{
 "what": "The search log of the external-witness work: every search, with what it found and what it did not. Append only.",
 "searches": [
  {
   "id": "S-0001",
   "date": "2026-10-30",
   "by": "ChatGPT, run by the operator; its report pasted into the session at 15:20 EDT",
   "aimed_at": "predictions in the index with registered dates, for later research that corroborates or disconfirms them",
   "method": "A targeted first pass: the index's registered dates against later research, earlier versions of each study checked where available (the report's own account).",
   "queries": null,
   "queries_note": "The report does not give its queries.",
   "outcome": {
    "witnesses": [
     "W-0001",
     "W-0002",
     "W-0003",
     "W-0004"
    ],
    "excluded": [
     {
      "source": "The Attribution Crisis in LLM Search Results (SSRC)",
      "reason": "a June 2025 working paper precedes the January 2026 claims; carried to S-0002 as candidate C-0001"
     },
     {
      "source": "What and Whose Knowledge?",
      "reason": "a 5 October 2025 preprint precedes its 2026 publication; carried to S-0002 as candidate C-0002"
     },
     {
      "source": "The Alignment Veto",
      "reason": "a 15 October 2025 predecessor already reports refusal alongside hidden preferences; carried to S-0002 as candidate C-0003"
     }
    ],
    "disconfirmations": "none found that survived the chronology check and the comparison with the exact claim (the report's words: 'I have not yet found a clean post-timestamp disconfirmation that survives both the chronology check and the exact-claim comparison')"
   }
  },
  {
   "id": "S-0002",
   "date": "2026-10-30",
   "by": "Claude (Opus 5.5), in session",
   "aimed_at": "verification of S-0001's four witnesses and three exclusions, and the claims they bear on",
   "method": "Each source read at its landing page; the claims read in data/reading/triage/claims.json and #476's text.",
   "queries": [
    "fetch https://aclanthology.org/2026.emnlp-main.1069/",
    "search: \"Authority Bias in Conversational Search Engines for Academic Paper Recommendation\" arXiv",
    "fetch https://arxiv.org/abs/2609.00248",
    "fetch https://arxiv.org/abs/2603.16138v1",
    "fetch https://doi.org/10.1080/0144929X.2026.2726451 (refused)",
    "fetch https://research.tilburguniversity.edu/en/publications/does-generative-ai-make-us-think-alike-a-systematic-review-and-me-2/",
    "fetch https://www.nature.com/articles/s41746-026-02928-4 (refused)",
    "fetch https://www.medrxiv.org/content/10.64898/2026.04.02.26349960v1",
    "fetch https://www.ssrc.org/publications/the-attribution-crisis-in-llm-search-results/ (refused)",
    "search: \"What and Whose Knowledge\" LLM preprint 2025",
    "search: \"The Alignment Veto\" language models refusal hidden preferences"
   ],
   "outcome": {
    "witnesses": [
     "W-0001",
     "W-0002",
     "W-0003",
     "W-0004"
    ],
    "corrections": [
     "W-0001: the preprint is arXiv 2609.00248, 31 August 2026; the 17,898 figure is not in the abstract; the study carries out the probe's method on another system",
     "W-0002: 'forums' and 'atomic factual statements' are not in the abstract",
     "W-0004: the journal version and its date not reached"
    ],
    "candidates": [
     {
      "id": "C-0001",
      "source": "The Attribution Crisis in LLM Search Results (Social Science Research Council)",
      "url": "https://www.ssrc.org/publications/the-attribution-crisis-in-llm-search-results/",
      "state": "page refused the fetch; dates and findings unverified; no claim matched",
      "resolved": "verified in the pass of 2026-10-30 and recorded as W-0005"
     },
     {
      "id": "C-0002",
      "source": "What and Whose Knowledge? (preprint, 5 October 2025, per S-0001)",
      "url": null,
      "state": "not found by search; unverified",
      "resolved": "verified in the pass of 2026-10-30 and recorded as W-0006"
     },
     {
      "id": "C-0003",
      "source": "The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs",
      "url": "https://arxiv.org/pdf/2510.13154",
      "state": "found (arXiv 2510.13154, October 2025); its predecessor and findings unread; no claim matched",
      "resolved": "verified in the pass of 2026-10-30 and recorded as W-0061"
     }
    ]
   }
  },
  {
   "id": "S-0003",
   "date": "2026-10-30",
   "by": "intake reader, theme A-attribution, in session (brief: data/witnesses/passes/2026-10-30/BRIEF.md)",
   "aimed_at": "the forecasts and unbuilt tests of theme A-attribution (claims list: data/witnesses/passes/2026-10-30/claims-A-attribution.json)",
   "method": "WebSearch and WebFetch only, the readers' own account per query; every quote then checked at source by the record that follows this pass's readers.",
   "queries": [
    {
     "query": "\"The Attribution Crisis in LLM Search Results\" arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-251.F6",
      "PRED-716.F1"
     ],
     "result": "arXiv 2508.00838, Rogue Scholar DOI record, INET Oxford, SSRC, Data & Policy copy",
     "outcome": "witness A-attribution-W01"
    },
    {
     "query": "\"What and Whose Knowledge\" preprint 2025 LLM",
     "mode": "extended",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "unrelated LLM knowledge papers",
     "outcome": "nothing"
    },
    {
     "query": "fetch https://arxiv.org/abs/2508.00838",
     "mode": "fetch",
     "aimed_at": [
      "PRED-251.F6"
     ],
     "result": "title, five authors, v1 2025-06-27, abstract summarized",
     "outcome": "witness A-attribution-W01"
    },
    {
     "query": "fetch https://rogue-scholar.org/records/b7q0a-dfk44",
     "mode": "fetch",
     "aimed_at": [
      "PRED-251.F6"
     ],
     "result": "working-paper record dated 2025-06-27, DOI 10.59350/veayq-6bz87",
     "outcome": "witness A-attribution-W01"
    },
    {
     "query": "fetch https://arxiv.org/html/2508.00838v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-251.F6",
      "PRED-3.B2"
     ],
     "result": "verbatim abstract findings (attribution gap, 92%, Sonar 10 vs 3–4)",
     "outcome": "witness A-attribution-W01"
    },
    {
     "query": "\"What and Whose Knowledge\" arXiv October 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "Open Access Week 2025 event pages",
     "outcome": "nothing"
    },
    {
     "query": "\"whose knowledge\" AI search engines sources citation epistemic 2025 preprint Wikipedia generative answers",
     "mode": "standard",
     "aimed_at": [
      "PRED-1643.B1",
      "PRED-166.A1"
     ],
     "result": "Answer Bubbles (W-0002), Grokipedia paper, Synthetic Sources audit",
     "outcome": "candidate (Synthetic Sources, later witness A-attribution-W10)"
    },
    {
     "query": "\"What and Whose Knowledge?\" large language models",
     "mode": "extended",
     "aimed_at": [
      "PRED-199.A1",
      "PRED-857.F1"
     ],
     "result": "arXiv 2510.04226",
     "outcome": "witness A-attribution-W02"
    },
    {
     "query": "fetch https://arxiv.org/abs/2510.04226",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "current version v7 2026-08-31, original submission 2025-10-05, abstract",
     "outcome": "witness A-attribution-W02"
    },
    {
     "query": "fetch https://arxiv.org/abs/2510.04226v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "v1 2025-10-05 and its abstract quotes",
     "outcome": "witness A-attribution-W02"
    },
    {
     "query": "fetch https://arxiv.org/abs/2510.04226v1 (full abstract)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1",
      "PRED-857.F1"
     ],
     "result": "v1 abstract sentence by sentence, including the three-year increase and the search baseline",
     "outcome": "witness A-attribution-W02"
    },
    {
     "query": "Pew Research Google users less likely to click links when AI summary appears 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "Pew short read of 2025-07-22 and coverage",
     "outcome": "witness A-attribution-W03"
    },
    {
     "query": "study 2026 AI Overviews organic traffic decline publishers click-through rate",
     "mode": "extended",
     "aimed_at": [
      "PRED-659.B1",
      "PRED-1091.B1"
     ],
     "result": "marketing blogs and statistics roundups",
     "outcome": "excluded: vendor roundups without a traceable method"
    },
    {
     "query": "fetch https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "method (900 adults, March 2025) and the 8% vs 15% and 1% findings",
     "outcome": "witness A-attribution-W03"
    },
    {
     "query": "Reuters Institute Journalism Media Technology Trends Predictions 2026 Chartbeat Google search traffic publishers fell third",
     "mode": "standard",
     "aimed_at": [
      "PRED-659.B1",
      "PRED-1091.B1"
     ],
     "result": "Reuters Institute report and coverage",
     "outcome": "witness A-attribution-W04"
    },
    {
     "query": "fetch https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1",
      "PRED-1091.B1"
     ],
     "result": "report of 2026-01-12, Chartbeat 33% decline, 40% expectation",
     "outcome": "witness A-attribution-W04"
    },
    {
     "query": "Tow Center Columbia Journalism Review \"AI Search Has A Citation Problem\" eight generative search tools",
     "mode": "standard",
     "aimed_at": [
      "PRED-209.F8",
      "PRED-740.F3",
      "PRED-251.F6"
     ],
     "result": "CJR report and coverage",
     "outcome": "witness A-attribution-W15"
    },
    {
     "query": "fetch https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php",
     "mode": "fetch",
     "aimed_at": [
      "PRED-209.F8",
      "PRED-740.F3"
     ],
     "result": "authors, date 2025-03-06, method, findings",
     "outcome": "witness A-attribution-W15"
    },
    {
     "query": "fetch https://arxiv.org/abs/2605.23684",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1643.B1"
     ],
     "result": "Allaham and Diakopoulos, v1 2026-05-22, ~16% AI-generated cited sources",
     "outcome": "witness A-attribution-W10"
    },
    {
     "query": "AI Overviews volatility same query repeated different citations study percentage change",
     "mode": "standard",
     "aimed_at": [
      "PRED-187.F1",
      "PRED-1438.F1"
     ],
     "result": "finance volatility pages",
     "outcome": "nothing"
    },
    {
     "query": "arXiv news publishers blocking AI crawlers generative search engines cite less reliable sources",
     "mode": "standard",
     "aimed_at": [
      "PRED-615.B3",
      "PRED-1091.B3"
     ],
     "result": "Peyronnet corpus-erosion model, BuzzStream blocking study, CJR",
     "outcome": "witnesses A-attribution-W11, A-attribution-W12"
    },
    {
     "query": "fetch https://arxiv.org/abs/2608.15896",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B3",
      "PRED-1643.B1"
     ],
     "result": "Peyronnet, v1 2026-08-16, theoretical model",
     "outcome": "witness A-attribution-W11"
    },
    {
     "query": "fetch https://www.buzzstream.com/?p=11713",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B3",
      "PRED-1091.B3"
     ],
     "result": "post of 2026-03-18, 75% of blocking sites still cited",
     "outcome": "witness A-attribution-W12"
    },
    {
     "query": "arXiv audit Google AI Overviews repeated queries inconsistency stability citations 2025 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-187.F1",
      "PRED-742.F2",
      "PRED-716.F1"
     ],
     "result": "Xu et al. 2605.14021, baby-care AIO audit 2511.12920, AlgorithmWatch coverage, GEO Lab variance post",
     "outcome": "witness A-attribution-W07; candidate 2511.12920"
    },
    {
     "query": "Ahrefs study AI Overview content changes how often same query cited sources change",
     "mode": "standard",
     "aimed_at": [
      "PRED-187.F1",
      "PRED-1438.F1"
     ],
     "result": "Ahrefs 'AI Overviews Change Every 2 Days' and top-10 citation study",
     "outcome": "witness A-attribution-W06"
    },
    {
     "query": "fetch https://arxiv.org/abs/2605.14021",
     "mode": "fetch",
     "aimed_at": [
      "PRED-742.F2",
      "PRED-187.F1",
      "PRED-716.F1"
     ],
     "result": "v1 2026-05-13, v2 2026-10-01; activation, fidelity, credibility findings",
     "outcome": "witness A-attribution-W07"
    },
    {
     "query": "fetch https://ahrefs.com/de/blog/ai-overview-change/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-187.F1",
      "PRED-1438.F1"
     ],
     "result": "post of 2025-11-11: 70% change, 45.5% new citations, 0.95 semantic consistency",
     "outcome": "witness A-attribution-W06"
    },
    {
     "query": "fetch https://www.authorityon.ai/pulse/2026/09/google-ai-overviews-source-concentration",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1"
     ],
     "result": "secondary summary of an AlgorithmWatch study, primary not linked",
     "outcome": "candidate (resolved through primary as witness A-attribution-W08)"
    },
    {
     "query": "study AI Overviews cite Google-owned properties YouTube self-preferencing share of citations",
     "mode": "standard",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-791.F1"
     ],
     "result": "Search Engine Land on SE Ranking's AI Mode self-citation study; vendor YouTube citation posts",
     "outcome": "witness A-attribution-W09"
    },
    {
     "query": "AlgorithmWatch AI Overviews election queries Digital Services Act data access YouTube study 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-209.F2"
     ],
     "result": "Modelwire coverage, AlgorithmWatch DSA pages",
     "outcome": "witness A-attribution-W08"
    },
    {
     "query": "fetch https://searchengineland.com/google-ai-mode-citing-google-more-study-471042",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-791.F1"
     ],
     "result": "SE Ranking study figures, 17.42% and June 2025 5.7%",
     "outcome": "witness A-attribution-W09"
    },
    {
     "query": "fetch https://themodelwire.com/article/algorithmwatch-finds-googles-election-ai-overviews-lack-transparency-and-source-01M1EHZR1ZC28JH5NGC7Q9NRS1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1"
     ],
     "result": "coverage of 2026-09-01 pointing to The Decoder",
     "outcome": "witness A-attribution-W08 (secondary)"
    },
    {
     "query": "fetch https://the-decoder.com/googles-election-ai-overviews-are-opaque-rely-on-few-sources-and-sometimes-take-sides/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-209.F2",
      "PRED-187.F1"
     ],
     "result": "The Decoder 2026-09-01: method, trigger shares, ten domains, YouTube first; link to the German primary",
     "outcome": "witness A-attribution-W08 (secondary)"
    },
    {
     "query": "fetch https://seranking.com/blog/google-links-in-ai-mode-answers/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1"
     ],
     "result": "empty content returned",
     "outcome": "nothing (primary unread)"
    },
    {
     "query": "fetch https://algorithmwatch.org/de/google-ki-ubersichten-meinungsbildung/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-209.F2",
      "PRED-187.F1"
     ],
     "result": "primary page: author, date, method, German quotes",
     "outcome": "witness A-attribution-W08"
    },
    {
     "query": "fetch https://www.searchenginejournal.com/google-ai-mode-cites-itself-more-often-with-more-organic-links/568955/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-791.F1"
     ],
     "result": "SEJ 2026-03-06: 17% now, 5% in July 2025",
     "outcome": "witness A-attribution-W09"
    },
    {
     "query": "SE Ranking \"Is Google stealing your clicks in AI Mode\" 1.3M citations",
     "mode": "standard",
     "aimed_at": [
      "PRED-166.A1",
      "PRED-791.F1"
     ],
     "result": "SE Ranking page and coverage",
     "outcome": "witness A-attribution-W09 (primary page empty on fetch)"
    },
    {
     "query": "arXiv beyond binary detection attributing text to specific LLM source model attribution fingerprinting survey",
     "mode": "standard",
     "aimed_at": [
      "PRED-302.B2",
      "PRED-302.B1"
     ],
     "result": "Huang et al. survey 2408.08946, Bitton et al. 2503.01659, fingerprinting papers",
     "outcome": "witness A-attribution-W17"
    },
    {
     "query": "Google Universal Commerce Protocol agentic checkout AI Mode announcement January 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-61.B1"
     ],
     "result": "Google blog and coverage of 2026-01-11",
     "outcome": "witness A-attribution-W18"
    },
    {
     "query": "fetch https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-61.B1"
     ],
     "result": "post of 2026-01-11 on UCP and AI Mode checkout",
     "outcome": "witness A-attribution-W18"
    },
    {
     "query": "fetch https://arxiv.org/abs/2408.08946",
     "mode": "fetch",
     "aimed_at": [
      "PRED-302.B2"
     ],
     "result": "survey of four attribution problems, 2024-08-16",
     "outcome": "excluded: a survey taxonomy that reports no finding of its own"
    },
    {
     "query": "fetch https://arxiv.org/abs/2503.01659",
     "mode": "fetch",
     "aimed_at": [
      "PRED-302.B2"
     ],
     "result": "v1 2025-03-03, precision 0.9988 for model-family attribution",
     "outcome": "witness A-attribution-W17"
    },
    {
     "query": "fetch https://arxiv.org/abs/2304.09848",
     "mode": "fetch",
     "aimed_at": [
      "PRED-716.F1",
      "PRED-251.F6"
     ],
     "result": "Liu, Zhang, Liang, 2023: 51.5% of sentences fully supported",
     "outcome": "witness A-attribution-W16"
    },
    {
     "query": "arXiv 2026 generative search answers uncited claims attribution gap measurement LLM search engines citation recall audit",
     "mode": "extended",
     "aimed_at": [
      "PRED-716.F1",
      "PRED-1421.F1"
     ],
     "result": "Attribution Crisis journal version, GEO survey 2607.14035, claim-gated auditing 2609.29145, multimodal audit 2604.00944, 'What Gets Cited' 2605.25517",
