{"path":"research/alignment-infrastructure-red-team/shared-assessments-as-coordination-devices-report-2026-11-02.md","content":"# Helper report, verbatim: shared assessments as coordination devices (2026-11-02)\n\n*An `Explore` helper on Sonnet, briefed with the founder's claim (a public record of reasoning could help\nhumans align with each other on the system that serves machine alignment), the angles to cover, the angles\nalready in the corpus (to skip) and a read-status rule. Saved unchanged the moment it arrived, before it was\nread further. Its figures are claims until the main note cites them with their tag; nothing here is ruled.\nThe synthesis is in `../alignment-infrastructure-red-team.md`.*\n\n---\n\nBOTTOM LINE\n\nShared assessments moved agreement when three things held: the parties who had to act accepted the assessment, a standing treaty process kept re-reading it, and a cheap alternative existed (ozone). They did little when a veto-holder's interests were threatened (BWC 2001, emissions since 1988, the 2026 lapse of New START). In AI, the clearest 2026 movement followed third-party measurement of one concrete capability (cyber), not the consensus reports. Many 2026 items rest on press or law-firm summaries; the tags show which.\n\n1. OTHER DOMAINS\n\nIPCC. The effect is argued, not measured. Agrawala (1998) held there would be no UNFCCC or Kyoto without the IPCC. De Pryck (2018) holds such links are claimed rather than shown (Lidskog and Sundqvist, in Hulme (ed.), A Critical Assessment of the IPCC, Cambridge, 2022: https://www.cambridge.org/core/books/critical-assessment-of-the-intergovernmental-panel-on-climate-change/influence/724F6497F4CFE10D4AEC99E7E2838BD2 [part]). The cleanest link: the Paris decision asked for the 1.5C special report (approved Oct 2018: about 45% cut from 2010 levels by 2030, net zero near 2050), and many net-zero pledges followed. No counterfactual study exists (UK Commons Library, https://commonslibrary.parliament.uk/global-net-zero-commitments [secondary]). Fossil CO2 went from about 24 Gt in 1990 (chart reading, https://datawrapper.dwcdn.net/H5DPX/1/ [secondary]) to a projected 38.1 Gt in 2025 (Global Carbon Budget 2025 via Carbon Brief, https://www.carbonbrief.org/analysis-fossil-fuel-co2-emissions-to-set-new-record-in-2025-as-land-sink-recovers/ [secondary]); my arithmetic gives about +58%. Governments still dispute AR6's status, best available science or one contribution (De Pryck 2025, Climatic Change, https://link.springer.com/article/10.1007/s10584-025-04004-4 [part]). Tol (2022) argues the IPCC is too close to policymakers to judge their failures (https://arxiv.org/html/2207.14724 [abstract only]). In January 2026 the US ordered withdrawal from 66 bodies including the IPCC and UNFCCC (Jurist 2026, https://www.jurist.org/news/2026/01/trump-orders-us-withdrawal-from-66-international-organizations-including-un-climate-bodies/ [secondary]). Cash et al. (2003, PNAS) say assessments work when seen as credible, salient and legitimate (https://pmc.ncbi.nlm.nih.gov/articles/PMC166186 [abstract only]).\n\nOzone. Haas (1992, International Organization 46(1)) credits an epistemic community acting mostly through the US and DuPont (https://www.cambridge.org/core/journals/international-organization/article/banning-chlorofluorocarbons-epistemic-community-efforts-to-protect-stratospheric-ozone/6325DC982B4573C378334D80F8835B72 [abstract only]). Critics: Litfin's Ozone Discourses calls that a partial picture (https://www.downtoearth.org.in/reviews/science-backs-the-politically-correct-29188 [secondary]); a 1999 Environmental Politics test found the community's role exaggerated (https://journals.sagepub.com/doi/ref/10.1111/1467-9256.00091 [secondary]). Parson (2003, OUP) still gives assessments a central role in constraining debate (https://academic.oup.com/book/32484 [secondary, publisher text]). Sequence: the Ozone Trends Panel (March 1988) reported 1.7-3% Northern Hemisphere loss; three weeks after refusing, DuPont said it would stop making CFCs as substitutes became available (https://web.njit.edu/~lipuma/Ozone/oadd15.htm [secondary]). London 1990 set phase-out by 2000 and Copenhagen moved it to 1996 (https://www.epa.gov/archive/epa/aboutepa/regulatory-history-cfcs-and-other-stratospheric-ozone-depleting-chemicals-1993.html [secondary]). Article 6 requires reassessment at least every four years (https://www.informea.org/ar/treaties/montreal/text [part]). Outcome: nearly 99% of banned substances phased out, Antarctic recovery about 2066 (UN News 2023, https://news.un.org/en/node/1132277 [secondary]). Assessment, plus a substitute, plus a built-in review clock.