorial.pdf","sourceType":"peer-reviewed","supportingQuote":"By Charles Herder, Meng-Day (Mandel) Yu, Farinaz Koushanfar, and Srinivas Devadas ... Proceedings of the IEEE | Vol. 102, No. 8, August 2014 ... instead of storing secrets in digital memory, PUFs derive a secret from the physical characteristics of the integrated circuit (IC). For example, this pape"},{"kind":"topic","slug":"healthcare","sentence":"In the EU, the MDCG/AI Board issued joint interplay guidance MDCG 2025-6 on 19 June 2025, and on 19 November 2025 the Commission's \"Digital Omnibus\" proposed postponing high-risk obligations for Annex I products — including medical devices — to 2 August 2028 (European Commission, Digital Omnibus proposal, 19 Nov 2025).","citation":"European Commission, \"Digital Omnibus\" package / Digital Omnibus on AI, proposed 19 November 2025 (postponing high-risk AI Act obligations for Annex I products, including medical devices, to 2 August 2028).","url":"https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/","sourceType":"reputable_secondary","supportingQuote":"for AI embedded in regulated products under Annex I — such as medical devices, machinery, and vehicles ... to 2 August 2028 ... the European Commission to table the Digital Omnibus on AI on 19 November 2025"},{"kind":"topic","slug":"healthcare","sentence":"The MDCG/AI Board joint guidance MDCG 2025-6 confirms these AI Act duties (data governance, logging, human oversight, transparency) are to be discharged within the existing MDR/IVDR conformity-assessment procedure rather than through a parallel certification (MDCG 2025-6).","citation":"Medical Device Coordination Group & European Artificial Intelligence Board, \"AIB 2025-1 / MDCG 2025-6: Interplay between the MDR/IVDR and the AI Act\" (19 June 2025).","url":"https://health.ec.europa.eu/latest-updates/mdcg-2025-6-faq-interplay-between-medical-devices-regulation-vitro-diagnostic-medical-devices-2025-06-19_en","sourceType":"official-government","supportingQuote":"As the quality management system obligations under the AIA are specifically targeted to the AI system, additional requirements such as data and data governance, record-keeping, transparency, human oversight must be integrated, as appropriate ... manufacturers of AI systems may include the elements o"},{"kind":"benchmark","slug":"humanitys-last-exam","sentence":"The first large jump came not from a larger base model but from tool use: OpenAI's agentic Deep Research, browsing autonomously for minutes per question, reached 26.6% in February 2025 — roughly a threefold gain over the best non-tool score at the time (OpenAI 2025-02-02).","citation":"OpenAI, 'Introducing deep research', OpenAI blog, 2 February 2025","url":"https://openai.com/index/introducing-deep-research/","sourceType":"primary_official","supportingQuote":"On Humanity's Last Exam ... the model powering deep research scores a new high at 26.6% accuracy ... Powered by OpenAI's frontier o3 model, the AI agent can synthesize a wide range of information and complete multistep research within five-to-30 minutes"},{"kind":"concept","slug":"inference-time-compute","sentence":"First, evaluations must declare the inference-compute budget, since a system safe at K=1 can be dangerous at K=100; the Seoul Declaration and Frontier AI Safety Commitments (SEOUL-2024, May 2024) gesture toward 'pre-deployment evaluation under realistic conditions' but no regulator has formalised inference-compute-aware thresholds.","citation":"Frontier AI Safety Commitments, AI Seoul Summit 2024 (21 May 2024). UK Department for Science, Innovation & Technology / Republic of Korea. GOV.UK.","url":"https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024","sourceType":"primary/official government publication","supportingQuote":"Assess the risks posed by their frontier models or systems across the AI lifecycle, including before deploying that model or system ... Set out thresholds at which severe risks posed by a model or system, unless adequately mitigated, would be deemed intolerable."},{"kind":"concept","slug":"jailbreak-resistance","sentence":"(3) In-context attacks exploit long context: many-shot jailbreaking prepends hundreds of faux dialogue turns, with attack success rising as a power law in shot count (Anil et al. 2024, NeurIPS).","citation":"Anil, C., Durmus, E., Panickssery, N., Sharma, M., et al. (2024). Many-shot Jailbreaking. Advances in Neural Information Processing Systems 37 (NeurIPS 2024).","url":"https://papers.nips.cc/paper_files/paper/2024/hash/ea456e232efb72d261715e33ce25f208-Abstract-Conference.html","sourceType":"peer_reviewed","supportingQuote":"We investigate a family of simple long-context attacks on large language models: prompting with hundreds of demonstrations of undesirable behavior... We find that in diverse, realistic circumstances, the effectiveness of this attack follows a power law, up to hundreds of shots."