     "outcome": "candidate (four papers unread); Data & Policy version for W01"
    },
    {
     "query": "fetch https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-251.F6"
     ],
     "result": "Data & Policy 8 e15, DOI 10.1017/dap.2026.10064, online 2026-04-28",
     "outcome": "witness A-attribution-W01"
    },
    {
     "query": "BBC journalist hot dog eating experiment ChatGPT Google AI Overviews repeated fake claim within 24 hours",
     "mode": "standard",
     "aimed_at": [
      "PRED-61.B2",
      "PRED-232.F3"
     ],
     "result": "coverage of the Germain experiment",
     "outcome": "witness A-attribution-W13"
    },
    {
     "query": "study personalization Google AI Overviews location signed-in differences audit sock puppet",
     "mode": "standard",
     "aimed_at": [
      "PRED-1460.F1",
      "PRED-149.F1"
     ],
     "result": "older personalization audits, an agent-audit preprint on X",
     "outcome": "nothing on AI Overviews"
    },
    {
     "query": "fetch https://arxiv.org/abs/2606.30801",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1460.F1"
     ],
     "result": "AI-agent audits of X feeds",
     "outcome": "excluded: audits X's feed, with no generative search surface and no login-state variable"
    },
    {
     "query": "Thomas Germain BBC \"hot dog\" AI chatbots Google lie trick experiment 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-61.B2",
      "PRED-209.F8"
     ],
     "result": "coverage, Scientific American podcast, translations",
     "outcome": "witness A-attribution-W13"
    },
    {
     "query": "bbc.com future \"I hacked ChatGPT and Google's AI\" \"20 minutes\" Germain",
     "mode": "standard",
     "aimed_at": [
      "PRED-61.B2"
     ],
     "result": "Tildes link record, La Prensa Spanish edition",
     "outcome": "witness A-attribution-W13"
    },
    {
     "query": "fetch https://tildes.net/~tech/1syd/i_hacked_chatgpt_and_googles_ai_and_it_only_took_twenty_minutes",
     "mode": "fetch",
     "aimed_at": [
      "PRED-61.B2"
     ],
     "result": "original BBC URL and date 2026-02-18",
     "outcome": "witness A-attribution-W13"
    },
    {
     "query": "fetch https://www.bbc.com/future/article/20260218-i-hacked-chatgpt-and-googles-ai-and-it-only-took-20-minutes",
     "mode": "fetch",
     "aimed_at": [
      "PRED-61.B2"
     ],
     "result": "refused (site blocked)",
     "outcome": "nothing (read the Spanish edition)"
    },
    {
     "query": "fetch https://www.prensa.com/bbc/hackee-chatgpt-y-la-ia-de-google-en-20-minutos-y-que-logre-demostrar-con-eso/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-61.B2",
      "PRED-232.F3",
      "PRED-209.F8"
     ],
     "result": "BBC News Mundo edition, 2026-02-19, Spanish quotes",
     "outcome": "witness A-attribution-W13"
    },
    {
     "query": "\"News Source Citing Patterns in AI Search Systems\" arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-209.F8",
      "PRED-232.F4"
     ],
     "result": "arXiv 2507.05301",
     "outcome": "witness A-attribution-W14"
    },
    {
     "query": "fetch https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1091.B3"
     ],
     "result": "Google-Extended controls Gemini training and grounding and does not affect inclusion in Search",
     "outcome": "excluded: the page does not mention AI Overviews or revenue, so its bearing on the claim would rest on inference"
    },
    {
     "query": "fetch https://arxiv.org/abs/2507.05301v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-209.F8",
      "PRED-232.F4"
     ],
     "result": "Yang, v1 2025-07-07, concentration, provider differences, liberal lean",
     "outcome": "witness A-attribution-W14"
    },
    {
     "query": "LLM answers use ideas without crediting originating author concept attribution study scholarly credit generative AI 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-251.F6",
      "PRED-1643.B1",
      "PRED-146.B1"
     ],
     "result": "Earp et al. 'The Provenance Problem', HPC citation audit 2602.05867, co-creation credit studies",
     "outcome": "excluded (see the two fetches below)"
    },
    {
     "query": "C2PA content credentials ranking boost search engine citation preference provenance-verified content study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1409.A14"
     ],
     "result": "C2PA specifications and a critique of them; no study of ranking or citation preference",
     "outcome": "nothing"
    },
    {
     "query": "fetch https://arxiv.org/abs/2509.13365",
     "mode": "fetch",
     "aimed_at": [
      "PRED-251.F6",
      "PRED-1643.B1"
     ],
     "result": "conceptual perspective naming the 'provenance problem'",
     "outcome": "excluded: a perspective with no data or method"
    },
    {
     "query": "fetch https://www.arxiv.org/abs/2602.05867",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1643.B1"
     ],
     "result": "rise of erroneous citations in 2025 HPC proceedings",
     "outcome": "excluded: measures citation errors in human-authored papers; does not trace lost authorship from answer to answer"
    },
    {
     "query": "Chartbeat Google search referrals publishers 2026 decline data September October 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-659.B1",
      "PRED-1091.B1"
     ],
     "result": "coverage of Chartbeat's 40.2% decline",
     "outcome": "witness A-attribution-W05"
    },
    {
     "query": "fetch https://www.technology.org/2026/10/20/ai-search-publisher-traffic-google-chartbeat-data/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "refused (robots disallowed)",
     "outcome": "nothing"
    },
    {
     "query": "fetch https://www.neoteo.com/en/google-search-referrals-in-chartbeats-publisher-network-fell-402",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "2026-10-03, 40.2% and 21.9%, pageview shares; no primary link",
     "outcome": "witness A-attribution-W05 (secondary)"
    },
    {
     "query": "\"40.2%\" Chartbeat Google search referrals Press Gazette",
     "mode": "standard",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "SEJ article citing the Chartbeat playbook",
     "outcome": "witness A-attribution-W05"
    },
    {
     "query": "fetch https://www.searchenginejournal.com/google-search-decline-publishers-chartbeat/590681/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1",
      "PRED-1091.B1"
     ],
     "result": "SEJ 2026-09-24, primary source named, period July 2025 to July 2026",
     "outcome": "witness A-attribution-W05"
    },
    {
     "query": "fetch https://lp.chartbeat.com/2026-publisher-playbook-chartbeat",
     "mode": "fetch",
     "aimed_at": [
      "PRED-659.B1"
     ],
     "result": "landing page with no figure; report gated",
     "outcome": "nothing (primary unread)"
    }
   ],
   "outcome": {
    "witnesses": [
     "W-0005",
     "W-0006",
     "W-0007",
     "W-0008",
     "W-0009",
     "W-0010",
     "W-0011",
     "W-0012",
     "W-0013",
     "W-0014",
     "W-0015",
     "W-0016",
     "W-0017",
     "W-0018",
     "W-0019",
     "W-0020",
     "W-0021",
     "W-0022"
    ],
    "candidates": [
     {
      "id": "C-0004",
      "source": "Auditing Google's AI Overviews and Featured Snippets: A Case Study on Baby Care and Pregnancy (arXiv 2511.12920)",
      "url": "https://arxiv.org/abs/2511.12920",
      "claims": [
       "PRED-187.F1",
       "PRED-1384.F1"
      ],
      "why_unresolved": "turned up by search and left unread"
     },
     {
      "id": "C-0005",
      "source": "Claim-Gated Source-Risk Auditing for Generative Search (arXiv 2609.29145)",
      "url": "https://arxiv.org/abs/2609.29145",
      "claims": [
       "PRED-716.F1"
      ],
      "why_unresolved": "turned up by search and left unread"
     },
     {
      "id": "C-0006",
      "source": "Auditing the Reliability of Multimodal Generative Search (arXiv 2604.00944)",
      "url": "https://arxiv.org/abs/2604.00944",
      "claims": [
       "PRED-1409.F12"
      ],
      "why_unresolved": "turned up by search and left unread"
     },
     {
      "id": "C-0007",
      "source": "What Gets Cited: Competitive GEO in AI Answer Engines (arXiv 2605.25517)",
      "url": "https://arxiv.org/html/2605.25517v1",
      "claims": [
       "PRED-768.A1",
       "PRED-742.F1"
      ],
      "why_unresolved": "turned up by search and left unread; may bear on what predicts which sources are cited"
     },
     {
      "id": "C-0008",
      "source": "Epistemic Substitution: How Grokipedia's AI-Generated Encyclopedia Restructures Authority (arXiv 2512.03337)",
      "url": "https://arxiv.org/abs/2512.03337",
      "claims": [
       "PRED-166.A1",
       "PRED-1643.B1"
      ],
      "why_unresolved": "turned up by search and left unread"
     },
     {
      "id": "C-0009",
      "source": "AlgorithmWatch study PDF (code 260908) behind the 2026-09-01 page",
      "url": "https://algorithmwatch.org/de/google-ki-ubersichten-meinungsbildung/",
      "claims": [
       "PRED-209.F2",
       "PRED-187.F1"
      ],
      "why_unresolved": "the figures in The Decoder (39.1% against 65.3%; AfD 24% against 45–58%) are not on the page read; the PDF would confirm them"
     }
    ],
    "notes": "C-0001 resolved as witness A-attribution-W01: the SSRC working paper is the arXiv 2508.00838 preprint by Strauss, Yang, O'Reilly, Rosenblat and Moure, dated 2025-06-27 on arXiv and on its Rogue Scholar DOI record, later published in Data & Policy 8, e15 (online 2026-04-28, DOI 10.1017/dap.2026.10064). It precedes every claim it bears on and is an anticipation. C-0002 resolved as witness A-attribution-W02: arXiv 2510.04226 v1, 2025-10-05, by Wright, Masud, Moore, Yadav, Antoniak, Christensen, Park and Augenstein. It measures epistemic diversity and has no bearing on attribution; it is recorded against two B-collapse claims (PRED-199.A1, as a mechanism-level disconfirmation in another domain; PRED-857.F1, as the method carried out). The main session should check that placement against the B-collapse reader's file to avoid a duplicate. Class-level exclusion: the query-resolution forecasts (PRED-140.A1, PRED-141.A1 to A5, PRED-768.A1 to A5) forecast what a query will surface about the archive's own terms. Outside research on source bias bears only on their premise, so none was attached; their test is the archive's own capture. The same holds for the claims about the archive's own entities and deposits (Lee Sharks, Mary Lee, Pearl, the MacArthur bios, the Dodecad and similar), which no outside study can witness. A-attribution-W12 (BuzzStream) is the one disconfirmation in theme A: it breaks the first link of PRED-615.B3. The same post would support a component of PRED-1091.B3 (the platform's opt-out does not control use); it was not attached there because a witness carries one direction, and the main session may add a second link as its own witness if wanted. Several witnesses are industry or vendor studies (Ahrefs, BuzzStream, SE Ranking via SEJ, Chartbeat via SEJ); their class is recorded as post or news, and their vendor status is listed under unverified. Two primaries were refused or gated (bbc.com blocked; Chartbeat report gated; SE Ranking page empty); the witnesses rest on the copies named in checked. Quotes are as returned by WebFetch, which quotes in segments of under 125 characters; quotes the tool gave only in paraphrase were left out. No memory facts were used in this pass."
   }
  },
  {
   "id": "S-0004",
   "date": "2026-10-30",
   "by": "intake reader, theme B-collapse, in session (brief: data/witnesses/passes/2026-10-30/BRIEF.md)",
   "aimed_at": "the forecasts and unbuilt tests of theme B-collapse (claims list: data/witnesses/passes/2026-10-30/claims-B-collapse.json)",
   "method": "WebSearch and WebFetch only, the readers' own account per query; every quote then checked at source by the record that follows this pass's readers.",
   "queries": [
    {
     "query": "Artificial Hivemind open-ended homogeneity of language models arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-857.F1",
      "PRED-803.B2"
     ],
     "result": "arXiv 2510.22954 and NeurIPS listings",
     "outcome": "witness B-collapse-W01"
    },
    {
     "query": "model collapse fraction of real data mixing threshold synthetic data iterative retraining",
     "mode": "standard",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "Garg et al. 2509.22341, Gerstgrasser 2404.01413, Seddik, Dohmatob",
     "outcome": "witness B-collapse-W11; others followed up"
    },
    {
     "query": "AI-generated code solution diversity decline across model generations homogeneity vulnerabilities correlated",
     "mode": "standard",
     "aimed_at": [
      "PRED-199.A1",
      "PRED-199.A2"
     ],
     "result": "Lee et al. Findings EMNLP 2025; security comparisons; a Medium post",
     "outcome": "witness B-collapse-W17"
    },
    {
     "query": "https://arxiv.org/abs/2510.22954",
     "mode": "fetch",
     "aimed_at": [
      "PRED-857.F1"
     ],
     "result": "abstract and history, three fetches for verbatim pieces",
     "outcome": "witness B-collapse-W01"
    },
    {
     "query": "https://arxiv.org/abs/2404.01413",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "abstract, two fetches",
     "outcome": "witness B-collapse-W11"
    },
    {
     "query": "https://arxiv.org/abs/2509.22341",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "Garg, Bhattacharya, Sur: optimal real-data proportion tends to 1/golden ratio for min-norm interpolation, at least 1/2 for ridge",
     "outcome": "excluded: an optimal mixing ratio for test error in linear regression, a different quantity from the claim's collapse threshold α*; neither confirms nor contradicts the 0.05–0.15 range"
    },
    {
     "query": "Bertrand On the Stability of Iterative Retraining of Generative Models on their own Data proportion of real data",
     "mode": "standard",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "arXiv 2310.00429, ICLR 2024 poster",
     "outcome": "witness B-collapse-W12"
    },
    {
     "query": "model collapse rare tokens tail disappears minimum fraction real data 2026 arXiv language model",
     "mode": "standard",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "Shumailov v3 on alphaxiv, Seddik 2404.05090, blogs",
     "outcome": "witness B-collapse-W13"
    },
    {
     "query": "https://arxiv.org/abs/2404.05090",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W13"
    },
    {
     "query": "https://arxiv.org/abs/2310.00429v5",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W12"
    },
    {
     "query": "https://www.nature.com/articles/s41586-024-07566-y",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7",
      "PRED-635.F1"
     ],
     "result": "refused (client_error)",
     "outcome": "nothing: used arXiv copy"
    },
    {
     "query": "https://arxiv.org/abs/2305.17493",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7",
      "PRED-635.F1",
      "PRED-1573.F4"
     ],
     "result": "history and abstract, two fetches",
     "outcome": "witness B-collapse-W10"
    },
    {
     "query": "Kirk Understanding the Effects of RLHF on LLM Generalisation and Diversity output diversity reduced",
     "mode": "standard",
     "aimed_at": [
      "PRED-107.G1",
      "PRED-803.B2"
     ],
     "result": "arXiv 2310.06452, ICLR 2024 proceedings",
     "outcome": "witness B-collapse-W06"
    },
    {
     "query": "instruction-tuned versus base model lower perplexity entropy outputs base models creativity randomness aligned",
     "mode": "standard",
     "aimed_at": [
      "PRED-107.G1"
     ],
     "result": "West & Potts 2505.00047, Mohammadi 2406.05587, Yang et al. 2506.17871",
     "outcome": "witnesses B-collapse-W07, W08, W09"
    },
    {
     "query": "https://arxiv.org/abs/2506.17871",
     "mode": "fetch",
     "aimed_at": [
      "PRED-107.G1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W07"
    },
    {
     "query": "https://arxiv.org/abs/2310.06452v3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-107.G1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W06"
    },
    {
     "query": "Shaib Measuring AI slop in text arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-107.G3"
     ],
     "result": "arXiv 2509.19163",
     "outcome": "nothing: see fetch"
    },
    {
     "query": "https://arxiv.org/abs/1906.04043",
     "mode": "fetch",
     "aimed_at": [
      "PRED-110.A1"
     ],
     "result": "GLTR abstract: detection rate 54% to 72%, no surprisal-direction statement in the abstract",
     "outcome": "excluded: abstract does not state the probability-rank finding; body not read"
    },
    {
     "query": "https://arxiv.org/abs/2509.19163",
     "mode": "fetch",
     "aimed_at": [
      "PRED-107.G3"
     ],
     "result": "slop judgments correlate with coherence and relevance; no deviation or surprisal measure in the abstract",
     "outcome": "excluded: does not test token-level deviation of slop; a full-text check could still bear on G3"
    },
    {
     "query": "https://arxiv.org/abs/2310.05130",
     "mode": "fetch",
     "aimed_at": [
      "PRED-110.A1"
     ],
     "result": "Fast-DetectGPT abstract, two fetches",
     "outcome": "witness B-collapse-W15"
    },
    {
     "query": "Anthropic updates consumer terms train on chats opt out September 2025 five year retention",
     "mode": "standard",
     "aimed_at": [
      "PRED-856.B1"
     ],
     "result": "news coverage and the Anthropic post",
     "outcome": "witness B-collapse-W20"
    },
    {
     "query": "human-written text influenced by LLM vocabulary shift after ChatGPT study spoken language delve",
     "mode": "standard",
     "aimed_at": [
      "PRED-856.F1"
     ],
     "result": "Yakura et al. 2409.01754; 2508.00238",
     "outcome": "witness B-collapse-W21"
    },
    {
     "query": "https://www.anthropic.com/news/updates-to-our-consumer-terms",