\n\nNuclear winter. Gorbachev (2000) called the US-Soviet models a stimulus to act; Reagan cited the idea in February 1985; arsenals fell from about 70,000 to about 12,000 (Robock et al. 2023, https://acp.copernicus.org/articles/23/6691/2023/ [part]; AMS 2026, https://headlines.ametsoc.org/2026/02/23/earth-in-flames/ [part]). Yet Badash (2009), the standard history, says it did not seem to affect Reagan-administration policy (https://doi.org/10.7551/mitpress/9780262012720.003.0022 [part]). Physics Today (Robock, Xia, Mian, July 2026): New START expired in February 2026 with no successor, and the May 2026 NPT review conference failed for the third time running (https://physicstoday.aip.org/features/willful-ignorance-of-the-global-impact-of-nuclear-war [part]). Real elite uptake, contested causal weight, no persistence.\n\nIAEA and CTBTO. In 2024 the IAEA applied safeguards in 190 states; 137 had both a comprehensive agreement and the Additional Protocol; the broader conclusion was drawn for 75 (EU statement 2025, https://www.eeas.europa.eu/delegations/vienna-international-organisations/eu-statement-iaea-board-governors-safeguards-implementation-report-2024-10-june-2025_en [secondary]). Safeguards missed Iraq's undeclared programme before 1991; the Additional Protocol followed on 15 May 1997 (https://www.osti.gov/biblio/1001740 [secondary]). Access still rests on consent: in February 2026 the IAEA said it could not report the size or location of Iran's stockpile, 440.9 kg of up to 60% enriched uranium before the June 2025 attacks (Al Jazeera 2026, https://aljazeera.com/news/2026/2/27/iaea-eyes-isfahan-nuclear-complex-as-it-urges-iran-to-allow [secondary]). CTBTO: 307 of 337 facilities certified, all six declared nuclear tests this century detected, 178 ratifications, nine more needed (Floyd, 30 Oct 2025, mirror copy, https://kndc.kz/index.php/en/home-en/Smart%20Search/Joomla!%20Update/262-statement-by-robert-floyd-executive-secretary-of-the-ctbto-press-release-statement-by-robert-floyd-executive-secretary-of-the-ctbto [full text]). Russia de-ratified in 2023; Trump ordered test resumption in October 2025 (https://www.armscontrol.org/act/2025-11/news-briefs/trump-says-us-will-resume-nuclear-testing [secondary]). Shared measurement made cheating detectable; it did not make anyone willing.\n\nBWC. A scientific group (VEREX) assessed verification, and a 24-session negotiation (1995-2001) produced a chair's composite text. The US rejected it on 25 July 2001, citing inadequate rigour and exposure of its biodefence and pharma sectors (Mathews, https://www.the-trench.org/wp-content/uploads/2022/07/HN003-BTWC-Protocol-Negotiations-Mathews.pdf [part]). Advocate Pearson says over 50 of about 55 negotiating states backed it (https://opbw.org/sbtwc/evalu22.pdf [part]). The support unit has four staff and about $2.1M a year, against about $94M for the OPCW in 2025 (NTI 2025, https://www.nti.org/risky-business/the-next-50-years-strengthening-the-biological-weapons-convention-explained/ [secondary]). In February 2026 the working group adopted only a procedural report and the chair circulated a further draft decision on the last day; results go to the 2027 review conference (BWPP reports 2026, https://www.cbw-events.org.uk/BWC26-06.pdf [secondary]). Technical consensus plus 50 states did not beat one veto. Biodiversity is a second miss: none of 20 Aichi targets fully achieved (Global Biodiversity Outlook 5, 2020, https://cieem.net/uns-fifth-global-biodiversity-outlook-highlights-failure-in-meeting-aichi-targets/ [secondary]).\n\n2. GATEWAY BELIEF AND CONSENSUS MESSAGING\n\nVan der Linden et al. (2015, PLOS ONE, N=1,104): the message raised perceived consensus by 12.8 points, and effects on belief and support ran through it; the authors say the controlled design may overstate real-world effect (https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0118489 [part]). Kahan (2017, JCOM): against a proper control, belief differed by 2.49 points (p=.17) and support by 2.12 (p=.11) on a 101-point scale, with Bayes factors 12 and 9 favouring no effect (https://jcom.sissa.it/archive/16/05/JCOM_1605_2017_A03 [part]). The 2019 replication (N=6,301, J. Env. Psych. 