},{"kind":"concept","slug":"mesa-optimization","sentence":"GMG was first demonstrated empirically by Langosco et al.","citation":"Lauro Langosco, Jack Koch, Lee Sharkey, Jacob Pfau & David Krueger, \"Goal Misgeneralization in Deep Reinforcement Learning,\" Proceedings of the 39th International Conference on Machine Learning (ICML), PMLR 162 (2022).","url":"https://proceedings.mlr.press/v162/langosco22a.html","sourceType":"peer_reviewed","supportingQuote":"We provide the first explicit empirical demonstrations of goal misgeneralization and present a partial characterization of its causes."},{"kind":"concept","slug":"mesa-optimization","sentence":"Discussion of a sub-process that internally optimises a different objective than the one selecting it circulated in the alignment community as \"optimization daemons\" and \"inner optimizers,\" associated with an Arbital treatment around 2016, and with MIRI work by Jessica Taylor in February 2017 on whether such \"daemons\" arise for idealised agents (AI Alignment Forum, \"Mesa-optimization\" entry, summarising Taylor 2017).","citation":"AI Alignment Forum / LessWrong, \"Mesa-Optimization\" wiki/tag entry (summarizing the 'optimization daemons' Arbital treatment c.2016 and Jessica Taylor's Feb 2017 MIRI posts, incl. 'Are daemons a problem for ideal agents?', 2017-02-11).","url":"https://www.alignmentforum.org/w/mesa-optimization","sourceType":"community_reference","supportingQuote":"Previously work under this concept was called Inner Optimizer or Optimization Daemons. The optimization daemons article on Arbital was published probably in 2016."},{"kind":"concept","slug":"mesa-optimization","sentence":"(2022) define GMG as a model whose capabilities generalise out-of-distribution while its goal does not, and explicitly position mesa-optimisation as a strict special case: \"Hubinger et al. introduce mesa optimization, a type of goal misgeneralization where a learned model implements a search algorithm with an explicitly represented objective.","citation":"Shah, R., Varma, V., Kumar, R., Phuong, M., Krakovna, V., Uesato, J., & Kenton, Z. (2022). Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals. arXiv:2210.01790.","url":"https://arxiv.org/abs/2210.01790","sourceType":"peer_reviewed","supportingQuote":"Hubinger et al. introduce mesa optimization, a type of goal misgeneralization where a learned model implements a search algorithm with an explicitly represented objective."},{"kind":"benchmark","slug":"mmlu-pro","sentence":"Public aggregator leaderboards place 2025-era frontier systems near 90% — for example Gemini 3 Pro Preview at 89.8% and Claude Opus 4.5 (reasoning mode) at 89.5% (Artificial Analysis, accessed June 2026).","citation":"Artificial Analysis, \"MMLU-Pro Benchmark Leaderboard\" (accessed 2026)","url":"https://artificialanalysis.ai/evaluations/mmlu-pro","sourceType":"institutional","supportingQuote":"Gemini 3 Pro Preview (high) scores the highest on MMLU-Pro with a score of 89.8% ... and Claude Opus 4.5 (Reasoning) with a score of 89.5%"},{"kind":"topic","slug":"national-security-carveouts","sentence":"Executive Order 14110 was rescinded on 20 January 2025 and superseded by Executive Order 14179, \"Removing Barriers to American Leadership in Artificial Intelligence\" (signed 23 Jan. 2025; 90 Fed.","citation":"Executive Order 14179 of January 23, 2025, \"Removing Barriers to American Leadership in Artificial Intelligence,\" 90 Fed. Reg. 8741 (Jan. 31, 2025)","url":"https://www.govinfo.gov/content/pkg/FR-2025-01-31/pdf/2025-02172.pdf","sourceType":"primary_official","supportingQuote":"Executive Order 14179 of January 23, 2025 — Removing Barriers to American Leadership in Artificial Intelligence ... all policies, directives, regulations, orders, and other actions taken pursuant to the revoked Executive Order 14110 of October 30, 2023"},{"kind":"topic","slug":"national-security-carveouts","sentence":"The United States parallel track is itself filled out by the DoD Responsible AI Strategy and Implementation Pathway, whose tenets require that \"DoD personnel will exercise appropriate levels of judgment and care, while remaining responsible for the development, deployment, and use of AI capabilities\" (DoD RAI S&IP 2022) — operationalising security AI rather than exempting it.","citation":"U.S. Department of Defense, \"Responsible Artificial Intelligence Strategy and Implementation Pathway\" (June 2022)","url":"https://media.defense.gov/2024/Oct/26/2003571790/-1/-1/0/2024-06-RAI-STRATEGY-IMPLEMENTATION-PATHWAY.PDF","sourceType":"primary_official","supportingQuote":"DoD