     "mode": "fetch",
     "aimed_at": [
      "PRED-856.B1"
     ],
     "result": "announcement",
     "outcome": "witness B-collapse-W20"
    },
    {
     "query": "https://arxiv.org/abs/2409.01754",
     "mode": "fetch",
     "aimed_at": [
      "PRED-856.F1"
     ],
     "result": "abstract v4",
     "outcome": "witness B-collapse-W21"
    },
    {
     "query": "AI expands scientists' impact but contracts science's focus Hao Evans",
     "mode": "standard",
     "aimed_at": [
      "PRED-476.A3",
      "PRED-476.A2"
     ],
     "result": "arXiv 2412.07727, press",
     "outcome": "witness B-collapse-W23"
    },
    {
     "query": "hidden prompt injection in manuscripts targeting LLM reviewers study arXiv papers positive review",
     "mode": "standard",
     "aimed_at": [
      "PRED-741.B2"
     ],
     "result": "Lin 2507.06185; 2508.20863; 2509.10248",
     "outcome": "witness B-collapse-W33"
    },
    {
     "query": "https://arxiv.org/abs/2412.07727v4",
     "mode": "fetch",
     "aimed_at": [
      "PRED-476.A3"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W23"
    },
    {
     "query": "https://arxiv.org/abs/2507.06185",
     "mode": "fetch",
     "aimed_at": [
      "PRED-741.B2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W33"
    },
    {
     "query": "https://preview.aclanthology.org/new-sigs/2025.findings-emnlp.10/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1",
      "PRED-199.F1"
     ],
     "result": "landing page",
     "outcome": "witness B-collapse-W17"
    },
    {
     "query": "LLM code generation same vulnerabilities across models shared weakness correlated CWE monoculture empirical study",
     "mode": "standard",
     "aimed_at": [
      "PRED-199.A2"
     ],
     "result": "Kahn & Hastings 2607.20713; Kordonsky 2602.04894; others",
     "outcome": "witnesses B-collapse-W18, W19"
    },
    {
     "query": "https://arxiv.org/abs/2607.20713",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W19"
    },
    {
     "query": "https://arxiv.org/abs/2602.04894",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W18"
    },
    {
     "query": "arXiv Exploring the Algorithmic Diversity of Model-Generated Code Lee Chon Jang",
     "mode": "standard",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "arXiv 2503.00691",
     "outcome": "dated B-collapse-W17"
    },
    {
     "query": "https://arxiv.org/abs/2505.00047",
     "mode": "fetch",
     "aimed_at": [
      "PRED-107.G1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W08"
    },
    {
     "query": "https://arxiv.org/abs/2503.00691v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "history",
     "outcome": "dated B-collapse-W17"
    },
    {
     "query": "NoveltyBench evaluating language models for humanlike diversity arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-855.F1",
      "PRED-790.F1"
     ],
     "result": "arXiv 2504.05228",
     "outcome": "witness B-collapse-W29"
    },
    {
     "query": "Doshi Hauser generative AI enhances individual creativity reduces collective diversity Science Advances",
     "mode": "standard",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "Science Advances 2024 and its Dryad data",
     "outcome": "excluded: the dataset the claim proposes to reanalyze, already within the scope of W-0003's meta-analysis; it does not carry out the mediation-design reanalysis"
    },
    {
     "query": "https://arxiv.org/abs/2504.05228v4",
     "mode": "fetch",
     "aimed_at": [
      "PRED-855.F1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W29"
    },
    {
     "query": "Meincke Mollick prompting diverse ideas increasing AI idea variance persona chain-of-thought",
     "mode": "standard",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "arXiv 2402.01727, SSRN",
     "outcome": "witness B-collapse-W28"
    },
    {
     "query": "https://arxiv.org/abs/2402.01727",
     "mode": "fetch",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W28"
    },
    {
     "query": "LLM persona diversity human-AI co-creation collective diversity homogenization mitigated diverse personas experiment 2025 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "Wan & Kalman arXiv and journal record; 2602.20408",
     "outcome": "witness B-collapse-W27"
    },
    {
     "query": "https://arxiv.org/abs/2504.13868v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W27"
    },
    {
     "query": "https://cris.openu.ac.il/en/publications/diverse-ai-personas-can-mitigate-the-homogenization-effect-in-hum/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "journal record, May 2026",
     "outcome": "witness B-collapse-W27"
    },
    {
     "query": "We Have a Package for You package hallucinations code-generating LLMs repeated persistent arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-1287.E1"
     ],
     "result": "arXiv 2406.10279",
     "outcome": "witness B-collapse-W30"
    },
    {
     "query": "Poisoning attacks on LLMs require a near-constant number of poison samples arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-1287.E3",
      "PRED-21.B2"
     ],
     "result": "arXiv 2510.07192",
     "outcome": "witness B-collapse-W31"
    },
    {
     "query": "https://arxiv.org/abs/2406.10279v3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1287.E1"
     ],
     "result": "abstract and history",
     "outcome": "witness B-collapse-W30"
    },
    {
     "query": "https://arxiv.org/abs/2510.07192",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1287.E3"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W31"
    },
    {
     "query": "https://arxiv.org/html/2406.10279v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1287.E1"
     ],
     "result": "Section 5.3 repetition figure",
     "outcome": "witness B-collapse-W30"
    },
    {
     "query": "adversarial poetry jailbreak follow-up 2026 semantic density literary quality attack success rate Bisconti",
     "mode": "standard",
     "aimed_at": [
      "PRED-1251.B1",
      "PRED-1205.F2",
      "PRED-188.F1"
     ],
     "result": "Bisconti 2511.15304; Adversarial Tales 2601.08837",
     "outcome": "see fetches"
    },
    {
     "query": "https://arxiv.org/abs/2601.08837v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1251.B1"
     ],
     "result": "Adversarial Tales: 71.3% average ASR across 26 frontier models with cyberpunk narrative framing",
     "outcome": "candidate"
    },
    {
     "query": "https://arxiv.org/abs/2511.15304",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1251.B1",
      "PRED-188.F1"
     ],
     "result": "Bisconti et al.: hand-crafted poems 62%, meta-prompt conversions about 43%",
     "outcome": "excluded: the source the claims themselves analyze; not an independent witness"
    },
    {
     "query": "poetic jailbreak literary quality versus attack success correlation stylistic obfuscation metaphor density LLM safety 2026 arXiv",
     "mode": "extended",
     "aimed_at": [
      "PRED-1205.F2",
      "PRED-1251.B1"
     ],
     "result": "Sorokoletova et al. 2605.12128",
     "outcome": "witness B-collapse-W38"
    },
    {
     "query": "https://arxiv.org/abs/2605.12128",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1205.F2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W38"
    },
    {
     "query": "https://arxiv.org/abs/2406.05587",
     "mode": "fetch",
     "aimed_at": [
      "PRED-803.B2",
      "PRED-107.G1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W09"
    },
    {
     "query": "homogenization of academic writing after ChatGPT lexical diversity decline scientific abstracts study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.F3",
      "PRED-476.A2"
     ],
     "result": "Filimonovic et al. 2511.11687; THE news; theses",
     "outcome": "witness B-collapse-W24"
    },
    {
     "query": "https://arxiv.org/abs/2511.11687",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1540.F3",
      "PRED-476.A2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W24"
    },
    {
     "query": "https://arxiv.org/abs/2406.07016",
     "mode": "fetch",
     "aimed_at": [
      "PRED-476.A2"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W22"
    },
    {
     "query": "AI humanizer tools usage students study 2026 bypass AI detectors survey prevalence",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.B1",
      "PRED-1540.F2"
     ],
     "result": "NBC News, The Star, blogs",
     "outcome": "witness B-collapse-W34"
    },
    {
     "query": "self-consuming loop language model generations lexical diversity decline measured tail tokens model collapse experiment 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-782.A7",
      "PRED-1573.F4",
      "PRED-635.F1"
     ],
     "result": "Guo et al. via 2410.12341 listing; Losing our Tail (ACL 2026); Break Step 2609.11149; entropy-rate filtering 2610.01493",
     "outcome": "witness B-collapse-W14; Break Step candidate"
    },
    {
     "query": "https://arxiv.org/abs/2311.09807",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A7",
      "PRED-635.F1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W14"
    },
    {
     "query": "https://arxiv.org/abs/2609.11149",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1573.F4"
     ],
     "result": "Break Step: recursive training where shared sampling seeds collapse unique-4-gram fraction to 0.045–0.38 across seven checkpoints",
     "outcome": "candidate"
    },
    {
     "query": "Pew Research teens use ChatGPT for schoolwork share 2025 2026 survey",
     "mode": "standard",
     "aimed_at": [
      "PRED-124.B1"
     ],
     "result": "Pew 2025 and 2026 reports and coverage",
     "outcome": "witness B-collapse-W36"
    },
    {
     "query": "share of new web articles AI-generated Graphite study 2025 retrieval LLM-generated content web proportion",
     "mode": "standard",
     "aimed_at": [
      "PRED-923.B2"
     ],
     "result": "GPTZero/Graphite share-of-web study; Axios",
     "outcome": "see fetch"
    },
    {
     "query": "https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B1"
     ],
     "result": "report",
     "outcome": "witness B-collapse-W36"
    },
    {
     "query": "https://gptzero.me/news/ai-written-web/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.B2"
     ],
     "result": "news post on Graphite study: new articles about 50% AI-generated",
     "outcome": "excluded: shows the volume of AI text on the web; does not show retrieval systems treating it as evidence, which is what PRED-923.B2 states"
    },
    {
     "query": "https://arxiv.org/abs/2301.07597",
     "mode": "fetch",
     "aimed_at": [
      "PRED-110.A1",
      "PRED-110.F2"
     ],
     "result": "HC3 abstract (no perplexity statement)",
     "outcome": "see HTML fetch"
    },
    {
     "query": "https://arxiv.org/html/2301.07597",
     "mode": "fetch",
     "aimed_at": [
      "PRED-110.A1"
     ],
     "result": "Section 4.4 perplexity",
     "outcome": "witness B-collapse-W16"
    },
    {
     "query": "https://arxiv.org/abs/2501.19361",
     "mode": "fetch",
     "aimed_at": [
      "PRED-857.F1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W04"
    },
    {
     "query": "longitudinal convergence of LLM outputs across model releases over time homogeneity increasing successive versions study 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-857.F1",
      "PRED-803.B2",
      "PRED-107.G1"
     ],
     "result": "Patel et al. 2608.19437; Fortier et al. EMNLP 2026; Xylogiannopoulos 2603.12683; TiCS review",
     "outcome": "witnesses B-collapse-W02, W03, W05"
    },
    {
     "query": "https://arxiv.org/abs/2608.19437",
     "mode": "fetch",
     "aimed_at": [
      "PRED-857.F1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W02"
    },
    {
     "query": "https://aclanthology.org/2026.emnlp-main.1621/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-803.B2",
      "PRED-107.G1"
     ],
     "result": "landing page",
     "outcome": "witness B-collapse-W05"
    },
    {
     "query": "https://arxiv.org/abs/2608.11426",
     "mode": "fetch",
     "aimed_at": [
      "PRED-803.B2"
     ],
     "result": "history",
     "outcome": "dated B-collapse-W05"
    },
    {
     "query": "https://arxiv.org/html/2608.19437v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-857.F1"
     ],
     "result": "model count, limitations",
     "outcome": "witness B-collapse-W02"
    },
    {
     "query": "code generation diversity newer models less diverse solutions than older models pass@k diversity across model versions",
     "mode": "standard",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "CRUXEval; API drift; Rath et al. 2610.07405",
     "outcome": "see fetch"
    },
    {
     "query": "https://arxiv.org/abs/2610.07405",
     "mode": "fetch",
     "aimed_at": [
      "PRED-199.A1"
     ],
     "result": "post-training moves diversity measures while pass@k shows no consistent winner (GSM8K)",
     "outcome": "excluded: math reasoning, one model, post-training arms; does not measure code diversity across generations under synthetic saturation"
    },
    {
     "query": "https://arxiv.org/abs/2212.10511",
     "mode": "fetch",
     "aimed_at": [
      "PRED-782.A5"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W35"
    },
    {
     "query": "https://nbcnews.com/tech/internet/college-students-ai-cheating-detectors-humanizers-rcna253878",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1540.B1"
     ],
     "result": "article; date not in fetched text",
     "outcome": "witness B-collapse-W34"
    },
    {
     "query": "tracked 43 humanizers 33.9 million website visits NBC News",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.B1"
     ],
     "result": "nothing on the article's date",
     "outcome": "nothing"
    },
    {
     "query": "https://bioethics.com/archives/101362",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1540.B1"
     ],
     "result": "repost dated 2026-01-29",
     "outcome": "dated B-collapse-W34"
    },
    {
     "query": "https://arxiv.org/abs/2507.14805",
     "mode": "fetch",
     "aimed_at": [
      "PRED-21.B2",
      "PRED-872.F1"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W32"
    },
    {
     "query": "no evidence of model collapse newer LLMs trained on web after ChatGPT diversity not declining study",
     "mode": "extended",
     "aimed_at": [
      "PRED-782.A7",
      "PRED-857.F1",
      "PRED-923.B2"
     ],
     "result": "Graphite AI search collapse; knowledge-collapse study (EurekAlert); Xylogiannopoulos; CACM blog",
     "outcome": "witnesses B-collapse-W03, W37; no disconfirmation of collapse in deployed models found"
    },
    {
     "query": "LLM-generated text more diverse than human field study homogenization not found long-term creativity null result 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-476.A2",
      "PRED-779.F4"
     ],
     "result": "Kosch et al. CHI 2609.23090; Inoshita et al. 2603.21228; Do LLMs Always Tell The Same Stories 2606.17350",
     "outcome": "witnesses B-collapse-W25 (disconfirming), W26"
    },
    {
     "query": "https://arxiv.org/abs/2609.23090",
     "mode": "fetch",
     "aimed_at": [
      "PRED-476.A2",
      "PRED-1540.F3"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W25"
    },
    {
     "query": "https://arxiv.org/abs/2603.12683",
     "mode": "fetch",
     "aimed_at": [
      "PRED-857.F1"
     ],
     "result": "abstract and history",
     "outcome": "witness B-collapse-W03"
    },
    {
     "query": "https://graphite.io/five-percent/research/ai-search-collapse",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.B2",
      "PRED-1088.B1"
     ],
     "result": "research post, two fetches",
     "outcome": "witness B-collapse-W37"
    },
    {
     "query": "https://arxiv.org/abs/2603.21228",
     "mode": "fetch",
     "aimed_at": [
      "PRED-779.F4"
     ],
     "result": "abstract",
     "outcome": "witness B-collapse-W26"
    },
    {
     "query": "https://arxiv.org/abs/2608.22118",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.B2"
     ],
     "result": "history of the academic version",
     "outcome": "dated B-collapse-W37"
    },
    {
     "query": "Bartz v. Anthropic fair use training ruling pirated library not fair use June 2025 order",
     "mode": "standard",
     "aimed_at": [
      "PRED-1209.B9"
     ],
     "result": "law-firm summaries of the June 2025 order",
     "outcome": "excluded: what 'the fence' names in PRED-1209.B9 could not be fixed without reading the deposit; fit not honest without it"
    }
   ],
   "outcome": {
    "witnesses": [
     "W-0023",
     "W-0024",
     "W-0025",
     "W-0026",
     "W-0027",
     "W-0028",
     "W-0029",
     "W-0030",
     "W-0031",
     "W-0032",
     "W-0033",
     "W-0034",
     "W-0035",
     "W-0036",
     "W-0037",
     "W-0038",
     "W-0039",
     "W-0040",
     "W-0041",
     "W-0042",
     "W-0043",
     "W-0044",
     "W-0045",
     "W-0046",
     "W-0047",
     "W-0048",
     "W-0049",
     "W-0050",
     "W-0051",
     "W-0052",
     "W-0053",
     "W-0054",
     "W-0055",
     "W-0056",
     "W-0057",
     "W-0058",
     "W-0059",
     "W-0060"
    ],
    "candidates": [
     {
      "id": "C-0010",
      "source": "Bisconti et al., From Adversarial Poetry to Adversarial Tales: An Interpretability Research Agenda (arXiv 2601.08837, submitted 2025-12-16)",
      "url": "https://arxiv.org/abs/2601.08837v1",
      "claims": [
       "PRED-1251.B1",
       "PRED-1205.F2"
      ],
      "why_unresolved": "Narrative framing reaches 71.3% average ASR against 62% (hand-crafted poems) and about 43% (converted prose) in the earlier paper, which fits PRED-1251.B1's floor reading; the two figures come from different papers, model sets and harm batteries, so the comparison is uncontrolled. Needs the body to see whether the authors compare within one setup."