62:49-58) says the model held (https://www.repository.cam.ac.uk/handle/1810/289506 [abstract only]); a pre-registered China replication (N=1,506, 2025) agrees (https://cris.maastrichtuniversity.nl/en/publications/the-gateway-belief-model-a-pre-registered-large-scale-replication/ [secondary]). Rode et al. (2021) meta-analysis: Hedges g = 0.09, 95% CI 0.05-0.13, no different from other interventions; Chinn and Hart (2021) find inconsistent results (https://pmc.ncbi.nlm.nih.gov/articles/PMC8380189 [secondary]). A 63-country tournament (Vlasceanu et al. 2024, Science Advances; 59,440 people, 11 interventions) found small effects, best +2.3% on belief and +2.6% on policy support (https://researchonline.lse.ac.uk/id/eprint/122079 [abstract only]). I found no experiment testing whether perceived expert consensus on AI risk moves policy support. The expert base is itself split: of 2,778 researchers, 38-51% gave at least a 10% chance to extinction-level outcomes, depending on the question (Grace et al. 2024, https://jair.org/index.php/jair/article/view/19087 [secondary]).\n\n3. AI-SPECIFIC\n\nInternational AI Safety Report. Second edition, 3 Feb 2026, chaired by Bengio: over 100 experts (93 named), an advisory panel of 29 countries plus the EU, OECD and UN, no policy recommendations. It notes that 12 companies published or updated frontier safety frameworks in 2025 and that models increasingly tell tests from deployment (https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026 [part]). The US did not back the 2026 edition (TIME headline, via https://en.wikipedia.org/wiki/India_AI_Impact_Summit_2026 [secondary]). I found no documented decision that cites it. First edition, Jan 2025: 96 experts, 30 countries (https://babl.ai/?p=7087 [secondary]).\n\nUN Scientific Panel. Created by A/RES/79/325 (26 Aug 2025); 40 members chosen from 2,600+ applications; Bengio and Ressa co-chairs since 3 Mar 2026; preliminary report 1 Jul 2026 says safeguards cannot keep pace with capabilities; first Global Dialogue Geneva 6-7 Jul 2026, second May 2027 (https://www.un.org/independent-international-scientific-panel-ai/en/faq [full text]; https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report [part]). A co-chair reportedly set a bar of broad scientific consensus for including findings (press summaries, unverified). No commitment followed that I could find.\n\nAISI network. Now the International Network for Advanced AI Measurement, Evaluation and Science, 10 members; first best-practice guidance posted 23 Jul 2026; no joint test results reported (https://www.aisi.gov.uk/blog/international-evaluation-best-practice-and-open-questions-in-ai-measurement [part]). UK AISI Trends Report (Dec 2025): 30+ models tested; apprentice-level cyber success 9% to 50%; universal jailbreaks in every system tested (https://www.aisi.gov.uk/blog/5-key-findings-from-our-first-frontier-ai-trends-report [part]).\n\nThe one measured chain (2026). UK AISI found Mythos averaged 22 of 32 steps in its network-attack simulation, against 16 for the next model (CMS law firm, 17 Apr 2026, https://cms.law/en/gbr/legal-updates/mythos-project-glasswing-and-regulating-catastrophic-risk-caused-by-ai-models [secondary]). On 4 May the White House was reported weighing pre-release review (https://www.csoonline.com/article/4166824/anthropic-mythos-spurs-white-house-to-weigh-pre-release-reviews-for-high-risk-ai-models.html [secondary]). On 5 May Google DeepMind, Microsoft and xAI signed CAISI pre-deployment testing agreements, joining OpenAI and Anthropic (https://labs.cloudsecurityalliance.org/research/csa-research-note-caisi-frontier-ai-testing-agreements-20260/ [secondary]). All voluntary. Weeks, not years, but on a measured capability that is also a US security interest.\n\nSummits and company pledges. Bletchley (Nov 2023): 28 countries plus the EU, which mandated the report (https://www.longtermwiki.com/knowledge-base/responses/governance/bletchley-declaration [secondary]). Seoul (May 2024): 16 companies signed, 20 by Feb 2025; by Dec 2025 only 12 companies had published frontier safety policies, not all of them signatories (METR 2025, https://metr.org/blog/2025-12-09-common-elements-of-frontier-ai-safety-policies/ [part]). Paris (Feb 2025): about 60 signed, US and UK did not (https://silicon.co.uk/cloud/ai/us-uk-refuse-to-sign-ai-declaration-599396 [secondary]). New Delhi (Feb 2026): 86-89 endorsements depending on source, US and China reportedly among them, thin on safety; 13 firms' voluntary pledges cover usage evidence and multilingual evaluation only (https://www.globalgovernmentforum.com/declaration-on-leveraging-ai-for-growth-and-good-endorsed-at-global-summit-in-india/ [part]). Next: Geneva 2027. The Frontier Model Forum lists best practices, research and information sharing, with no dated commitments (https://www.frontiermodelforum.org/ [part]).