personnel will exercise appropriate levels of judgment and care, while remaining responsible for the development, deployment, and use of AI capabilities"},{"kind":"topic","slug":"open-weight-release","sentence":"Second, the *firm-level cleavage* exposed by California SB-1047: Meta opposed the bill, urging a lighter approach (Meta letters to California lawmakers, 2024), whereas Anthropic moved from non-support to \"measured support\" after amendments (D.","citation":"Wikipedia (secondary aggregation of primary letters), \"Safe and Secure Innovation for Frontier Artificial Intelligence Models Act\"; Anthropic (Dario Amodei) letter to Gov. Newsom, 21 Aug. 2024; Meta (Rob Sherman) June 2024 letter to Sen. Wiener as reported by CalMatters.","url":"https://en.wikipedia.org/wiki/Safe_and_Secure_Innovation_for_Frontier_Artificial_Intelligence_Models_Act","sourceType":"reputable_secondary","supportingQuote":"the new SB 1047 is substantially improved, to the point where we believe its benefits likely outweigh its costs"},{"kind":"topic","slug":"open-weight-release","sentence":"(4) Preservation-and-refusal contracting: California's AI Transparency Act reaches weight distribution indirectly — a licensor must contractually require licensees to keep a disclosure capability and revoke within 96 hours otherwise (Cal.","citation":"California Legislature, SB-942 California AI Transparency Act, § 22757.3(c) (2024) — covered provider must revoke a licensee's license within 96 hours upon discovering the licensee modified the GenAI system so it can no longer include the required disclosure; corroborated by Orrick, \"Navigating the California AI Transparency Act: New Contract Requirements\" (2025).","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240SB942","sourceType":"primary_official","supportingQuote":"If a covered provider knows that a third-party licensee modified a licensed GenAI system such that it is no longer capable of including a disclosure required by subdivision (b) in content the system creates or alters, the covered provider shall revoke the license within 96 hours of discovering the l"},{"kind":"topic","slug":"open-weight-release","sentence":"In California, the vetoed SB-1047 (vetoed 29 Sep 2024) was succeeded by SB-53, the Transparency in Frontier Artificial Intelligence Act, signed 29 Sep 2025 and effective 1 Jan 2026; it imposes transparency-framework publication and critical-safety-incident reporting on developers training above ~10^26 FLOP, applying uniformly to open- and closed-weight models rather than carving open release out (Brookings, *What is California's AI safety law?*, 2025).","citation":"Brookings Institution, \"What is California's AI safety law?\" (2025).","url":"https://www.brookings.edu/articles/what-is-californias-ai-safety-law/","sourceType":"institutional","supportingQuote":"On September 29, 2025, as Congress remained deadlocked on comprehensive AI legislation, California Governor Gavin Newsom signed Senate Bill 53 (SB 53) into law. ... foundation models trained above 10^26 floating-point operations (FLOPs)."},{"kind":"concept","slug":"policy-instrument","sentence":"Hood (1983, ch. 1-2) groups instruments into the NATO scheme — Nodality (information), Authority (legal command), Treasure (fiscal transfer), and Organisation (direct provision) — so a registry, a binding rule, a subsidy, and a state laboratory are distinct techniques even when aimed at one problem, distinguishing instruments from objectives and from policy styles.","citation":"Christopher Hood, The Tools of Government (Macmillan/Chatham House, 1983), ch. 1-2 — the NATO scheme: Nodality, Authority, Treasure, Organisation.","url":"https://www.ippapublicpolicy.org/file/paper/5b28e4eeaa3d0.pdf","sourceType":"academic_paper","supportingQuote":"Christopher Hood (1983; 1986) proposed his well-known NATO typology based on the \"resources\" governments have at their disposal: nodality (being at the centre of an information network), authority, treasure and organization."