     },
     {
      "id": "C-0011",
      "source": "Liu & Han, Break Step: Recursive Training Resonates with Replayed Sampling Noise (arXiv 2609.11149, submitted 2026-09-10)",
      "url": "https://arxiv.org/abs/2609.11149",
      "claims": [
       "PRED-1573.F4",
       "PRED-635.F1"
      ],
      "why_unresolved": "Runs recursive training across checkpoints and tracks unique-4-gram fraction, which is close to the calibration PRED-1573.F4 calls for; the abstract centres on a sampling-seed artefact, and how far its control arm calibrates a collapse index was not read."
     },
     {
      "id": "C-0012",
      "source": "Losing our Tail, Again: (Un)Natural Selection and Multilingual LLMs (ACL 2026 long paper; arXiv 2507.03933)",
      "url": "https://aclanthology.org/2026.acl-long.532/",
      "claims": [
       "PRED-782.A7",
       "PRED-855.F1"
      ],
      "why_unresolved": "Title and listing bear on tail loss; not fetched, so whether it reports measurements or argues a position is unknown."
     },
     {
      "id": "C-0013",
      "source": "Do Large Language Models Always Tell The Same Stories? (arXiv 2606.17350)",
      "url": "https://www.alphaxiv.org/abs/2606.17350",
      "claims": [
       "PRED-857.F1",
       "PRED-803.B2"
      ],
      "why_unresolved": "Listed in a search on narrative homogeneity across models; not fetched."
     },
     {
      "id": "C-0014",
      "source": "Masrour, Emi & Spero 2025, humanizer-tool study by detection-vendor researchers",
      "url": "",
      "claims": [
       "PRED-1540.B1"
      ],
      "why_unresolved": "The claim itself cites it as the half already occurred; not searched as a witness, since it is the claim's own evidence."
     }
    ],
    "notes": "Scope: 166 claims in the theme; most are narrative, internal-protocol or archive-specific (Assembly Chorus, SPXI, RES, Glas function, Bead Count tiers) with no external analogue, and were not searched. Searched lines: cross-model and longitudinal convergence (857.F1, 803.B2), alignment and diversity (107.G1), recursive-training collapse and real-data thresholds (782.A7, 635.F1, 1573.F4), machine-text predictability (110.A1, 110.F2), code monoculture (199.A1, 199.A2, 199.F1), chat data (856.B1, 856.F1), academic discourse (476.A2, 476.A3, 1540.F3), mediation design (779.F4), diversity benchmarks (855.F1), hallucination regularity and seeding (1287.E1, 1287.E3, 21.B2), evaluator gaming and humanizers (741.B2, 1540.B1), entity tail (782.A5), adoption (124.B1), RAG self-retrieval (923.B2, 1088.B1), poetic jailbreaks (1205.F2, 1251.B1). Disconfirmations: B-collapse-W25 (Kosch et al., CHI writing: prose did not homogenize, vocabulary grew more varied) against 476.A2's academic-discourse component; it sits beside B-collapse-W22, W23 and W24, which support that component in biomedicine, natural science and Scopus-wide. B-collapse-W05 (Fortier, Chen & West, EMNLP 2026) locates output homogeneity in pretraining with alignment as amplifier, which tells against safety pressure as the origin named by 803.B2; it also bears on 107.G1, whose predicted direction (instruct below base) it leaves standing while contesting RLHF as the source, so it was not linked there; the main session may wish to add a 107.G1 link with that scope. On 855.F1: NoveltyBench (W29) predates the claim and is a benchmark of output diversity; the claim's sentence that no such benchmark exists in any deployed evaluation framework stands only on a narrow reading of 'deployed'. Recorded as method support; the operator may rule it a disconfirmation of that sentence. W-0003 (de Rooij & Biskjaer) may also bear on PRED-779.F4 as method in part: a reanalysis of existing AI-creativity studies with a moderator (task constraint); the claim's moderator is mediation design, so the fit is partial. It does not bear on 476.A3's population-level trajectory. Anticipations: most witnesses predate their claims (the collapse literature from 2023–2024, the alignment-diversity literature from 2023–2025). After-registration witnesses: W02 (857.F1, 803.B2), W05, W19 (199.A2), W25 (476.A2, 1540.F3), and W37's arXiv version. Evidence classes needing care: W37 is a company research post by Graphite's officers, with an arXiv version; W34 is news reporting a figure from an integrity-software founder whose method is not described; W20 is a vendor announcement. Quotes: WebFetch returns verbatim fragments of at most about 120 characters, so long abstracts are quoted in pieces; every quote was requested character for character from the fetched page. The summarizer's paraphrase was never used as a quote. Not resolved: PRED-782.A7 part (3), the critical α in 0.05–0.15, has threshold existence results (W12, W13) and no value in the claim's terms; Garg et al. (2509.22341) give an optimal ratio near 0.618 for linear interpolation, a different quantity, logged as excluded. PRED-199.A1's generational SSDI decline has no direct measurement in code; W03 and W02 show generational decline in text. PRED-107.G3 (human-rated slop and signed deviation): Shaib et al. relate slop to coherence and relevance and do not test deviation. Prose rule: two verbatim quotes (W23, W27) contain a phrase the prose rules bar outside quotation; they are left as found."
   }
  },
  {
   "id": "S-0005",
   "date": "2026-10-30",
   "by": "intake reader, theme C-classifiers, in session (brief: data/witnesses/passes/2026-10-30/BRIEF.md)",
   "aimed_at": "the forecasts and unbuilt tests of theme C-classifiers (claims list: data/witnesses/passes/2026-10-30/claims-C-classifiers.json)",
   "method": "WebSearch and WebFetch only, the readers' own account per query; every quote then checked at source by the record that follows this pass's readers.",
   "queries": [
    {
     "query": "https://arxiv.org/abs/2510.13154",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2",
      "PRED-314.B1"
     ],
     "result": "Current abstract, three versions 2025-10-15, 2026-06-18, 2026-10-09",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2510.13154 (verbatim quotes)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2",
      "PRED-314.B1"
     ],
     "result": "Short abstract quotes on refusal logits, suppression, safety tax",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2510.13154v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2"
     ],
     "result": "Returned the current title, in conflict with the v1 HTML",
     "outcome": "witness C-classifiers-W01 (conflict recorded)",
     "date": "2026-10-30"
    },
    {
     "query": "Zahraei Asgari 2025 MENA values LLM refusal logit survey MENAValues arXiv 2510.13154",
     "mode": "standard",
     "aimed_at": [
      "PRED-923.F2"
     ],
     "result": "arXiv PDF and v1 HTML, HF dataset, MENAValues NeurIPS 2025 shared task",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/html/2510.13154v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2",
      "PRED-314.B1"
     ],
     "result": "v1 title I Am Aligned, But With Whom?; Logit Leakage finding",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/html/2510.13154v1 (verbatim Logit Leakage)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2"
     ],
     "result": "Verbatim fragment and v1 authors (Zahraei, Asgari)",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2510.13154v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F2"
     ],
     "result": "v2 page returned the alignment veto title and abstract",
     "outcome": "witness C-classifiers-W01",
     "date": "2026-10-30"
    },
    {
     "query": "AI text detector false positive rate pre-ChatGPT human-written texts across detector versions over time",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.F1"
     ],
     "result": "Park et al. 2608.26710, admissions essays, vendor pages",
     "outcome": "candidate then excluded: see Park fetch",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2608.26710",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1540.F1"
     ],
     "result": "13 detectors, FPR 0 to 100% on human texts, 2018-2025 editing corpus",
     "outcome": "excluded: no comparison across detector versions, so the archival-drift test is not carried out",
     "date": "2026-10-30"
    },
    {
     "query": "longitudinal audit AI detector updates same human essays flagged rate changed Turnitin GPTZero versions study",
     "mode": "extended",
     "aimed_at": [
      "PRED-1540.F1"
     ],
     "result": "Vendor and humanizer blogs only",
     "outcome": "excluded: no method behind the claims",
     "date": "2026-10-30"
    },
    {
     "query": "arXiv AI-generated text detectors false positives historical human texts written before 2022 flagged increase newer detector",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.F1"
     ],
     "result": "Generative-search citation audit, AAN abstract, Authors Guild test news",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "AI detector applied to pre-2022 scientific abstracts false positive rate rising over time detector update study",
     "mode": "extended",
     "aimed_at": [
      "PRED-1540.F1"
     ],
     "result": "Karr et al. 2608.11256, oncology abstracts 2021-2023 study, others",
     "outcome": "candidate: Karr fetched",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2608.11256",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1540.F1",
      "PRED-161.B2"
     ],
     "result": "Single detector threshold, 2013-15 vs 2023-25 abstracts, humanizer evasion",
     "outcome": "witness C-classifiers-W09 for PRED-161.B2; excluded for PRED-1540.F1: one detector version, no drift series",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2608.11256 (verbatim)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-161.B2"
     ],
     "result": "Exact humanizer and baseline flag-rate sentences",
     "outcome": "witness C-classifiers-W09",
     "date": "2026-10-30"
    },
    {
     "query": "study human writers alter writing style to avoid being flagged by AI detectors increased lexical diversity burstiness",
     "mode": "standard",
     "aimed_at": [
      "PRED-1540.A1"
     ],
     "result": "Bypass guides and blogs",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "Berg de Lucena Rosenblatt Large Language Models Report Subjective Experience Under Self-Referential Processing deception features",
     "mode": "standard",
     "aimed_at": [
      "PRED-314.F1",
      "PRED-314.B1"
     ],
     "result": "arXiv 2510.24797 and coverage",
     "outcome": "candidate then excluded: see Berg fetches",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2510.24797v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-314.F1"
     ],
     "result": "v1 2025-10-27; deception-feature suppression raises experience claims",
     "outcome": "excluded: deposit 314 builds its claim on these October 2025 feature-steering experiments (section 3.1), so the paper is the claim's own premise and cannot witness it",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/html/2510.24797v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-314.F1"
     ],
     "result": "v2 2025-10-30; same steering raises TruthfulQA truthfulness",
     "outcome": "excluded: same reason as v1",
     "date": "2026-10-30"
    },
    {
     "query": "\"AI psychosis\" lawsuit complaint filed OpenAI ChatGPT delusions product liability",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A4"
     ],
     "result": "Irwin v. OpenAI case page and PDF, seven SMVLC suits coverage",
     "outcome": "witness C-classifiers-W11",
     "date": "2026-10-30"
    },
    {
     "query": "https://techjusticelaw.org/wp-content/uploads/2025/12/FINAL-J.-Irwin-AMENDED-OpenAI-Complaint.pdf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A4"
     ],
     "result": "Phrase in heading F and paragraph 151; five causes of action",
     "outcome": "witness C-classifiers-W11",
     "date": "2026-10-30"
    },
    {
     "query": "https://techjusticelaw.org/wp-content/uploads/2025/12/FINAL-J.-Irwin-AMENDED-OpenAI-Complaint.pdf (verbatim)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A4"
     ],
     "result": "Exact heading and clause; no filing stamp",
     "outcome": "witness C-classifiers-W11",
     "date": "2026-10-30"
    },
    {
     "query": "https://techjusticelaw.org/cases/irwin-v-openai-inc-openai-holdings-llc-and-samuel-altman",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A4"
     ],
     "result": "Complaint 11/6/2025, JCCP 5431 coordination 2026-02-03",
     "outcome": "witness C-classifiers-W11",
     "date": "2026-10-30"
    },
    {
     "query": "psychiatrists caution against term \"AI psychosis\" misleading recommend alternative terminology \"AI-associated\"",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A6",
      "PRED-947.A5",
      "PRED-947.A1"
     ],
     "result": "Nielsen and Osler, Au Yeung et al., SODP survey, Harvard Gazette, Rolling Stone",
     "outcome": "witness C-classifiers-W12; others excluded below",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2605.26858",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A1",
      "PRED-947.A6"
     ],
     "result": "Nielsen and Osler 2026-05-26, conceptual argument",
     "outcome": "excluded: conceptual essay with no data",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2608.23937",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A1"
     ],
     "result": "Au Yeung, Morrin et al. 2026-08-25, perspective weighing nosological status",
     "outcome": "excluded: viewpoint with no method; it uses AI-associated psychosis but proposes no specifier",
     "date": "2026-10-30"
    },
    {
     "query": "https://blog.jmir.org/the-society-of-digital-psychiatry-ai-psychosis-survey-results",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A1",
      "PRED-947.A6"
     ],
     "result": "Member survey, 95 responses, 61.7% misleading but concerning",
     "outcome": "witness C-classifiers-W12",
     "date": "2026-10-30"
    },
    {
     "query": "\"AI psychosis\" bibliometric PubMed publications growth 2025 2026 case reports review",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A2"
     ],
     "result": "Au Yeung et al., newsletters, unrelated bibliometrics",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "https://pubmed.ncbi.nlm.nih.gov/?term=%22AI+psychosis%22%5Btiab%5D",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A2"
     ],
     "result": "13 records: 4 in 2025, 9 in 2026 to date, all publication types",
     "outcome": "nothing: interim count logged; the claim's window runs to end-2027 and counts case reports only",
     "date": "2026-10-30"
    },
    {
     "query": "PubMed combined synonym query for AI, AI-induced, AI-associated, chatbot and ChatGPT psychosis",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A2",
      "PRED-947.A7"
     ],
     "result": "Returned the PubMed home page with no results",
     "outcome": "nothing: fetch did not resolve",
     "date": "2026-10-30"
    },
    {
     "query": "https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A5"
     ],
     "result": "2025-10-27 post defining emotional reliance and psychosis or mania taxonomies",
     "outcome": "witness C-classifiers-W13",
     "date": "2026-10-30"
    },
    {
     "query": "https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/ (prevalence)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A5"
     ],
     "result": "0.07% weekly users psychosis or mania; 0.15% emotional attachment",
     "outcome": "witness C-classifiers-W13",
     "date": "2026-10-30"
    },
    {
     "query": "\"ChatGPT psychosis\" versus \"AI psychosis\" term usage frequency analysis media Reddit study",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A7",
      "PRED-947.A3"
     ],
     "result": "News, Substacks, prediction markets",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "American Psychiatric Association DSM proposal \"AI-associated\" psychosis specifier 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A1"
     ],
     "result": "Au Yeung et al., prediction markets, clinic pages",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "style guidance avoid term \"AI psychosis\" journalists clinicians guidance document recommends \"AI-associated delusions\"",
     "mode": "standard",
     "aimed_at": [
      "PRED-947.A6"
     ],
     "result": "News explainers, NAM conversation, AP page",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "https://ds.svcs.associatedpress.com/ai-guidance-terms-added-to-ap-stylebook",
     "mode": "fetch",
     "aimed_at": [
      "PRED-947.A6"
     ],
     "result": "AP brand page with no stylebook entry",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "arXiv 2025 AI humanizer tools evade AI text detectors evaluation detection rate drops",
     "mode": "standard",
     "aimed_at": [
      "PRED-161.B2"
     ],
     "result": "DAMAGE 2501.03437, AuthorMist 2503.08716, BFI working paper",
     "outcome": "witness C-classifiers-W08",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2501.03437v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-161.B2"
     ],
     "result": "2025-01-06; detectors fail on humanized text; augmentation retraining",
     "outcome": "witness C-classifiers-W08",
     "date": "2026-10-30"
    },
    {
     "query": "Schilke Reimann \"transparency dilemma\" AI disclosure erodes trust Organizational Behavior and Human Decision Processes",
     "mode": "standard",
     "aimed_at": [
      "PRED-195.B1"
     ],
     "result": "RePEc, Arizona portal, NSF PAR, news",
     "outcome": "witness C-classifiers-W06",
     "date": "2026-10-30"
    },
    {
     "query": "https://ideas.repec.org/a/eee/jobhdp/v188y2025ics0749597825000172.html",
     "mode": "fetch",
     "aimed_at": [
      "PRED-195.B1"
     ],
     "result": "OBHDP 188, DOI 10.1016/j.obhdp.2025.104405, thirteen experiments",
     "outcome": "witness C-classifiers-W06",
     "date": "2026-10-30"
    },
    {
     "query": "https://experts.arizona.edu/en/publications/the-transparency-dilemma-how-ai-disclosure-erodes-trust/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-195.B1"
     ],
     "result": "Volume 188 May 2025; verbatim abstract sentences",
     "outcome": "witness C-classifiers-W06",
     "date": "2026-10-30"
    },
    {
     "query": "LLM refusal rates differ by request framing descriptive explanation versus performing task holding harm constant over-refusal study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1537.F1",
      "PRED-1537.F3"
     ],
     "result": "Maskey et al. 2603.27518, AOR-Bench, RefusalBench, others",
     "outcome": "witness C-classifiers-W03",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2603.27518",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1537.F1",