\n\nEU Code of Practice. Published 10 Jul 2025; the Commission lists 22 signatories (Euronews counted about 26 on 1 Aug 2025); Meta refused; xAI signed only the safety chapter (https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai [part]; https://www.euronews.com/next/2025/08/01/some-26-tech-companies-sign-up-to-eu-commissions-ai-code [part]). The Commission's power to fine (up to 3% of turnover) began 2 Aug 2026 (https://cms.law/en/gbr/legal-updates/mythos-project-glasswing-and-regulating-catastrophic-risk-caused-by-ai-models [part]); I found no enforcement case. California's frontier AI act, in force from 1 Jan 2026, also requires large developers to assess catastrophic risk and report incidents (same source, [part]). Law, not the reports, is what made a shared risk framing binding.\n\nAutumn 2026.\n- US-China: after the late-September Washington summit, an AI incident channel and a first AI dialogue in November; no breakthroughs; Trump said the US is \"not going to be putting on brakes\" (https://www.pbs.org/newshour/world/china-and-u-s-agree-to-establish-ai-safety-channel-and-continue-trade-and-military-talks [part]).\n- Amodei's September essay urged pacing, chip controls and cooperation with China; Altman and Musk backed it; China's foreign ministry called it fear-mongering; Trump and Zuckerberg rejected it (https://restofworld.org/2026/china-rejects-dario-amodei-anthropic-ai-slowdown-trump-xi/ [part]; outlets differ on details).\n- Finland/Norway-led call, 21 Sep 2026, 33 signatories including von der Leyen: explore an international institution to set standards, enable verification and convene at capability thresholds; no design, verification method or threshold definition (https://www.presidentti.fi/en/a-call-for-control-of-frontier-ai-models/ [full text]).\n- Berkeley working group (26 researchers, 200 pages, 9 Oct 2026): a verified pause of 10 years or more is feasible by halting training-chip production; no government response reported (https://news.berkeley.edu/2026/10/09/a-global-hardwired-pause-of-frontier-ai-training-is-feasible-says-new-report/ [part]).\n- Chip Security Act: House committee vote 42-0 in March 2026, chip industry opposed; I found no enactment by mid-October (https://www.nbcnews.com/tech/tech-news/chips-security-act-gains-industry-support-letter-rcna350500 [part]).\n- Red Lines call (UNGA Sept 2025; goal: agreement by end 2026): 300+ figures, 90+ organisations, no government adoption reported (https://red-lines.ai/ [part]). The Oct 2025 Statement on Superintelligence sets its lift condition as broad scientific consensus plus strong public buy-in (https://futureoflife.org/open-letter/awos-signatories [part]).\n- I found no treaty negotiation or verification regime.\n\n4. THEORY AND THE REALIST COUNTER\n\nFarrell and Rabin (1996, JEP 10(3):103-118) say both \"talk always works\" and \"talk never works\" are wrong (https://ideas.repec.org/a/aea/jecper/v10y1996i3p103-18.html [abstract only]). Aumann (1990): in a stag hunt, announcing stag is not self-enforcing; Farrell and Rabin expect talk to help anyway; experiments find such agreements fragile (Clark, Kay, Sefton 2001, https://ideas.repec.org/a/spr/jogath/v29y2001i4p495-515.html [secondary]). Skyrms (2004) makes the stag hunt the social-contract game, with signalling one route to trust (https://www.librarything.com/work/205158 [secondary]). Fearon (1995) lists private information, commitment problems and indivisibility as causes of failed bargains; shared assessments address the first (https://ideas.repec.org/a/cup/intorg/v49y1995i03p379-414_03.html [secondary]; the commitment point is from background knowledge, not re-read). Schelling's focal points: background knowledge, not re-read.