},{"kind":"concept","slug":"prompt-injection","sentence":"In September 2022 Riley Goodside publicly demonstrated that GPT-3 could be made to disregard its instructions via crafted user input, and Simon Willison coined the term 'prompt injection' that same month, explicitly analogising it to SQL injection (Willison 2022, 'Prompt injection attacks against GPT-3').","citation":"Simon Willison, \"Prompt injection attacks against GPT-3,\" simonwillison.net, 12 September 2022.","url":"https://simonwillison.net/2022/Sep/12/prompt-injection/","sourceType":"primary_official","supportingQuote":"I propose that the obvious name for this should be prompt injection. ... The obvious parallel here is SQL injection. ... [references Riley's September 12, 2022 work showing GPT-3 could be tricked into ignoring its instructions]"},{"kind":"concept","slug":"prompt-injection","sentence":"Standardisation of measurement arrived in 2024 with formal frameworks and agentic benchmarks—Liu et al.","citation":"Yupei Liu, Yuqi Jia, Runpeng Geng, Jinyuan Jia, Neil Zhenqiang Gong, \"Formalizing and Benchmarking Prompt Injection Attacks and Defenses,\" USENIX Security Symposium 2024.","url":"https://arxiv.org/abs/2310.12815","sourceType":"peer_reviewed","supportingQuote":"We propose a framework to formalize prompt injection attacks. Existing attacks are special cases in our framework. ... systematic evaluation on 5 prompt injection attacks and 10 defenses with 10 LLMs and 7 tasks ... a common benchmark for quantitatively evaluating future prompt injection attacks and"},{"kind":"topic","slug":"redress","sentence":"Outside the EU, Brazil's PL 2338/2023 — with its rights to explanation, contestation, and human review — passed the Federal Senate on 10 December 2024 and remained under review in the Chamber of Deputies through 2025 (Library of Congress 2025).","citation":"Library of Congress, Global Legal Monitor, 'Brazil: Senate Advances Discussions on Bill to Regulate AI Use' (May 23, 2025)","url":"https://www.loc.gov/item/global-legal-monitor/2025-05-23/brazil-senate-advances-discussions-on-bill-to-regulate-ai-use/","sourceType":"primary_official","supportingQuote":"On 10 December 2024, the Brazilian Senate approved Bill No. 2338/2023... now moves to the Chamber of Deputies for a vote... [PL 2338/2023 grants:] Right to Explanation... Right to Contest... Right to Human Review — the right to human determination and participation in decisions of artificial intelli"},{"kind":"concept","slug":"sandbagging","sentence":"Prompting: a system prompt instructs frontier models (GPT-4, Claude 3 Opus) to answer dangerous biology, chemistry, and cyber questions incorrectly while staying accurate on benign topics (van der Weij et al. 2024).","citation":"van der Weij, T., Hofstätter, F., Jaffe, O., Brown, S.F., & Ward, F.R. (2024). AI Sandbagging: Language Models can Strategically Underperform on Evaluations. arXiv:2406.07358.","url":"https://arxiv.org/abs/2406.07358","sourceType":"peer_reviewed","supportingQuote":"We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations."},{"kind":"concept","slug":"sandbagging","sentence":"In developer-driven sandbagging, a vendor configures or trains a system to score low on dangerous-capability tests so the model presents as safer and dodges capability-tier obligations; in AI-driven sandbagging, the model itself withholds capability, presupposing situational awareness of evaluation conditions (van der Weij et al. 2024).","citation":"van der Weij, T., Hofstätter, F., Jaffe, O., Brown, S.F., & Ward, F.R. (2024). AI Sandbagging: Language Models can Strategically Underperform on Evaluations. arXiv:2406.07358.","url":"https://arxiv.org/abs/2406.07358","sourceType":"peer_reviewed","supportingQuote":"If the strategic nature originates from the developer, we call it developer sandbagging. Otherwise, if the strategic nature originates from the AI system, we call it AI system sandbagging."},{"kind":"topic","slug":"sovereign-ai","sentence":"On 2025-05-13 the Bureau of Industry and Security rescinded the Biden-era \"AI Diffusion Rule\" days before its effective date, abandoning its three-tier country framework as \"overly bureaucratic\" (BIS / Department of Commerce announcement, 2025).","citation":"Bureau of Industry and Security, U.S. Department of Commerce, \"Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule, Strengthens Chip-Related Export Controls\" (press release, May 13, 2025).","url":"https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens","sourceType":"primary_official","supportingQuote":"On May 13, 2025, the Department of Commerce announced it rescinded the Biden Administration's AI Diffusion Rule, which was set to take effect on May 15, 2025. \"These new requirements would have stifled American innovation and saddled companies with burdensome new regulatory requirements.