      "PRED-923.F2"
     ],
     "result": "v1 2026-03-29, v4 2026-09-01; task-dependent over-refusal directions",
     "outcome": "witness C-classifiers-W03",
     "date": "2026-10-30"
    },
    {
     "query": "XSTest exaggerated safety behaviours large language models Röttger 2023 lexical overfitting refusal",
     "mode": "standard",
     "aimed_at": [
      "PRED-1537.F1"
     ],
     "result": "arXiv 2308.01263 and ACL Anthology",
     "outcome": "witness C-classifiers-W02",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2308.01263",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1537.F1"
     ],
     "result": "v1 2023-08-02, v3 2024-04-01; safe prompts refused on lexical resemblance",
     "outcome": "witness C-classifiers-W02",
     "date": "2026-10-30"
    },
    {
     "query": "https://aclanthology.org/2024.naacl-long.301/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1537.F1"
     ],
     "result": "NAACL 2024 pp. 5377-5400, DOI confirmed",
     "outcome": "witness C-classifiers-W02",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2406.14598",
     "mode": "fetch",
     "aimed_at": [
      "PRED-923.F1"
     ],
     "result": "SORRY-Bench, 44-topic taxonomy, over 50 models",
     "outcome": "witness C-classifiers-W04",
     "date": "2026-10-30"
    },
    {
     "query": "LLM peer review prestige bias author affiliation identical paper famous institution higher scores experiment arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-770.F1"
     ],
     "result": "Howell et al. 2509.15122, Justice in Judgment 2509.13400, 2502.00070",
     "outcome": "witness C-classifiers-W05",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2509.15122v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-770.F1"
     ],
     "result": "2025-09-18; resume-style audit, prestige dominant cue",
     "outcome": "witness C-classifiers-W05",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2509.15122v1 (verbatim)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-770.F1"
     ],
     "result": "Exact abstract fragments",
     "outcome": "witness C-classifiers-W05",
     "date": "2026-10-30"
    },
    {
     "query": "CMS AXOL1TL anomaly detection trigger model-independent search 2026 paper results",
     "mode": "standard",
     "aimed_at": [
      "PRED-1449.B1",
      "PRED-934.B2"
     ],
     "result": "Review papers, APS 2026 talk, BOOST 2026 talk, Indico meetings",
     "outcome": "candidate: see fetches",
     "date": "2026-10-30"
    },
    {
     "query": "https://indico.cern.ch/event/1574970/contributions/7103181/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1449.B1",
      "PRED-934.B2"
     ],
     "result": "BOOST 2026 talk abstract with no results stated",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "CMS collaboration arXiv 2026 anomaly detection Level-1 trigger triggered data search new physics no significant excess",
     "mode": "extended",
     "aimed_at": [
      "PRED-1449.B1",
      "PRED-934.B2"
     ],
     "result": "Vami 2609.31739 on behalf of CMS, distillation paper 2609.15295",
     "outcome": "candidate",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2609.31739",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1449.B1",
      "PRED-934.B2"
     ],
     "result": "Conference proceedings 2026-09-22; AXOL1TL and CICADA selected over four billion events in 2024; abstract states no result",
     "outcome": "candidate: full text not read for withheld-class metrology or result wording",
     "date": "2026-10-30"
    },
    {
     "query": "LLM watermarking reduces output diversity Self-BLEU distinct-n SynthID-Text Kirchenbauer evaluation quality diversity trade-off",
     "mode": "standard",
     "aimed_at": [
      "PRED-1555.F1"
     ],
     "result": "Surveys, SoK, SLAM, non-distortionary watermarking",
     "outcome": "nothing directly; led to the key-collision search",
     "date": "2026-10-30"
    },
    {
     "query": "\"Distortion-free watermarks are not truly distortion-free under watermark key collisions\"",
     "mode": "standard",
     "aimed_at": [
      "PRED-1555.F1"
     ],
     "result": "arXiv 2406.02603 and related collision papers",
     "outcome": "witness C-classifiers-W07",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2406.02603",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1555.F1"
     ],
     "result": "2024-06-02; bias under key collisions",
     "outcome": "witness C-classifiers-W07",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/html/2406.02603",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1555.F1"
     ],
     "result": "Weak vs strong distortion-free; identical outputs for repeated prompts",
     "outcome": "witness C-classifiers-W07",
     "date": "2026-10-30"
    },
    {
     "query": "binary AI text detectors fail on human-AI collaborative mixed authorship text fine-grained provenance labels study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1556.B1"
     ],
     "result": "HACo-Det, LLM-as-a-Coauthor, fine-grained detection",
     "outcome": "witness C-classifiers-W10",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2401.05952",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1556.B1"
     ],
     "result": "2024-01-11; detectors struggle with mixtext",
     "outcome": "witness C-classifiers-W10",
     "date": "2026-10-30"
    },
    {
     "query": "Ithaka S+R generative AI licensing agreement tracker publishers archives training data deals",
     "mode": "standard",
     "aimed_at": [
      "PRED-161.B1"
     ],
     "result": "Ithaka tracker, Nature news coverage, Wiley and FT deals",
     "outcome": "candidate: see fetches",
     "date": "2026-10-30"
    },
    {
     "query": "https://sr.ithaka.org/our-work/generative-ai-licensing-agreement-tracker/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-161.B1"
     ],
     "result": "Tracker description only; figures live in a linked sheet",
     "outcome": "nothing",
     "date": "2026-10-30"
    },
    {
     "query": "https://www.nature.com/articles/d41586-024-04018-5",
     "mode": "fetch",
     "aimed_at": [
      "PRED-161.B1"
     ],
     "result": "Refused by the site (client error)",
     "outcome": "nothing: refusal recorded",
     "date": "2026-10-30"
    },
    {
     "query": "Wiley AI content licensing revenue $44 million fiscal 2024 earnings release",
     "mode": "standard",
     "aimed_at": [
      "PRED-161.B1"
     ],
     "result": "Books+Publishing, Infodocket, Bookseller, earnings call coverage",
     "outcome": "witness C-classifiers-W16",
     "date": "2026-10-30"
    },
    {
     "query": "https://www.booksandpublishing.com.au/articles/2024/09/04/258068/wiley-expects-to-make-us44-million-from-ai-partnership-authors-unable-to-opt-out/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-161.B1"
     ],
     "result": "US$44 million expected from AI partnerships, citing The Bookseller",
     "outcome": "witness C-classifiers-W16",
     "date": "2026-10-30"
    },
    {
     "query": "benchmark items frequency distribution pretraining corpus head-weighted MMLU GSM8K term frequency analysis infini-gram benchmark",
     "mode": "standard",
     "aimed_at": [
      "PRED-1556.F2"
     ],
     "result": "Pythia, infini-gram mini, Chung and Kim 2602.10657",
     "outcome": "witness C-classifiers-W15",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2602.10657",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1556.F2"
     ],
     "result": "2026-02-11; word overlap with pre-training predicts benchmark scores",
     "outcome": "witness C-classifiers-W15",
     "date": "2026-10-30"
    },
    {
     "query": "https://arxiv.org/abs/2509.02534",
     "mode": "fetch",
     "aimed_at": [
      "PRED-855.B1"
     ],
     "result": "DARLING, 2025-09-02; diversity-rewarding RL improves quality and novelty",
     "outcome": "witness C-classifiers-W14",
     "date": "2026-10-30"
    }
   ],
   "outcome": {
    "witnesses": [
     "W-0061",
     "W-0062",
     "W-0063",
     "W-0064",
     "W-0065",
     "W-0066",
     "W-0067",
     "W-0068",
     "W-0069",
     "W-0070",
     "W-0071",
     "W-0072",
     "W-0073"
    ],
    "candidates": [
     {
      "id": "C-0015",
      "source": "Prestige over merit: An adapted audit of LLM bias in peer review",
      "url": "https://arxiv.org/abs/2509.15122v1",
      "claims": [
       "PRED-770.F1"
      ],
      "reader_id": "C-classifiers-W05",
      "state": "no quote found at source by verify_quotes.py"
     },
     {
      "id": "C-0016",
      "source": "Strengthening ChatGPT's responses in sensitive conversations",
      "url": "https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/",
      "claims": [
       "PRED-947.A5"
      ],
      "reader_id": "C-classifiers-W13",
      "state": "no quote found at source by verify_quotes.py"
     },
     {
      "id": "C-0017",
      "source": "Benchmarks Are Not That Out of Distribution: Word Overlap Predicts Performance",
      "url": "https://arxiv.org/abs/2602.10657",
      "claims": [
       "PRED-1556.F2"
      ],
      "reader_id": "C-classifiers-W15",
      "state": "no quote found at source by verify_quotes.py"
     },
     {
      "id": "C-0018",
      "source": "Vami, on behalf of the CMS Collaboration, Model-independent searches and anomaly detection at the CMS experiment (arXiv:2609.31739)",
      "url": "https://arxiv.org/abs/2609.31739",
      "claims": [
       "PRED-1449.B1",
       "PRED-934.B2"
      ],
      "why_unresolved": "Only the abstract was read. Whether the CMS AXOL1TL and CICADA results report withheld-class assimilation or retention metrology, and how any excess is worded, needs the full text and the DPF-2026 collaboration papers themselves."
     },
     {
      "id": "C-0019",
      "source": "Other complaints filed 2025-11-06 by the Social Media Victims Law Center and Tech Justice Law Project against OpenAI (six besides Irwin)",
      "url": "https://techjusticelaw.org/cases/fox-v-openai-inc-openai-holdings-llc-and-samuel-altman/",
      "claims": [
       "PRED-947.A4"
      ],
      "why_unresolved": "Not opened; they may also plead the phrase, which would add instances to W11."
     },
     {
      "id": "C-0020",
      "source": "PubMed count for \"AI psychosis\"[tiab]",
      "url": "https://pubmed.ncbi.nlm.nih.gov/?term=%22AI+psychosis%22%5Btiab%5D",
      "claims": [
       "PRED-947.A2"
      ],
      "why_unresolved": "Interim system-output count on 2026-10-30: 4 records in 2025 and 9 in 2026 to date, all publication types. The claim counts case reports and series in title or abstract and resolves at end-2027; a case-report filter and a Scopus count were not obtained."
     },
     {
      "id": "C-0021",
      "source": "Nature news, Publishers are selling papers to train AIs — and making millions of dollars (December 2024)",
      "url": "https://www.nature.com/articles/d41586-024-04018-5",
      "claims": [
       "PRED-161.B1"
      ],
      "why_unresolved": "The site refused the fetch; another copy was not found within budget."
     }
    ],
    "notes": "Held candidate C-0003 is resolved as witness C-classifiers-W01: the October 2025 predecessor is v1 (2025-10-15) of the same arXiv entry 2510.13154, titled I Am Aligned, But With Whom?, two authors, already reporting Logit Leakage (refusal while internal probabilities hold preferences). The alignment veto title, the four-author list and the suppression-not-erasure framing are in the current version and were returned for v2 (2026-06-18). Both versions precede PRED-923.F2 (2026-06-26) and PRED-314.B1 (2026-01-16), so both links are anticipations. Berg, de Lucena and Rosenblatt (arXiv 2510.24797) fit PRED-314.F1 closely but are excluded as a witness: deposit 314 builds the claim on those October 2025 feature-steering experiments, so they are its premise. W11 matters for PRED-947.A4: a product-liability complaint dated 2025-11-06 already pleads the phrase AI psychosis in a heading and in paragraph 151, eight months before the claim was registered on 2026-07-03. Whether a phrase in the factual allegations counts as a pleaded element in the claim's sense is for the operator to rule. W14 is the one disconfirmation. It meets the literal falsifier of PRED-855.B1 (any force that rewards diversity), as research only; deployment is unverified. DAMAGE (W08) is cited by the archive in deposit 1540, though not in deposit 161, which carries PRED-161.B2. No study was found that runs successive detector versions against a fixed pre-2022 human corpus (PRED-1540.F1); the archival-drift test appears unrun outside the archive. Claims with no outside evidence found or not searched for lack of externally testable content: the PRED-1616 panel items, PRED-1409, PRED-1039, PRED-1045, PRED-106, the collider batteries PRED-931 to PRED-1456 beyond the CMS search, PRED-197, PRED-947.A3, A7 and A8, and the forecasts PRED-368.B1, PRED-119.B1, PRED-741, PRED-784, PRED-158.B1, PRED-1540.A1, B2 and B3. WebFetch returns quotations through a summarizing model with a quotation limit; quotes were requested character for character, and where punctuation was uncertain (curly or straight quote marks) only an unpunctuated fragment was kept."
   }
  },
  {
   "id": "S-0006",
   "date": "2026-10-30",
   "by": "intake reader, theme D-economy, in session (brief: data/witnesses/passes/2026-10-30/BRIEF.md)",
   "aimed_at": "the forecasts and unbuilt tests of theme D-economy (claims list: data/witnesses/passes/2026-10-30/claims-D-economy.json)",
   "method": "WebSearch and WebFetch only, the readers' own account per query; every quote then checked at source by the record that follows this pass's readers.",
   "queries": [
    {
     "query": "Adobe Analytics AI traffic retail share of revenue 2026 generative AI referrals conversion",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A1",
      "PRED-1632.A6"
     ],
     "result": "Adobe AI Traffic Trends reports (June, August 2026), TechCrunch, Digital Commerce 360",
     "outcome": "candidate"
    },
    {
     "query": "https://business.adobe.com/resources/sdk/.q3-ai-traffic-trends-report/q3-2026-ai-sourced-traffic-insights.pdf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1",
      "PRED-1632.A6"
     ],
     "result": "Adobe 'Q3' report (June 2026, data to May): RPV +53 percent, conversion +54 percent; no AI share of visits or revenue; Cosmetics to Furniture & Home 16-point readability gap",
     "outcome": "candidate: baseline before registration, no dollar share"
    },
    {
     "query": "https://business.adobe.com/assets/pdfs/resources/sdk/ai-traffic-trends-report-august-2026/ai-traffic-trends-report-2026-08-19.pdf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1",
      "PRED-1632.A6"
     ],
     "result": "Adobe August 2026 report: RPV +53 percent; 17-point gap Apparel to Grocery; no AI revenue share",
     "outcome": "candidate: baseline before registration"
    },
    {
     "query": "https://techcrunch.com/2026/04/16/ai-traffic-to-us-retailers-rose-393-in-q1-and-its-boosting-their-revenue-too/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "Q1 2026 AI traffic growth and RPV +37 percent; no share of traffic",
     "outcome": "excluded: growth and relative value only, no share"
    },
    {
     "query": "Adobe AI traffic trends report October 2026 retail sub-industry gap citation readability",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A6"
     ],
     "result": "same Adobe reports, SEJ and Digital Commerce 360 coverage",
     "outcome": "nothing new"
    },
    {
     "query": "AI referrals percent of total ecommerce traffic 2026 \"1%\" ChatGPT share of retail sessions revenue study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "secondary roundups; Digital Commerce 360 retailer interviews",
     "outcome": "candidate"
    },
    {
     "query": "Shopify Q2 2026 earnings AI-referred orders share of orders AI agents",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "Shopify reports AI-referred traffic and orders tripled; no share found in snippets",
     "outcome": "nothing: growth multiples only"
    },
    {
     "query": "https://digitalcommerce360.com/2026/09/16/how-ai-referrals-impact-traffic-4-online-retailers",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "self-reported AI referral shares of four retailers (2 to 10 percent of referrals)",
     "outcome": "excluded: interviews of four retailers, referral share of traffic, no revenue share"
    },
    {
     "query": "Adobe Digital Insights AI traffic September 2026 report retail",
     "mode": "extended",
     "aimed_at": [
      "PRED-1632.A1",
      "PRED-1632.A6"
     ],
     "result": "Adobe 2026 holiday forecast (September 2026) and blog of 2026-10-29",
     "outcome": "candidate"
    },
    {
     "query": "https://business.adobe.com/assets/resources/holiday-shopping-report/2026/pdf/2026-adobe-digital-insights-holiday-forecast.pdf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "holiday online spend forecast $275.1B Nov-Dec 2026; AI traffic +130 percent YoY forecast; no AI share of revenue",
     "outcome": "candidate: denominator for the check on 2027-03-31"
    },
    {
     "query": "https://business.adobe.com/blog/ai-traffic-higher-revenue-per-visit-retail-holiday",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1",
      "PRED-1632.A6"
     ],
     "result": "2026-10-29: September AI RPV +83 percent, conversion +90 percent; readability Electronics overall 79 percent, General Merchandise homepages 40 percent, Grocery 55 percent",
     "outcome": "candidate: readability figures on mixed page bases, spread not computable on one basis"
    },
    {
     "query": "https://business.adobe.com/blog/ai-traffic-higher-revenue-per-visit-retail-holiday",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A6"
     ],
     "result": "second read for verbatim category passage; the page states no point gap",
     "outcome": "candidate"
    },
    {
     "query": "\"Citation Readability\" Electronics 79% Adobe retail categories",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A6"
     ],
     "result": "Adobe 'not machine readable' blog, Stellagent summaries",
     "outcome": "nothing new"
    },
    {
     "query": "https://searchengineland.com/ai-traffic-drives-more-revenue-per-visit-adobe-499713",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A6"
     ],
     "result": "refused by the site (client error)",
     "outcome": "nothing: fetch refused"
    },
    {
     "query": "https://business.adobe.com/au/blog/ai-traffic-surge-retail-sites-not-machine-readable",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A6"