\n\nRealist counter. I found no rigorous 2024-26 realist paper on AI racing that I would call definitive. The strongest case is behavioural: experts in both countries acknowledge extreme risks (https://arxiv.org/html/2407.16903 [part]), yet both national action plans treat safety as secondary, and neither wants strong safety measures a rival might ignore (Singer and Zvenyhorodskyi, Just Security 2025, https://www.justsecurity.org/119509/us-chinese-ai-playbooks/ [part]). Theory: a neorealist paper (Bik 2026, unreviewed) says AI compresses decision time, blurs offence and defence, and resists verification (https://www.researchgate.net/publication/415154739_Artificial_Intelligence_and_the_Security_Dilemma_Extending_Neorealist_Theory_into_the_Digital_Age [part]). Delaney (IAPS 2025) notes that offensive realism makes war more acceptable than backing down, and that deterrence stability depends on verification (https://www.iaps.ai/research/crucial-considerations-in-asi-deterrence [part]). Against it: defensive-realist models make a superintelligence race a trust dilemma that verification can fix (https://arxiv.org/html/2501.14749v1 [part]); Roussel et al. (2026) say a moratorium serves self-interest once the perceived loss-of-control cost is high enough (https://arxiv.org/abs/2605.01297 [abstract only]); Ó hÉigeartaigh (2025) calls the race narrative partly self-fulfilling (https://papers.ssrn.com/abstract=5278644 [abstract only]).\n\nTABLE\n\n| Case | Shared assessment | Coordination followed | What did not follow | Status |\n|---|---|---|---|---|\n| IPCC 1988-2026 | Periodic consensus reports; SR1.5 (2018) | UNFCCC and Paris asked for reports; net-zero pledges (link unproven) | Fossil CO2 about +58% since 1990; US exit 2026 | [part], [secondary] |\n| Ozone | Ozone Trends Panel 1988; four-yearly panels | CFC phase-out; about 99% of banned substances gone | Needed a substitute and a US/DuPont shift | [abstract only], [secondary] |\n| Nuclear winter | 1983 US and Soviet climate models | Leader statements 1985; arsenal cuts (weight contested) | Badash: no effect on Reagan policy; New START lapsed 2026 | [part] |\n| IAEA safeguards | Common measurement of declared material | Broader conclusion for 75 states | Missed Iraq pre-1991; Iran stockpile unverified 2026 | [secondary] |\n| CTBTO | Global monitoring network, 307 of 337 sites | All six 21st-century tests detected | Treaty not in force, nine ratifications short | [full text] |\n| BWC | VEREX; 2001 composite text | 50-plus states backed it | US rejection 2001; four-person support unit | [part] |\n| Consensus messaging | Perceived-consensus message | Perceived consensus +12.8 points (2015) | Belief and support: null (Kahan) or g=0.09 | [part], [secondary] |\n| IASR 2025-26, UN Panel 2026 | Annual syntheses, no recommendations | No citing decision found | US withheld backing 2026; no UN commitment | [part] |\n| AISI network, Mythos chain 2026 | Third-party evaluation (22 of 32 steps) | Five labs under CAISI testing, May 2026 | Voluntary; Meta not party; no joint test results | [secondary] |\n| Summits, Seoul, FMF, EU Code | Declarations, frameworks, code | 12 firms with frameworks; 22 listed Code signers; EU fines from Aug 2026 | Meta refused; US and UK skipped Paris; Delhi thin on safety | [part], [secondary] |\n| Autumn 2026 proposals | Nordic-led call, Berkeley pause report, red lines | US-China incident channel, November dialogue | No treaty, no verification regime | [full text], [part] |\n\nWHAT THIS SAYS FOR A PUBLIC CLAIM GRAPH\n\nThe record says a shared assessment coordinates when the parties who must act accept it, a standing process keeps re-reading it, it measures something observable, and no veto-holder's core interest is threatened; where those failed (BWC 2001, the 2026 US exits, the 2026 AI proposals), belief was shared and coordination still failed. A public claim graph is therefore a weak device for persuading the public (consensus-messaging effects are small and contested) or restraining states in a race (both sides already acknowledge the risks), but a credible one for the missing layer: an inspectable ledger tying each commitment to the evidence it is conditional on, so a measured result like the Mythos evaluation maps to who agreed to what within weeks. The most testable use is to record conditional commitments (Seoul-style thresholds, the Statement on Superintelligence's lift condition, the EU Code's risk duties) against third-party measurements, and to judge the graph by whether any actor's stated commitment changes when the measurement does.\n"}