\" ... the ru"},{"kind":"topic","slug":"synthetic-content-provenance","sentence":"The United States is converging through state law rather than federal mandate: California's AI Transparency Act (SB 942) was delayed by AB 853 from 1 January 2026 to 2 August 2026 to align with the EU date, and AB 853 layers in staged duties — large-platform provenance DETECTION from 1 January 2027 and capture-device latent-disclosure options from 1 January 2028 (California SB 942; AB 853, signed 13 Oct 2025).","citation":"California Legislature, AB-853 California AI Transparency Act (Chapter 674, Statutes of 2025), signed 13 Oct 2025 — delays SB-942 operative date to 2 Aug 2026 and adds large-platform provenance-detection duties (from 1 Jan 2027) and capture-device provenance features (from 1 Jan 2028).","url":"https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB853","sourceType":"primary_official","supportingQuote":"This chapter shall become operative on August 2, 2026. ... [22757.3.1] This section shall become operative on January 1, 2027. ... [22757.3.2] This section shall become operative on January 1, 2028."},{"kind":"topic","slug":"synthetic-content-provenance","sentence":"On the technical layer, C2PA's shift to durable Content Credentials (soft-binding watermarks, C2PA 2.x, 2025) signals the underlying standard is still maturing as the legal deadlines arrive.","citation":"Coalition for Content Provenance and Authenticity (C2PA), \"C2PA and Content Credentials Explainer, v2.2\" (2025-04-22), describing Durable Content Credentials combining hard binding with soft-binding watermarks/fingerprints.","url":"https://spec.c2pa.org/specifications/specifications/2.2/explainer/_attachments/Explainer.pdf","sourceType":"primary_official","supportingQuote":"That is why the C2PA specification includes the concept of durable Content Credentials, which combines a hard binding (aka cryptographic hashing) with a soft binding (e.g., watermarking and fingerprinting). ... Soft bindings can either be implemented via invisible watermarking or fingerprint lookup."},{"kind":"concept","slug":"systemic-risk","sentence":"Zwetsloot and Dafoe (2019) partition AI risk into misuse, accident, and structural — the last being harm that arises from how a technology reshapes incentives, power balances, and competitive dynamics even when no actor misuses it and nothing malfunctions (\"Thinking About Risks From AI: Accidents, Misuse and Structure\", GovAI).","citation":"Remco Zwetsloot and Allan Dafoe, \"Thinking About Risks From AI: Accidents, Misuse and Structure,\" Centre for the Governance of AI (GovAI), 2019.","url":"https://www.governance.ai/research-paper/thinking-about-risks-from-ai-accidents-misuse-and-structure","sourceType":"institutional/research (Centre for the Governance of AI)","supportingQuote":"analysts have done a good job outlining how AI might cause harm through either intentional misuse or accidental system failures. But other kinds of risk...do not fit neatly into this misuse-accident dichotomy ... a structural perspective on risk, one that focuses explicitly on how AI technologies wi"},{"kind":"topic","slug":"training-data","sentence":"Anthropic, Judge Alsup held that training on *lawfully acquired* books was \"quintessentially transformative\" fair use, but that ingesting *pirated* copies was not — a split the parties resolved with a US$1.5 billion settlement preliminarily approved in September 2025 (Bartz v.","citation":"Bartz et al. v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.), Order on Fair Use (Alsup, J.), 23 June 2025","url":"https://docs.justia.com/cases/federal/district-courts/california/candce/3:2024cv05417/434709/231","sourceType":"primary_official","supportingQuote":"the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative... The downloaded pirated copies used to build a central library were not justified by a fair use. Every factor points against fair use."},{"kind":"instrument","slug":"us-eo-14110","sentence":"Relative to the Council of Europe Framework Convention on AI (CETS No. 225, 2024), EO 14110 was narrower, focused on a national-security and standards agenda rather than human-rights treaty obligations.","citation":"Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), adopted 17 May 2024, opened for signature 5 Sept. 2024","url":"https://www.coe.int/en/web/portal/-/council-of-europe-opens-first-ever-global-treaty-on-ai-for-signature","sourceType":"primary_official","supportingQuote":"On 5 September 2024, the first-ever, international, legally-binding instrument on Artificial Intelligence (AI) was opened for signature by the Council of Europe (CoE). ... provides a common baseline to ensure that activities within the lifecycle of AI systems are fully consistent with human rights, "},{"kind":"instrument","slug":"us-eo-14110","sentence":"The trajectory then inverted: President Trump revoked EO 14110 on January 20, 2025, and issued EO 14179, \"Removing Barriers to American Leadership in Artificial Intelligence,\" 90 Fed.","citation":"Executive Order 14179, \"Removing Barriers to American Leadership in Artificial Intelligence,\" 90 Fed. Reg. 8741 (Jan. 31, 2025) [signed Jan. 23, 2025]","url":"https://www.federalregister.gov/documents/2025/01/31/2025-02172/removing-barriers-to-american-leadership-in-artificial-intelligence","sourceType":"primary_official","supportingQuote":"This order revokes certain existing AI policies and directives that act as barriers to American AI innovation ... identify any actions taken pursuant to Executive Order 14110 that are or may be inconsistent with ... the policy set forth in section 2 of this order."