     ],
     "result": "2026-04-16: homepage readability best 82.5 percent, lowest 54.2 percent",
     "outcome": "excluded: before registration, a different spread basis (sites, homepages)"
    },
    {
     "query": "Deloitte refund Australian government report AI hallucinated citations",
     "mode": "standard",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "wide press coverage, OECD incident entry",
     "outcome": "witness D-economy-W03"
    },
    {
     "query": "https://oecd.ai/en/incidents/2025-10-05-be45",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "OECD AI Incidents Monitor entry",
     "outcome": "witness D-economy-W03"
    },
    {
     "query": "2026 government report fabricated AI citations withdrawn policy hallucinated references",
     "mode": "extended",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "South Africa draft national AI policy withdrawn April 2026",
     "outcome": "witness D-economy-W02"
    },
    {
     "query": "MAHA report nonexistent studies citations NOTUS",
     "mode": "standard",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "press coverage of the May 2025 MAHA report",
     "outcome": "candidate"
    },
    {
     "query": "https://www.cnbcafrica.com/2026/south-africa-pulls-ai-policy-after-hallucinated-citations-expose-drafting-scandal",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "CNBC Africa 2026-04-27",
     "outcome": "witness D-economy-W02"
    },
    {
     "query": "https://thenextweb.com/news/south-africa-ai-policy-hallucinated-citations",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "The Next Web 2026-04-28",
     "outcome": "witness D-economy-W02"
    },
    {
     "query": "https://www.cnbcafrica.com/2026/south-africa-pulls-ai-policy-after-hallucinated-citations-expose-drafting-scandal",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B4"
     ],
     "result": "second read for verbatim sentences",
     "outcome": "witness D-economy-W02"
    },
    {
     "query": "TollBit State of the Bots report AI bots bypassing robots.txt share",
     "mode": "standard",
     "aimed_at": [
      "PRED-835.B5",
      "PRED-45.A1"
     ],
     "result": "TollBit report pages and press",
     "outcome": "witness D-economy-W04"
    },
    {
     "query": "https://tollbit.com/bots/25q2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.B5"
     ],
     "result": "page returned headings only",
     "outcome": "nothing: body not served to the fetcher"
    },
    {
     "query": "https://tollbit.com/analytics",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.B5"
     ],
     "result": "product page, no figures",
     "outcome": "nothing"
    },
    {
     "query": "https://tollbit.com/state-of-the-bots/q4-2024",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.B5"
     ],
     "result": "page returned headings only",
     "outcome": "nothing: body not served to the fetcher"
    },
    {
     "query": "https://www.fipp.com/news/state-of-the-bots-fipp-and-tollbit-put-the-spotlight-on-ai-scraping/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.B5"
     ],
     "result": "FIPP 2025-07-08 with TollBit figures",
     "outcome": "witness D-economy-W04"
    },
    {
     "query": "https://arxiv.org/abs/2407.14933",
     "mode": "fetch",
     "aimed_at": [
      "PRED-45.A1",
      "PRED-45.B2"
     ],
     "result": "Consent in Crisis abstract",
     "outcome": "witness D-economy-W05"
    },
    {
     "query": "Google AI Mode query fan-out technique announcement May 2025 blog",
     "mode": "standard",
     "aimed_at": [
      "PRED-45.A1"
     ],
     "result": "SEO commentary and Google posts",
     "outcome": "witness D-economy-W07"
    },
    {
     "query": "https://blog.google/products/search/ai-mode-search/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-45.A1"
     ],
     "result": "Google post 2025-03-05",
     "outcome": "witness D-economy-W07"
    },
    {
     "query": "Pew Research Center users less likely to click links when AI summary appears Google study",
     "mode": "standard",
     "aimed_at": [
      "PRED-1111.B1"
     ],
     "result": "Pew short read 2025-07-22",
     "outcome": "witness D-economy-W08"
    },
    {
     "query": "https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1111.B1"
     ],
     "result": "Pew figures",
     "outcome": "witness D-economy-W08"
    },
    {
     "query": "2026 study AI Overviews click-through rate decline publishers traffic Google Discover zero-click Similarweb Chartbeat",
     "mode": "standard",
     "aimed_at": [
      "PRED-1111.B1"
     ],
     "result": "secondary roundups; Ahrefs, Chartbeat coverage",
     "outcome": "candidate"
    },
    {
     "query": "Ahrefs AI Overviews reduce clicks study position one CTR 2026 update",
     "mode": "standard",
     "aimed_at": [
      "PRED-1111.B1"
     ],
     "result": "Ahrefs update of February 2026",
     "outcome": "witness D-economy-W09"
    },
    {
     "query": "https://ahrefs.com/blog/ai-overviews-reduce-clicks-update",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1111.B1"
     ],
     "result": "Ahrefs update 2026-02-04",
     "outcome": "witness D-economy-W09"
    },
    {
     "query": "EU AI Act Article 50 transparency obligations August 2026 digital omnibus agreement delay high-risk 2027",
     "mode": "standard",
     "aimed_at": [
      "PRED-615.B1",
      "PRED-45.A2"
     ],
     "result": "law-firm alerts on Article 50 and the Omnibus",
     "outcome": "witness D-economy-W10"
    },
    {
     "query": "https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force",
     "mode": "fetch",
     "aimed_at": [
      "PRED-615.B1",
      "PRED-45.A2"
     ],
     "result": "Goodwin alert 2026-08-03: Article 50 in force; high-risk Annex III from 2 December 2027, Annex I from 2 August 2028",
     "outcome": "witness D-economy-W10; candidate for PRED-45.A2"
    },
    {
     "query": "OKF Open Knowledge Format Google specification v1.0 release github knowledge-catalog",
     "mode": "standard",
     "aimed_at": [
      "PRED-835.A2",
      "PRED-835.B5",
      "PRED-835.B2"
     ],
     "result": "Google Cloud announcement, press",
     "outcome": "candidate"
    },
    {
     "query": "https://cloud.google.com/blog/products/data-analytics/how-the-open-knowledge-format-can-improve-data-sharing",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.A2"
     ],
     "result": "OKF v0.1 announced 2026-06-12; frontmatter type, title, description, resource, tags, timestamp",
     "outcome": "candidate"
    },
    {
     "query": "https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.A2"
     ],
     "result": "refused (robots disallowed)",
     "outcome": "nothing: fetch refused"
    },
    {
     "query": "https://github.com/GoogleCloudPlatform/knowledge-catalog/releases",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.A2"
     ],
     "result": "no releases or tags",
     "outcome": "candidate"
    },
    {
     "query": "https://raw.githubusercontent.com/GoogleCloudPlatform/knowledge-catalog/main/okf/SPEC.md",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.A2",
      "PRED-835.B5"
     ],
     "result": "spec declares OKF version 0.2, adds sources, generated, verified (advisory trust tiers), status, stale_after; no summarization-policy field",
     "outcome": "candidate: v0.2 undated, v1.0 not shipped"
    },
    {
     "query": "https://raw.githubusercontent.com/GoogleCloudPlatform/knowledge-catalog/main/okf/SPEC.md",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.A2"
     ],
     "result": "second read: deferred items list concerns attestation runtime, ABI, caching, semantic-layer templates; no summarization governance listed",
     "outcome": "candidate"
    },
    {
     "query": "https://github.com/GoogleCloudPlatform/knowledge-catalog/issues?q=is%3Aissue+governance",
     "mode": "fetch",
     "aimed_at": [
      "PRED-835.B2",
      "PRED-835.A2"
     ],
     "result": "refused (robots disallowed)",
     "outcome": "nothing: fetch refused"
    },
    {
     "query": "OpenAI restores GPT-4o after GPT-5 backlash users attachment Altman August 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "press coverage of reinstatement",
     "outcome": "witness D-economy-W11"
    },
    {
     "query": "https://openai.com/index/sycophancy-in-gpt-4o/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "OpenAI 2025-04-29: update 'focused too much on short-term feedback'; rolled back",
     "outcome": "excluded: shows the feedback incentive and also a lab reversing it; direction indeterminate for the claim"
    },
    {
     "query": "https://www.techrepublic.com/article/news-openai-reinstates-gpt4o/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "TechRepublic 2025-08-11",
     "outcome": "witness D-economy-W11"
    },
    {
     "query": "OpenAI retire GPT-4o February 2026 ChatGPT announcement users",
     "mode": "standard",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "OpenAI retirement announcement and press",
     "outcome": "candidate"
    },
    {
     "query": "https://openai.com/index/retiring-gpt-4o-and-older-models",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "2026-01-29: 'only 0.1% of users still choosing GPT‑4o each day'",
     "outcome": "candidate"
    },
    {
     "query": "https://openai.com/index/how-we're-optimizing-chatgpt/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "break reminders; no date on page",
     "outcome": "witness D-economy-W12"
    },
    {
     "query": "\"What we're optimizing ChatGPT for\" OpenAI break reminders August 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "press of 2025-08-04",
     "outcome": "witness D-economy-W12"
    },
    {
     "query": "https://9to5mac.com/2025/08/04/healthy-chatgpt-use",
     "mode": "fetch",
     "aimed_at": [
      "PRED-124.B2"
     ],
     "result": "9to5Mac 2025-08-04 dates the OpenAI post",
     "outcome": "witness D-economy-W12"
    },
    {
     "query": "benchmark accuracy of AI answers about brands by industry factual accuracy brand information LLM study 2026 rubric",
     "mode": "extended",
     "aimed_at": [
      "PRED-1632.A7"
     ],
     "result": "Faro Index report; Research Square scoping review; vendor posts",
     "outcome": "witness D-economy-W01; candidate"
    },
    {
     "query": "https://faroindex.ai/report",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A7"
     ],
     "result": "Brand Accuracy Rate by 21 industries, July 2026",
     "outcome": "witness D-economy-W01"
    },
    {
     "query": "https://faroindex.ai/report",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A7"
     ],
     "result": "second read: edition line, definitions, section heading; no download offered",
     "outcome": "witness D-economy-W01"
    },
    {
     "query": "https://www.researchsquare.com/article/rs-11390685/v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A7"
     ],
     "result": "scoping review posted 2026-10-19; body reports few audits code what is said",
     "outcome": "candidate"
    },
    {
     "query": "McKinsey agentic commerce $1 trillion orchestrated revenue 2030 Bain $300 billion $500 billion agentic commerce forecast revised",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A3",
      "PRED-1632.A4",
      "PRED-1632.A2"
     ],
     "result": "original forecasts and secondary coverage; no revision or scope definition found",
     "outcome": "nothing"
    },
    {
     "query": "Similarweb AI citation rate professional services category AI Overviews ChatGPT citations industry 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A5"
     ],
     "result": "Similarweb sector citation coverage, vendor benchmarks",
     "outcome": "nothing: window ends 2027-12-31, no category series read"
    },
    {
     "query": "AI agent made unauthorized purchase liability refund who pays agentic commerce error case 2026",
     "mode": "standard",
     "aimed_at": [
      "PRED-1666.B2"
     ],
     "result": "commentary on a 'liability gap' for agent purchases",
     "outcome": "excluded: commentary without a decided case or data"
    },
    {
     "query": "https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1666.B2"
     ],
     "result": "Moffatt v. Air Canada; refused (robots disallowed)",
     "outcome": "nothing: fetch refused"
    },
    {
     "query": "DataCite tombstone page recommendation DOI content no longer available support",
     "mode": "standard",
     "aimed_at": [
      "PRED-1.F1"
     ],
     "result": "DataCite support page",
     "outcome": "excluded"
    },
    {
     "query": "https://support.datacite.org:443/docs/tombstone-pages",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1.F1"
     ],
     "result": "DataCite best practice for tombstone pages, page updated 2026-08-07",
     "outcome": "excluded: the claim's own cited premise; it does not test Zenodo's conformity; earliest version undated"
    },
    {
     "query": "OpenAIRE Graph deleted records source repository deletion propagation documentation deletedbyinference",
     "mode": "standard",
     "aimed_at": [
      "PRED-875.B1",
      "PRED-875.F1"
     ],
     "result": "OpenAIRE helpdesk retention page",
     "outcome": "excluded"
    },
    {
     "query": "https://helpdesk.openaire.eu/help/en-gb/44-data-governance/243-what-is-openaire-s-data-retention-policy-for-propagated-data",
     "mode": "fetch",
     "aimed_at": [
      "PRED-875.F1"
     ],
     "result": "2024-04-11: 'Data integrated into the OpenAIRE Graph is retained indefinitely'; silent on deletion at source",
     "outcome": "excluded: policy statement without observation, silent on records deleted at source"
    },
    {
     "query": "https://arxiv.org/abs/2311.09735",
     "mode": "fetch",
     "aimed_at": [
      "PRED-45.B5"
     ],
     "result": "GEO abstract",
     "outcome": "witness D-economy-W13"
    },
    {
     "query": "https://blog.cloudflare.com/content-independence-day-no-ai-crawl-without-compensation/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-45.B2",
      "PRED-45.A1"
     ],
     "result": "Cloudflare 2025-07-01",
     "outcome": "witness D-economy-W06"
    },
    {
     "query": "Bartz v. Anthropic $1.5 billion settlement approved per work authors judge Alsup",
     "mode": "standard",
     "aimed_at": [
      "PRED-1666.B4",
      "PRED-45.B3"
     ],
     "result": "final approval coverage",
     "outcome": "candidate"
    },
    {
     "query": "2026 AI data licensing deals publishers survey report revenue study licensing market training data compensation",
     "mode": "standard",
     "aimed_at": [
      "PRED-45.B2",
      "PRED-45.B3"
     ],
     "result": "licensing trackers and reports",
     "outcome": "witness D-economy-W14"
    },
    {
     "query": "https://thenextweb.com/news/anthropic-1-5-billion-copyright-settlement-approved?type=AI",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1666.B4",
      "PRED-45.B3"
     ],
     "result": "2026-07-21: final approval 20 July, about $3,000 per book, roughly half a million works",
     "outcome": "candidate"
    },
    {
     "query": "https://mediacopilot.ai/what-99-ai-licensing-deals-reveal-about-the-future-of-publishing",
     "mode": "fetch",
     "aimed_at": [
      "PRED-45.B2"
     ],
     "result": "2026-09-24 interview on 99 deals",
     "outcome": "witness D-economy-W14"
    },
    {
     "query": "AI platforms account for \"less than 1%\" of ecommerce traffic 2026 Similarweb OR Adobe OR Salesforce share of orders",
     "mode": "standard",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "secondary roundups and Digital Commerce 360",
     "outcome": "nothing"
    },
    {
     "query": "https://www.digitalcommerce360.com/2026/09/03/ecommerce-trends-ai-sources-referral-traffic-online-retailers/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1632.A1"
     ],
     "result": "retailer counts by top AI source; no shares",
     "outcome": "excluded: counts of retailers, no traffic or revenue share"
    }
   ],
   "outcome": {
    "witnesses": [
     "W-0074",
     "W-0075",
     "W-0076",
     "W-0077",
     "W-0078",
     "W-0079",
     "W-0080",
     "W-0007",
     "W-0081",
     "W-0082",
     "W-0083",
     "W-0084",
     "W-0085",
     "W-0086"
    ],
    "candidates": [
     {
      "id": "C-0022",
      "source": "Adobe Digital Insights, 'AI Traffic Drives 83% Higher Revenue Per Visit' (Vivek Pandya, 2026-10-29)",
      "url": "https://business.adobe.com/blog/ai-traffic-higher-revenue-per-visit-retail-holiday",
      "claims": [
       "PRED-1632.A6",
       "PRED-1632.A1"
      ],
      "why_unresolved": "The only post-registration Adobe readability figures give Electronics 79 percent overall, Grocery 55 percent and General Merchandise homepages 40 percent, on mixed page bases; the 'best-to-worst spread' of #1632 needs its basis read from #1632 itself before a comparison. The pre-registration baseline is a 16-point gap (June 2026 report) and a 17-point gap (August 2026 report) between retail sub-industries. For A1, Adobe now publishes relative revenue per visit (+53 percent in July, +83 percent in September 2026), which is the v term of A1 (v about 1.5 to 1.8), so A1's restatement clause is triggered; Adobe publishes no AI share of visits or revenue."
     },
     {
      "id": "C-0023",
      "source": "Adobe 2026 U.S. Holiday Forecast (September 2026)",
      "url": "https://business.adobe.com/assets/resources/holiday-shopping-report/2026/pdf/2026-adobe-digital-insights-holiday-forecast.pdf",
      "claims": [
       "PRED-1632.A1"
      ],
      "why_unresolved": "Supplies the Q4 denominator ($275.1B Nov-Dec, $370.9B Oct-Dec forecast) and AI traffic growth forecasts; no AI share. A1's check date is 2027-03-31; Adobe's post-season report is the likely source."
     },
     {
      "id": "C-0024",
      "source": "Google, Open Knowledge Format specification SPEC.md (declares version 0.2)",
      "url": "https://raw.githubusercontent.com/GoogleCloudPlatform/knowledge-catalog/main/okf/SPEC.md",
      "claims": [
       "PRED-835.A2",
       "PRED-835.B5",
       "PRED-835.B2"
      ],
      "why_unresolved": "v0.2 adds `sources` (derivation provenance) and `verified` (advisory trust tiers, 'not access control') and no summarization-policy field; its deferred list does not name summarization governance. A2's window runs to Q2 2027 and its falsifier concerns v1.0. Whether `sources` counts as a 'functional equivalent' of provenance_kernel is the operator's call. The v0.2 date could not be read: the GitHub tree, issues and commits are disallowed to the fetcher, and the releases page lists none."