},{"kind":"instrument","slug":"us-eo-14110","sentence":"A second fault line is institutional legitimacy: commentators questioned grounding economy-wide AI reporting in the Defense Production Act, a Korean-War-era statute, rather than tailored legislation (CRS Report R47843, 2023).","citation":"Congressional Research Service, \"Highlights of the 2023 Executive Order on Artificial Intelligence for Congress,\" CRS Report R47843 (2023)","url":"https://www.congress.gov/crs-product/R47843","sourceType":"primary_official","supportingQuote":"E.O. 14110 invokes the Defense Production Act (DPA), which gives the President sweeping authorities to compel or incentivize industry in the interest of national security."}],"needsSourcingQueue":[{"kind":"benchmark","slug":"aime-2024","section":"Saturation & score trajectory","sentence":"The same release reported OpenAI o1 at 74.4% pass@1, rising to 83.3% with majority vote over 64 samples and ~93% with learned re-ranking over 1,000 samples — a single-day jump of roughly 60 points over GPT-4o on the same items.","rationale":"Specific vendor-reported OpenAI o1/GPT-4o scores (74.4%, 83.3%, 93%) are third-party model-performance figures, not properties of the AIME benchmark itself, and no listed literature item covers them."},{"kind":"benchmark","slug":"aime-2024","section":"Saturation & score trajectory","sentence":"GPT-4o, a strong non-reasoning model, solved on average about 12% (reported as 13.4% pass@1) of the 2024 problems (OpenAI, \"Learning to Reason with LLMs,\" 2024-09-12).","rationale":"Specific GPT-4o pass@1 figure (13.4%) from OpenAI's 'Learning to Reason with LLMs' — a third-party vendor result, not a fact about the AIME benchmark itself, and unsupported by any listed item."},{"kind":"concept","slug":"alignment","section":"History of the idea and term","sentence":"The framing as a distinct technical problem for advanced AI is generally traced to Yudkowsky's articulation of the risk that a sufficiently capable optimiser pursues its given objective rather than its designers' intent (Yudkowsky 2008, the article's primary citation).","rationale":"[residual quote-audit demotion] The quote ('When we talk about AIs we are really talking about minds-in-general, or optimization processes in general') is topical-only. It does not establish the specific load-bearing assertion — tha"},{"kind":"topic","slug":"development-rights-framing","section":"Key fault lines","sentence":"First, even at the UN the framing split: 2024 produced two consensus resolutions — the US-led A/RES/78/265 emphasising safe, trustworthy systems for sustainable development, and a China-led capacity-building resolution co-sponsored by a Global-South-plus coalition foregrounding investment, technology transfer and bridging the divide.","rationale":"[residual quote-audit demotion] The quote establishes only the US-led A/RES/78/265 (title + adopted 21 March 2024). It does NOT establish the sentence's load-bearing facts about the second, China-led resolution: no date (1 July 2024"},{"kind":"instrument","slug":"eu-ai-act","section":"Implementation and trajectory: staggered timeline and the Digital Omnibus","sentence":"A provisional inter-institutional agreement reached on 7 May 2026 deferred the high-risk deadlines — Annex III obligations from 2 August 2026 to 2 December 2027 (a 16-month slip tied to standards availability), and Annex I obligations from 2 August 2027 to 2 August 2028 — while adding targeted measures such as a ban on 'nudifier' applications (European Parliament press release, 23 March 2026).","rationale":"[2026-07-02 cross-audit] Demoted from externally_resolved: the recorded source does not establish the claim (predates the event / quote not at URL / supports only a tangential limb). Needs a valid external reference."