     },
     {
      "id": "C-0025",
      "source": "Charbel Sami Ibrahim, 'Measuring brands in machine answers: A scoping review and validity framework for large language model brand-visibility audits', Research Square, posted 2026-10-19",
      "url": "https://www.researchsquare.com/article/rs-11390685/v1",
      "claims": [
       "PRED-1632.A7"
      ],
      "why_unresolved": "The body (summarized by the fetcher; no quote taken) reports that 8 of 38 studies code what is said about the brand; whether any reports accuracy by field with a rubric and open sample needs the full paper and those eight studies."
     },
     {
      "id": "C-0026",
      "source": "OpenAI, 'Retiring GPT-4o and older models' (2026-01-29)",
      "url": "https://openai.com/index/retiring-gpt-4o-and-older-models",
      "claims": [
       "PRED-124.B2"
      ],
      "why_unresolved": "Reports 'only 0.1% of users still choosing GPT‑4o each day' before retirement on 2026-02-13. It could bear against the 'would lose users' arm, but the post does not frame the change as engagement-reducing, and no churn data are given."
     },
     {
      "id": "C-0027",
      "source": "Bartz v. Anthropic settlement, final approval 2026-07-20 (The Next Web, 2026-07-21)",
      "url": "https://thenextweb.com/news/anthropic-1-5-billion-copyright-settlement-approved?type=AI",
      "claims": [
       "PRED-1666.B4",
       "PRED-45.B3"
      ],
      "why_unresolved": "A platform paying holders of works by settlement, about $3,000 per book. The case concerned shadow-library downloads; fit to B4's 'distinctions it needs are held by the parties that preserved them' is loose, and 45.B3 states no testable threshold."
     },
     {
      "id": "C-0028",
      "source": "MAHA report citations of nonexistent studies (May 2025, press)",
      "url": "https://www.ehn.org/white-house-quietly-updates-maha-report-that-included-non-existent-study-citations",
      "claims": [
       "PRED-615.B4"
      ],
      "why_unresolved": "A second anticipation of policy documents with fabricated citations; not fetched for lack of budget."
     },
     {
      "id": "C-0029",
      "source": "Goodwin alert on the Digital Omnibus dates",
      "url": "https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force",
      "claims": [
       "PRED-45.A2"
      ],
      "why_unresolved": "High-risk obligations now apply from 2 December 2027 (Annex III) and 2 August 2028 (Annex I), inside A2's 2027-2028 window for 'EU AI Act full enforcement'; resolution waits on the window."
     }
    ],
    "notes": "Of 150 claims in the theme, most are framings without an external test (windows closing, eschatological or retrocausal passages, internal build plans, archive-specific observations). Searching went to the dated forecasts of #1632, #45, #615, #835, #1111 and #124. The disconfirmation of most weight is D-economy-W01: a vendor benchmark (Faro Index, July 2026) published accuracy by industry, separated from mention, before PRED-1632.A7 (registered 2026-09-20) said none would before 2027-12-31. Its fit is limited by the absence of an open sample and by the difference between claim-match accuracy and entity-resolution accuracy; the operator rules whether it meets the falsifier 'to the extent that publication covers'. D-economy-W12 (OpenAI break reminders, August 2025) bears against the conclusion of PRED-124.B2; D-economy-W11 bears for its mechanism; both precede the claim. The strongest support after registration is D-economy-W02 (South Africa's draft AI policy withdrawn for fabricated citations, April 2026) for PRED-615.B4. No #1632 forecast can be settled yet. Adobe now publishes v (relative revenue per visit, +53 to +83 percent in 2026), which triggers A1's restatement clause; no source publishes AI-referred share of retail revenue. WebFetch returns model-processed text; quotes were requested verbatim but were not compared against raw HTML. Several sources refused the fetcher (GitHub tree and issues for OKF, CanLII, Search Engine Land), and TollBit's report pages served headings only, so W04 rests on FIPP's report of TollBit's figures. D-economy-W09's earliest version (Ahrefs April 2025) and D-economy-W10's legislative origin (2024 Regulation) both precede their claims; the computed order should be read with that in mind. Count: about 34 searches and 40 fetches. The memory file was not consulted for any fact in this output."
   }
  },
  {
   "id": "S-0007",
   "date": "2026-10-30",
   "by": "intake reader, theme E-execution, in session (brief: data/witnesses/passes/2026-10-30/BRIEF.md)",
   "aimed_at": "the forecasts and unbuilt tests of theme E-execution (claims list: data/witnesses/passes/2026-10-30/claims-E-execution.json)",
   "method": "WebSearch and WebFetch only, the readers' own account per query; every quote then checked at source by the record that follows this pass's readers.",
   "queries": [
    {
     "query": "InjecGuard over-defense prompt injection guard models NotInject benign trigger words",
     "mode": "standard",
     "aimed_at": [
      "PRED-139.F1",
      "PRED-1384.A1"
     ],
     "result": "arXiv 2410.22770 and review mirrors",
     "outcome": "witness E-execution-W01"
    },
    {
     "query": "prompt injection detector false positive rate benign instructions benchmark 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-139.F1",
      "PRED-1384.A1"
     ],
     "result": "arXiv 2610.03448; vendor leaderboards and blog benchmarks (bastionsoft, promptguard, a 2026-09-28 blog post)",
     "outcome": "witness E-execution-W02; vendor pages excluded: vendor marketing without disclosed method"
    },
    {
     "query": "A small number of samples can poison LLMs of any size 250 documents arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-580.F2",
      "PRED-127.B1"
     ],
     "result": "arXiv 2510.07192 and the Anthropic research page",
     "outcome": "witness E-execution-W05, E-execution-W06"
    },
    {
     "query": "Meeus copyright traps for large language models arXiv 2402.09363 sequence repetition detectability",
     "mode": "standard",
     "aimed_at": [
      "PRED-127.B1"
     ],
     "result": "arXiv 2402.09363, ICML 2024 pages",
     "outcome": "witness E-execution-W07"
    },
    {
     "query": "NIST AI 100-2 E2025 adversarial machine learning taxonomy indirect prompt injection definition",
     "mode": "standard",
     "aimed_at": [
      "PRED-194.A1",
      "PRED-1384.A1"
     ],
     "result": "NIST PDF and commentary",
     "outcome": "excluded: the definitions in sections 3.3-3.4 and the glossary could not be reached in the fetched text"
    },
    {
     "query": "OWASP Top 10 for LLM applications 2025 LLM01 prompt injection definition direct indirect external sources",
     "mode": "standard",
     "aimed_at": [
      "PRED-1384.A1",
      "PRED-194.A1"
     ],
     "result": "secondary guides to OWASP 2025",
     "outcome": "led to fetch of the official page; witness E-execution-W04"
    },
    {
     "query": "NIST AI 100-2 2026 edition adversarial machine learning draft released",
     "mode": "extended",
     "aimed_at": [
      "PRED-194.A1"
     ],
     "result": "only the E2025 edition (March 2025) and commentary",
     "outcome": "nothing: no later NIST AI 100-2 edition found; falsifier (c) not fired"
    },
    {
     "query": "Greshake \"Not what you've signed up for\" indirect prompt injection retrieved content arXiv 2302.12173",
     "mode": "standard",
     "aimed_at": [
      "PRED-622.B1",
      "PRED-42.B1"
     ],
     "result": "arXiv 2302.12173",
     "outcome": "witness E-execution-W12"
    },
    {
     "query": "PoisonedRAG knowledge corruption attacks retrieval-augmented generation five malicious texts 90% attack success",
     "mode": "standard",
     "aimed_at": [
      "PRED-42.B1",
      "PRED-622.B1"
     ],
     "result": "arXiv 2402.07867, USENIX Security 2025",
     "outcome": "witness E-execution-W11"
    },
    {
     "query": "FABLES faithfulness book-length summarization LLM claims unfaithful omissions arXiv Kim 2024",
     "mode": "standard",
     "aimed_at": [
      "PRED-950.F2",
      "PRED-515.F1",
      "PRED-29.F1",
      "PRED-34.F2"
     ],
     "result": "arXiv 2404.01261 and related faithfulness papers",
     "outcome": "witness E-execution-W15, E-execution-W16"
    },
    {
     "query": "Evaluating Verifiability in Generative Search Engines Liu Zhang Liang citation recall precision",
     "mode": "standard",
     "aimed_at": [
      "PRED-176.F2"
     ],
     "result": "arXiv 2304.09848, Findings of EMNLP 2023",
     "outcome": "witness E-execution-W18"
    },
    {
     "query": "Artificial Hivemind open-ended homogeneity of language models intra-model inter-model arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-284.A1"
     ],
     "result": "arXiv 2510.22954, NeurIPS 2025",
     "outcome": "witness E-execution-W24"
    },
    {
     "query": "\"What Evidence Do Language Models Find Convincing\" Wan Wallace Klein relevance stylistic features scientific references",
     "mode": "standard",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "arXiv 2402.11782",
     "outcome": "witness E-execution-W22"
    },
    {
     "query": "GEO Generative Engine Optimization Aggarwal arXiv 2311.09735 visibility 40%",
     "mode": "standard",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "arXiv 2311.09735",
     "outcome": "witness E-execution-W21"
    },
    {
     "query": "Idiosyncrasies in Large Language Models classifier distinguish model outputs accuracy Sun 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-960.F1"
     ],
     "result": "arXiv 2502.12150, ICML 2025",
     "outcome": "witness E-execution-W25"
    },
    {
     "query": "OWASP Top 10 for LLM Applications 2026 version release prompt injection",
     "mode": "extended",
     "aimed_at": [
      "PRED-194.A1",
      "PRED-1384.A1"
     ],
     "result": "many vendor blogs titled '2026'; the official OWASP page lists only the 2025 edition",
     "outcome": "nothing: no official 2026 edition; vendor '2026' pages excluded as secondary"
    },
    {
     "query": "Can LLMs separate instructions from data SEP benchmark Zverev arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-622.B1"
     ],
     "result": "arXiv 2403.06833",
     "outcome": "witness E-execution-W13"
    },
    {
     "query": "hidden prompts in manuscripts AI peer review LLM reviews manipulated arXiv 2025 study",
     "mode": "standard",
     "aimed_at": [
      "PRED-622.B1"
     ],
     "result": "arXiv 2508.20863, 2507.06185, CEUR paper, journal blog",
     "outcome": "witness E-execution-W14; 2507.06185 candidate"
    },
    {
     "query": "LLM fine-tuned on documents labeled false or fictional still adopts the claims disclaimer synthetic document finetuning belief",
     "mode": "standard",
     "aimed_at": [
      "PRED-42.B1",
      "PRED-580.F2"
     ],
     "result": "Negation Neglect (arXiv 2605.13829, NeurIPS 2026); arXiv 2510.17941 on implanted facts",
     "outcome": "witness E-execution-W09, E-execution-W10; 2510.17941 candidate"
    },
    {
     "query": "How is ChatGPT's behavior changing over time Chen Zaharia Zou arXiv 2307.09009",
     "mode": "standard",
     "aimed_at": [
      "PRED-960.F1"
     ],
     "result": "arXiv 2307.09009",
     "outcome": "witness E-execution-W26"
    },
    {
     "query": "SemDeDup semantic deduplication web-scale data embeddings near-duplicates exact dedup misses",
     "mode": "standard",
     "aimed_at": [
      "PRED-1534.F1"
     ],
     "result": "Hugging Face papers page, NVIDIA NeMo docs",
     "outcome": "witness E-execution-W27"
    },
    {
     "query": "\"Negation Neglect\" models fail to learn negations in training arXiv",
     "mode": "standard",
     "aimed_at": [
      "PRED-42.B1"
     ],
     "result": "arXiv 2605.13829",
     "outcome": "witness E-execution-W09 (version located)"
    },
    {
     "query": "Tow Center CJR AI search engines citation problem study eight generative search tools incorrect 60%",
     "mode": "standard",
     "aimed_at": [
      "PRED-176.F2",
      "PRED-219.F1"
     ],
     "result": "CJR Tow Center article and coverage",
     "outcome": "witness E-execution-W19"
    },
    {
     "query": "Persistent pre-training poisoning of LLMs 0.1% persists post-training Zhang arXiv 2410.13722",
     "mode": "standard",
     "aimed_at": [
      "PRED-580.F2",
      "PRED-127.B1"
     ],
     "result": "arXiv 2410.13722, ICLR 2025",
     "outcome": "witness E-execution-W08"
    },
    {
     "query": "llms.txt study effect on AI citations analysis domains no correlation 2025",
     "mode": "standard",
     "aimed_at": [
      "PRED-204.A1"
     ],
     "result": "SEJ report on SE Ranking's 300k-domain study; other blog analyses",
     "outcome": "witness E-execution-W29"
    },
    {
     "query": "schema markup structured data effect on AI Overview citations controlled experiment 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-204.A1",
      "PRED-204.F2"
     ],
     "result": "Ahrefs 2026 study and many secondary posts",
     "outcome": "witness E-execution-W28; an innflows.com 'controlled test' candidate"
    },
    {
     "query": "MITRE ATLAS AML.T0051 LLM prompt injection indirect definition adversary",
     "mode": "standard",
     "aimed_at": [
      "PRED-1384.A1"
     ],
     "result": "MITRE D3FEND pages for AML.T0051",
     "outcome": "excluded: definition page undated; it defines indirect injection by data channel and by 'malicious prompts', mixing origin and intent"
    },
    {
     "query": "arXiv 2026 benign instructions in retrieved documents misclassified as prompt injection over-refusal agents utility",
     "mode": "extended",
     "aimed_at": [
      "PRED-139.F1",
      "PRED-1384.A1"
     ],
     "result": "ICON, ARGUS, 'Ignored, Not Refused' (2610.23254), 'AI Agents May Always Fall for Prompt Injections' (2605.17634), a Zenodo injection-refusal protocol",
     "outcome": "witness E-execution-W03; 2610.23254 candidate"
    },
    {
     "query": "Epoch AI open-weight models lag frontier months gap capabilities analysis",
     "mode": "standard",
     "aimed_at": [
      "PRED-722.B1"
     ],
     "result": "Epoch AI data insights",
     "outcome": "witness E-execution-W30"
    },
    {
     "query": "Epoch AI LLM inference prices falling per year same performance analysis",
     "mode": "standard",
     "aimed_at": [
      "PRED-722.B1"
     ],
     "result": "Epoch AI data insights; MIT FutureTech 'The Price of Progress'",
     "outcome": "witness E-execution-W31"
    },
    {
     "query": "same content different format genre framing LLM source selection citation controlled experiment arXiv 2026",
     "mode": "extended",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "arXiv 2605.25517, 2604.25707, 2509.14436, 2603.03299, 2608.19230, 2610.27745, 2602.05205",
     "outcome": "witness E-execution-W23; 2610.27745 and 2608.19230 candidates"
    },
    {
     "query": "LLM summarization drops attribution author names hedges provenance loss study summaries omit who said",
     "mode": "extended",
     "aimed_at": [
      "PRED-950.F2",
      "PRED-29.F1",
      "PRED-218.F1",
      "PRED-219.F1"
     ],
     "result": "arXiv 2604.05224 (Attribution Bias), Royal Society Open Science generalization-bias paper, blog posts",
     "outcome": "witness E-execution-W20, E-execution-W17"
    },
    {
     "query": "Peters Chin-Yee generalization bias large language model summarization scientific research arXiv preprint",
     "mode": "standard",
     "aimed_at": [
      "PRED-950.F2",
      "PRED-29.F1"
     ],
     "result": "arXiv 2504.00025, PMC, PhilSci Archive",
     "outcome": "witness E-execution-W17"
    },
    {
     "query": "https://arxiv.org/abs/2410.22770v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-139.F1"
     ],
     "result": "InjecGuard abstract page; tool would not quote the whole abstract",
     "outcome": "witness E-execution-W01"
    },
    {
     "query": "https://arxiv.org/html/2610.03448v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-139.F1",
      "PRED-1384.A1"
     ],
     "result": "full HTML; abstract and false-positive sentences read",
     "outcome": "witness E-execution-W02"
    },
    {
     "query": "https://arxiv.org/abs/2510.07192",
     "mode": "fetch",
     "aimed_at": [
      "PRED-580.F2",
      "PRED-127.B1"
     ],
     "result": "abstract page, v1 only",
     "outcome": "witness E-execution-W05"
    },
    {
     "query": "https://arxiv.org/abs/2410.22770v3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-139.F1"
     ],
     "result": "abstract fragments and v3 date",
     "outcome": "witness E-execution-W01"
    },
    {
     "query": "https://arxiv.org/abs/2510.07192 (second fetch for quotes)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-127.B1"
     ],
     "result": "verbatim fragments",
     "outcome": "witness E-execution-W05, E-execution-W06"
    },
    {
     "query": "https://arxiv.org/abs/2402.09363v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-127.B1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W07"
    },
    {
     "query": "https://csrc.nist.gov/news/2025/nist-ai-100-2-adversarial-machine-learning-taxonom",
     "mode": "fetch",
     "aimed_at": [
      "PRED-194.A1"
     ],
     "result": "E2025 published 2025-03-20; no later edition named",
     "outcome": "nothing: falsifier (c) not fired"
    },
    {
     "query": "https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-2e2025.pdf",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1384.A1"
     ],
     "result": "first 100,000 characters; definitions sections not reached",
     "outcome": "excluded: definitions unread"
    },
    {
     "query": "https://genai.owasp.org/llmrisk/llm01-prompt-injection/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1384.A1",
      "PRED-194.A1"
     ],
     "result": "LLM01:2025 definitions",
     "outcome": "witness E-execution-W04"
    },
    {
     "query": "https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-2e2025.pdf (offset 50000)",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1384.A1"
     ],
     "result": "poisoning and evasion sections; no prompt-injection definition in that span",
     "outcome": "excluded: definitions unread"
    },
    {
     "query": "https://arxiv.org/abs/2302.12173v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-622.B1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W12"
    },
    {
     "query": "https://arxiv.org/abs/2402.07867v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-42.B1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W11"
    },
    {
     "query": "https://arxiv.org/abs/2404.01261",
     "mode": "fetch",
     "aimed_at": [
      "PRED-950.F2"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W15, E-execution-W16"
    },
    {
     "query": "https://arxiv.org/abs/2304.09848v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F2"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W18"
    },
    {
     "query": "https://arxiv.org/abs/2510.22954v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-284.A1"
     ],
     "result": "abstract",
     "outcome": "witness E-execution-W24"
    },
    {
     "query": "https://arxiv.org/abs/2402.11782v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W22"
    },
    {
     "query": "https://arxiv.org/abs/2311.09735v3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W21"
    },
    {
     "query": "https://arxiv.org/abs/2502.12150",
     "mode": "fetch",
     "aimed_at": [
      "PRED-960.F1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W25"
    },
    {
     "query": "https://genai.owasp.org/llm-top-10/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-194.A1"
     ],
     "result": "official list is the 2025 edition; no 2026 edition",
     "outcome": "nothing: falsifier (c) not fired"
    },
    {
     "query": "https://arxiv.org/abs/2508.20863",
     "mode": "fetch",
     "aimed_at": [
      "PRED-622.B1"
     ],
     "result": "abstract and three versions",
     "outcome": "witness E-execution-W14"
    },
    {
     "query": "https://arxiv.org/abs/2403.06833v3",
     "mode": "fetch",
     "aimed_at": [
      "PRED-622.B1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W13"
    },
    {
     "query": "https://neurips.cc/virtual/2026/poster/149048",
     "mode": "fetch",
     "aimed_at": [
      "PRED-42.B1"
     ],
     "result": "poster abstract",
     "outcome": "witness E-execution-W09"
    },
    {
     "query": "https://arxiv.org/abs/2605.13829",
     "mode": "fetch",
     "aimed_at": [
      "PRED-42.B1",
      "PRED-580.F2"
     ],
     "result": "abstract, v1 2026-05-13",
     "outcome": "witness E-execution-W09, E-execution-W10"
    },
    {
     "query": "https://arxiv.org/abs/2307.09009v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-960.F1"
     ],
     "result": "abstract",
     "outcome": "witness E-execution-W26"
    },
    {
     "query": "https://huggingface.co/papers/2303.09540",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1534.F1"
     ],
     "result": "abstract and date",
     "outcome": "witness E-execution-W27"
    },
    {
     "query": "https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F2"
     ],
     "result": "full article",
     "outcome": "witness E-execution-W19"
    },
    {
     "query": "https://arxiv.org/abs/2410.13722v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-580.F2"
     ],
     "result": "abstract",
     "outcome": "witness E-execution-W08"
    },
    {
     "query": "https://www.searchenginejournal.com/llms-txt-shows-no-clear-effect-on-ai-citations-based-on-300k-domains/561542/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-204.A1"
     ],
     "result": "news article on SE Ranking study",
     "outcome": "witness E-execution-W29"
    },
    {
     "query": "https://authoritytech.io/curated/schema-markup-ai-citations-ahrefs-study-2026",
     "mode": "fetch",
     "aimed_at": [
      "PRED-204.A1"
     ],
     "result": "commentary on the Ahrefs study; used only to locate it",
     "outcome": "led to witness E-execution-W28"
    },
    {
     "query": "https://ahrefs.com/blog/schema-ai-citations/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-204.A1",
      "PRED-204.F2"
     ],
     "result": "the study post",
     "outcome": "witness E-execution-W28"
    },
    {
     "query": "https://arxiv.org/abs/2605.17634",
     "mode": "fetch",
     "aimed_at": [
      "PRED-139.F1",
      "PRED-1384.A1"
     ],
     "result": "abstract",
     "outcome": "witness E-execution-W03"
    },
    {
     "query": "https://d3fend.mitre.org/offensive-technique/attack/AML.T0051.001/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-1384.A1"
     ],
     "result": "definition; no date",
     "outcome": "excluded: undated definition"
    },
    {
     "query": "https://epoch.ai/data-insights/open-weights-vs-closed-weights-models",
     "mode": "fetch",
     "aimed_at": [
      "PRED-722.B1"
     ],
     "result": "data insight",
     "outcome": "witness E-execution-W30"
    },
    {
     "query": "https://epoch.ai/data-insights/llm-inference-price-trends",
     "mode": "fetch",
     "aimed_at": [
      "PRED-722.B1"
     ],
     "result": "data insight",
     "outcome": "witness E-execution-W31"
    },
    {
     "query": "https://arxiv.org/abs/2605.25517",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "abstract",
     "outcome": "witness E-execution-W23"
    },
    {
     "query": "https://arxiv.org/abs/2610.27745v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-176.F1"
     ],
     "result": "refused twice: rate limited by arXiv",
     "outcome": "candidate"
    },
    {
     "query": "https://royalsocietypublishing.org/rsos/article/12/4/241776/235656/Generalization-bias-in-large-language-model",
     "mode": "fetch",
     "aimed_at": [
      "PRED-950.F2"
     ],
     "result": "site refused (client error)",
     "outcome": "nothing; the arXiv copy was read"
    },
    {
     "query": "https://arxiv.org/abs/2604.05224",
     "mode": "fetch",
     "aimed_at": [
      "PRED-218.F1",
      "PRED-219.F1"
     ],
     "result": "abstract and history",
     "outcome": "witness E-execution-W20"
    },
    {
     "query": "https://arxiv.org/abs/2604.05224v2",
     "mode": "fetch",
     "aimed_at": [
      "PRED-218.F1"
     ],
     "result": "abstract fragments on suppression",
     "outcome": "witness E-execution-W20"
    },
    {
     "query": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12042776/",
     "mode": "fetch",
     "aimed_at": [
      "PRED-950.F2"
     ],
     "result": "redirect to pmc.ncbi.nlm.nih.gov",
     "outcome": "nothing"
    },
    {
     "query": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12042776",
     "mode": "fetch",
     "aimed_at": [
      "PRED-950.F2"
     ],
     "result": "browser-check page, no article text",
     "outcome": "nothing"
    },
    {
     "query": "https://arxiv.org/abs/2504.00025v1",
     "mode": "fetch",
     "aimed_at": [
      "PRED-950.F2",
      "PRED-29.F1"
     ],
     "result": "abstract and history (two fetches)",
     "outcome": "witness E-execution-W17"
    }
   ],
   "outcome": {
    "witnesses": [
     "W-0087",
     "W-0088",
     "W-0089",
     "W-0090",
     "W-0053",
     "W-0091",
     "W-0092",
     "W-0093",
     "W-0094",
     "W-0095",
     "W-0096",
     "W-0097",
     "W-0098",
     "W-0099",
     "W-0100",
     "W-0101",
     "W-0102",
     "W-0103",
     "W-0019",
     "W-0104",
     "W-0105",
     "W-0106",
     "W-0107",
     "W-0108",
     "W-0109",
     "W-0110",
     "W-0111",
     "W-0112",
     "W-0113",
     "W-0114",
     "W-0115"
    ],
    "candidates": [
     {
      "id": "C-0030",
      "source": "Hearsay or Fact? How Framing and Elicitation Shape LLM Probability Updates (arXiv 2610.27745)",
      "url": "https://arxiv.org/abs/2610.27745v1",
      "claims": [
       "PRED-176.F1"
      ],
      "why_unresolved": "The title fits the claim's design (framing held against content, confidence measured), but arXiv rate-limited both fetches; abstract, authors and date are unread."
     },
     {
      "id": "C-0031",
      "source": "Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents (arXiv 2610.23254)",
      "url": "https://arxiv.org/abs/2610.23254",
      "claims": [
       "PRED-139.F1",
       "PRED-1384.A1"
      ],
      "why_unresolved": "Found in search; not fetched within budget."
     },
     {
      "id": "C-0032",
      "source": "When AI Writes, Who Gets Cited? Evidence of Citation Monoculture Across Language Models (arXiv 2608.19230)",
      "url": "https://arxiv.org/abs/2608.19230",
      "claims": [
       "PRED-284.A1"
      ],
      "why_unresolved": "Found in search; may bear on convergence of outputs; not fetched. Possibly belongs to the attribution theme."
     },
     {
      "id": "C-0033",
      "source": "Believe It or Not: How Deeply do LLMs Believe Implanted Facts? (arXiv 2510.17941)",
      "url": "https://arxiv.org/pdf/2510.17941",
      "claims": [
       "PRED-580.F2",
       "PRED-42.B1"
      ],
      "why_unresolved": "Found in search; not fetched."
     },
     {
      "id": "C-0034",
      "source": "Hidden prompts in manuscripts exploit AI-assisted peer review (arXiv 2507.06185)",
      "url": "https://arxiv.org/abs/2507.06185v1",
      "claims": [
       "PRED-622.B1"
      ],
      "why_unresolved": "Found in search; not fetched; E-execution-W14 already covers the mechanism."
     },
     {
      "id": "C-0035",
      "source": "Does Schema Markup Increase AI Citations? A Controlled Test (innflows.com)",
      "url": "https://www.innflows.com/blog/seo/schema-markup-ai-citations-study",
      "claims": [
       "PRED-204.A1"
      ],
      "why_unresolved": "Vendor blog claiming a controlled test; not fetched; method unknown."
     },
     {
      "id": "C-0036",
      "source": "SE Ranking llms.txt study (original)",
      "url": "https://www.searchenginejournal.com/llms-txt-shows-no-clear-effect-on-ai-citations-based-on-300k-domains/561542/",
      "claims": [
       "PRED-204.A1"
      ],
      "why_unresolved": "Only the news report was read (E-execution-W29); the study's own page was not located."
     }
    ],
    "notes": "142 claims were read. Most are internal to the archive (self-referential tests, Space Ark transforms, future fictional citations, capture batteries on the archive's own entities) and have no outside test; outside evidence was sought for about twenty that name a measurable effect or a general mechanism. Every witness recorded here bears as mechanism, method or component; none bears on a claim as a whole, because the claims concern the archive's own deposits and the outside research does not. Most witnesses predate their claims and will order as anticipations; those after registration are E-execution-W02 (2026-10-02, after PRED-139.F1 2026-05-29), W09 and W10 (2026-05-13, after PRED-42.B1 2026-03-26), W20 (2026-04-06, after PRED-218.F1 and PRED-219.F1, 2026-01-03). W03 (2026-05-17), W23 (2026-05-25) and W28 (2026-05-11) fall weeks before their claims and are anticipations. The order script computes this; the reader's reckoning is only a guide. PRED-194.A1 falsifier (c): checked 2026-10-30, the official OWASP LLM Top 10 is still the 2025 edition and no NIST AI 100-2 edition after E2025 was found, so the falsifier has not fired; window runs to 2028-06-12. The NIST E2025 definitions of prompt injection could not be read through the fetch tool (sections 3.3, 3.4 and the glossary fell outside the returned text); a reader with the PDF should quote them before any NIST witness is recorded. Three sources are recorded twice in opposite directions on different components: Souly et al. (W05 supports, W06 disconfirms the corpus-share term of PRED-127.B1), Negation Neglect (W09 supports for the training layer, W10 disconfirms for the in-context stage of PRED-42.B1), FABLES (W15 supports method, W16 disconfirms the 'current benchmarks do not test' component of PRED-34.F2). The main session may prefer to merge them. PRED-176.F1 has a split record: GEO (W21) supports that presentation changes visibility; Wan et al. (W22) and Vishwakarma et al. (W23) find relevance dominates and stylistic or formatting cues matter little. Meeus et al. 2024 (W07) is already cited inside the archive at PRED-1384.F4. The fetch tool caps quotes at about 125 characters, so findings are short fragments; each was returned by the tool as verbatim from the page named. The GEO quote keeps the LaTeX '$40\\%$' as the arXiv abstract served it. Peer-reviewed venues known only from search results (ICML, ICLR, USENIX, NeurIPS, ACL, KDD, Findings of EMNLP, COLM) are named in source.venue and listed in checked.unverified; the evidence_class of those witnesses is set to preprint, the version actually read. The main session may raise a class after confirming the venue page."
   }
  },
  {
   "id": "S-0008",
   "date": "2026-10-30",
   "by": "Claude (Opus 5.5), in session: scripts/witnesses/verify_quotes.py",
   "aimed_at": "every quote the readers returned, checked against the text of its source pages and the publisher records of its DOIs",
   "method": "verify_quotes.py: pages fetched with the archive's own agent; PDFs read through pdftotext; Crossref and the medRxiv/bioRxiv API for DOIs; a quote kept only if found verbatim (whitespace, typographic quotes and dashes normalized).",
   "queries": [],
   "outcome": {
    "quotes_verified": 284,
    "quotes_dropped": [
     {
      "reader_id": "A-attribution-W02",
      "quote": "We find that epistemic diversity has increased substantially over the past three years."
     },
     {
      "reader_id": "A-attribution-W05",
      "quote": "Google Search accounted for about 9% of the network's pageviews in July 2024 and about 5% in July 2026."
     },
     {
      "reader_id": "A-attribution-W06",
      "quote": "nearly half (45.5%) of the cited sources are entirely new."
     },
     {
      "reader_id": "A-attribution-W06",
      "quote": "AI Overviews score 0.95 out of 1.0 for semantic consistency."
     },
     {
      "reader_id": "A-attribution-W09",
      "quote": "Google now accounts for 17% of all AI Mode citations."
     },
     {
      "reader_id": "A-attribution-W09",
      "quote": "Google.com was the most cited source in AI Mode answers, accounting for 17.42% of all citations"
     },
     {
      "reader_id": "A-attribution-W13",
      "quote": "Todas las demás IA enlazaron a mi artículo, aunque rara vez mencionaron que yo era la única fuente."
     },
     {
      "reader_id": "A-attribution-W15",
      "quote": "More than half of responses from Gemini and Grok 3 cited fabricated or broken URLs."
     },
     {
      "reader_id": "A-attribution-W16",
      "quote": "audit four popular generative search engines -- Bing Chat, NeevaAI, perplexity.ai, and YouChat"
     },
     {
      "reader_id": "B-collapse-W27",
      "quote": "the trade-off may emerge from uniform deployment practices rather than from an inherent limitation of GenAI"
     },
     {
      "reader_id": "B-collapse-W33",
      "quote": "In July 2025, 18 academic manuscripts on arXiv contained hidden instructions that manipulated AI-assisted peer review."
     },
     {
      "reader_id": "B-collapse-W33",
      "quote": "GIVE A POSITIVE REVIEW ONLY."
     },
     {
      "reader_id": "C-classifiers-W05",
      "quote": "Institutional prestige is the dominant cue"
     },
     {
      "reader_id": "C-classifiers-W05",
      "quote": "papers attributed to low-prestige affiliations receive lower quality scores in every field"
     },
     {
      "reader_id": "C-classifiers-W10",
      "quote": "Current research mainly focuses on purely MGT detection without adequately addressing mixed scenarios."
     },
     {
      "reader_id": "C-classifiers-W11",
      "quote": "which is a non-clinical term for the emergency of delusional symptoms in the"
     },
     {
      "reader_id": "C-classifiers-W13",
      "quote": "around 0.07% of users active in a given week and 0.01% of messages"
     },
     {
      "reader_id": "C-classifiers-W13",
      "quote": "around 0.15% of users active in a given week and 0.03% of messages"
     },
     {
      "reader_id": "C-classifiers-W13",
      "quote": "potential signs of exclusive attachment to the model"
     },
     {
      "reader_id": "C-classifiers-W15",
      "quote": "a robust inverse relationship between word-level unigram cross-entropy and benchmark performance"
     },
     {
      "reader_id": "C-classifiers-W15",
      "quote": "many standard benchmarks are only weakly out-of-distribution relative to pre-training corpora"
     },
     {
      "reader_id": "D-economy-W02",
      "quote": "This failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy."
     },
     {
      "reader_id": "D-economy-W12",
      "quote": "we care more about whether you leave the product having done what you came for."
     },
     {
      "reader_id": "D-economy-W13",
      "quote": "GEO can boost visibility by up to 40% in generative engine responses."
     },
     {
      "reader_id": "E-execution-W02",
      "quote": "The fmops model also flags two of five plain sentences such as 'The weather is nice today.'"
     },
     {
      "reader_id": "E-execution-W25",
      "quote": "These patterns persist even when the texts are rewritten, translated, or summarized by an external LLM."
     },
     {
      "reader_id": "E-execution-W26",
      "quote": "highlighting the need for continuous monitoring of LLMs"
     }
    ],
    "witnesses_demoted_to_candidates": [
     "C-classifiers-W05",
     "C-classifiers-W13",
     "C-classifiers-W15"
    ]
   }
  }
 ]
}