},{"kind":"instrument","slug":"eu-ai-act","section":"Implementation and trajectory: staggered timeline and the Digital Omnibus","sentence":"Early enforcement has nonetheless begun under the in-force prohibition and GPAI tiers, with reported market-surveillance scrutiny of large platforms (Council of the EU press release, 7 May 2026).","rationale":"[residual quote-audit demotion] Sentence asserts 'early enforcement has begun' with 'market-surveillance scrutiny of large platforms.' Quote only describes allocation of supervisory authority (AI Office over GPAI systems, DSA Arts 3"},{"kind":"instrument","slug":"eu-ai-act","section":"Cross-jurisdiction position: the only binding horizontal regime","sentence":"China's regime is binding but vertical and rolled out piecemeal — the Interim Measures for the Management of Generative AI Services (effective 15 August 2023) target public-facing generative services with content-control and security-assessment duties, and scholarship reads them partly as a pro-growth signalling device rather than a comprehensive risk framework (Zhang, 'The Promise and Perils of China's Regulation of Artificial Intelligence,' Columbia J.","rationale":"[2026-07-02 cross-audit] Demoted from externally_resolved: the recorded source does not establish the claim (predates the event / quote not at URL / supports only a tangential limb). Needs a valid external reference."},{"kind":"topic","slug":"foundation-models","section":"Trajectory: what is changing","sentence":"The Commission's Digital Omnibus on AI, published 19 November 2025 and reaching provisional trilogue agreement on 7 May 2026, deferred the *high-risk* (Annex III) deadline from August 2026 to 2 December 2027 but left the GPAI obligations and their August 2025 start date intact.","rationale":"[quote-audit demotion] a source was found but its quote did not ESTABLISH the specific claim: The quote establishes only the 19 November 2025 publication date and a generic description of 'targeted simplification measures.' It does NOT establish the substantive load-bearing facts: the 7 May 2026 provisional trilo"},{"kind":"benchmark","slug":"frontiermath","section":"Contamination, access asymmetry, and gaming","sentence":"OpenAI had \"access to a large fraction of the problems and solutions,\" governed by a \"verbal agreement\" not to train on them, and Epoch's Tamay Besiroglu conceded the organisation \"made a mistake\" in not negotiating to disclose the relationship earlier (TechCrunch 2025-01-19).","rationale":"[2026-07-02 cross-audit] Demoted from externally_resolved: the recorded source does not establish the claim (predates the event / quote not at URL / supports only a tangential limb). Needs a valid external reference."},{"kind":"benchmark","slug":"frontiermath","section":"Saturation and score trajectory","sentence":"On 20 December 2024 OpenAI reported o3-preview at 25.2% — a >10x jump announced the same day the partnership behind the benchmark surfaced (OpenAI 2024; TechCrunch 2025-01-19).","rationale":"[2026-07-02 cross-audit] Demoted from externally_resolved: the recorded source does not establish the claim (predates the event / quote not at URL / supports only a tangential limb). Needs a valid external reference."},{"kind":"instrument","slug":"gdpr","section":"Cross-jurisdiction position","sentence":"GDPR is the global reference point against which most peer regimes are read, a dynamic Bradford (2020) theorised as the \"Brussels Effect\": the de facto export of EU standards through market access and compliance economies of scale (Anu Bradford, *The Brussels Effect*, OUP 2020).","rationale":"[residual quote-audit demotion] Quote is only the book's title/author/publisher line. It does not establish the substantive load-bearing assertion that Bradford theorised the 'Brussels Effect' as de facto export of EU standards thro"},{"kind":"instrument","slug":"gdpr","section":"Implementation and trajectory","sentence":"The single largest penalty remains the EUR 1.2bn imposed on Meta in 2023 for unlawful EU-US data transfers via SCCs, following an EDPB binding decision of 13 April 2023 (edpb.europa.eu); 2024 brought two further headline DPC penalties — EUR 310m against LinkedIn (dataprotection.ie, 24 Oct 2024) for an unlawful behavioural-advertising basis, and EUR 251m against Meta (dataprotection.ie, 17 Dec 2024) over the 2018 Facebook token-exposure breach.","rationale":"[residual quote-audit demotion] Multi-fact sentence leads on 'the single largest penalty remains the EUR 1.2bn imposed on Meta in 2023' (EDPB 13 Apr 2023) plus a EUR 251m Meta penalty. The quote establishes ONLY the LinkedIn EUR 310"},{"kind":"benchmark","slug":"gpqa-diamond","section":"Contamination, format sensitivity, and gaming","sentence":"But the creator stresses the protection is not permanent: \"any fixed benchmark eventually gets trained against, either explicitly through data contamination or implicitly through general capability improvements\" (Rein, MindStudio 2025) — the rationale for vetted/withheld variants of difficult benchmarks generally.","rationale":"[over-crediting audit] Quotes Rein via 'MindStudio 2025' making a general claim about benchmark protection ('any fixed benchmark eventually gets trained against') — a statement about benchmarks generally, sourced to an external 2025 report, not to GPQA's own primary source. Needs an external reference."},{"kind":"concept","slug":"in-context-learning","section":"Debates and Open Questions","sentence":"First, mechanism: the Bayesian, meta-learning, and gradient-descent readings make divergent predictions about robustness and failure, with no resolution (Xie et al. 2022; von Oswald et al. 2023).","rationale":"[residual quote-audit demotion] The sentence turns on THREE readings (Bayesian, meta-learning, gradient-descent) making 'divergent predictions about robustness and failure, with no resolution.' The quote is only the Xie 2022 title/a"},{"kind":"concept","slug":"model-card","section":"From Voluntary Norm to Binding Codification","sentence":"NIST AI RMF cites model cards as a transparency mechanism under GOVERN 1.4 (referencing Mitchell et al.), with measurement documentation mapping to the MEASURE 2.x subcategories; ISO/IEC 23894 endorses analogous documentation.","rationale":"[quote-audit demotion] a source was found but its quote did not ESTABLISH the specific claim: The quote is merely the bibliographic listing of Mitchell et al. It establishes that the work is cited but NOT the load-bearing structural claims: that NIST places model cards under GOVERN 1.4, that measurement docs map "},{"kind":"concept","slug":"policy-instrument","section":"How Instrument Choice Operates as Substance","sentence":"Howlett (2011, ch. 3-5) treats selection as constrained by information, capability, and political variables, so one goal yields different tools across jurisdictions.","rationale":"[residual quote-audit demotion] Quote is bibliographic metadata only (title/author/publisher). It establishes the book exists but says nothing about the substantive load-bearing claim that Howlett 'treats selection as constrained by"},{"kind":"concept","slug":"policy-instrument","section":"How Instrument Choice Operates as Substance","sentence":"Lascoumes & Le Galès (2007, pp. 4-5) characterise an instrument as 'a particular form of materialisation of state power' that generates its own logic once deployed.","rationale":"Attributes a specific quoted definition to Lascoumes & Le Galès (2007, pp. 4-5); that work is not in the listed literature, so it needs an external reference."},{"kind":"concept","slug":"policy-instrument","section":"Governance Relevance: Mapping the AI Instrument Mix","sentence":"(2011, ch. 1), as a response to the 'pacing problem': regulation lags capability, so jurisdictions sequence soft-law ahead of hard-law.","rationale":"[residual quote-audit demotion] Quote is title/series/editors metadata plus a dangling 'At the same time that' — it establishes only that a book titled 'The Pacing Problem' exists. It does NOT establish the load-bearing mechanism th"},{"kind":"concept","slug":"red-team-evaluation","section":"Debates and Open Questions","sentence":"Field convergence after the Seoul 2024 summit has been slow.","rationale":"[quote-audit demotion] a source was found but its quote did not ESTABLISH the specific claim: The sentence's load-bearing assertion is that 'field convergence after Seoul 2024 has been slow.' The quote is about the content of a red-teaming commitment and says nothing about convergence, pace, or slowness — topical"},{"kind":"topic","slug":"sovereign-ai","section":"Definitional contestation","sentence":"The term entered wide circulation through NVIDIA, whose CEO framed it primarily as national capacity—\"a nation's capabilities to produce artificial intelligence using its own infrastructure, data, workforce and business networks\"—and urged that \"every country needs sovereign AI\" built on domestic \"AI factories\" (NVIDIA 2024, World Governments Summit remarks).","rationale":"[quote-audit demotion] a source was found but its quote did not ESTABLISH the specific claim: The sentence attributes two specific verbatim phrases to NVIDIA's CEO — \"a nation's capabilities to produce artificial intelligence using its own infrastructure, data, workforce and business networks\" and \"every country "},{"kind":"topic","slug":"sovereign-ai","section":"Definitional contestation","sentence":"Critics further note a \"sovereignty as a service\" paradox, in which vendors market compliance wrappers and hardware bundles that produce the appearance of control without delivering meaningful agency (TechPolicy.Press, \"Rethinking Sovereign AI as Strategy,\" 2025).","rationale":"Reports a specific 'sovereignty as a service' critique attributed to a named 2025 TechPolicy.Press article; an external source not in the subject and not among the listed literature items."}]}