{
 "what": "External witnesses of the prediction index: research and evidence from outside the archive bearing on its claims. Protocol: data/witnesses/WITNESS-PROTOCOL.md. A witness never changes a claim's status.",
 "adopted": "2026-10-30",
 "evidence_classes": [
  "peer-reviewed article",
  "peer-reviewed conference paper",
  "preprint",
  "working paper",
  "report",
  "dataset",
  "book",
  "thesis",
  "documentation",
  "post",
  "news",
  "system output",
  "other"
 ],
 "directions": [
  "supports",
  "disconfirms"
 ],
 "bears_on": [
  "whole",
  "component",
  "mechanism",
  "method"
 ],
 "witnesses": [
  {
   "id": "W-0001",
   "direction": "supports",
   "evidence_class": "peer-reviewed conference paper",
   "claims": [
    {
     "claim": "PRED-209.F5",
     "bears_on": "method",
     "scope": "The probe's method, the same content under different authority signals, carried out by others on LLM paper recommendation. PRED-209.F5 specifies Google's crawl speed, ranking position and summarization treatment, which this study does not measure."
    }
   ],
   "source": {
    "authors": [
     "Uthman Jinadu",
     "Parsa Ghazvinian",
     "Anjila Budathoki",
     "Benjamin M. Ampel",
     "Rajshekhar Sunderraman",
     "Yi Ding"
    ],
    "title": "Authority Bias in Conversational Search Engines for Academic Paper Recommendation",
    "venue": "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Budapest, pp. 24548–24568",
    "doi": "10.18653/2026.emnlp-main.1069",
    "urls": [
     "https://aclanthology.org/2026.emnlp-main.1069/",
     "https://arxiv.org/abs/2609.00248"
    ]
   },
   "versions": [
    {
     "date": "2026-08-31",
     "kind": "preprint",
     "where": "arXiv 2609.00248 v1",
     "url": "https://arxiv.org/abs/2609.00248"
    },
    {
     "date": "2026-10",
     "kind": "peer-reviewed conference paper",
     "where": "EMNLP 2026",
     "url": "https://aclanthology.org/2026.emnlp-main.1069/"
    }
   ],
   "finding": [
    {
     "quote": "Holding title and abstract constant, we vary authority metadata across three counterfactual conditions",
     "from": "abstract, ACL Anthology"
    },
    {
     "quote": "debiasing instructions suppress authority mentions far faster than authority-driven flips",
     "from": "abstract, ACL Anthology"
    }
   ],
   "shows": "Across eight LLMs (five open-weight, three closed), recommendations of the same paper shift with its authority metadata (author prestige, venue, citations), substantially and directionally; instructions against bias remove the mention of authority faster than the authority-driven choice, so auditing what the model says understates what it does.",
   "does_not_show": "Google indexing, crawl speed or ranking; the archive's own entities; any effect size the archive registered.",
   "checked": {
    "by": "Claude (Opus 5.5), in session, at the operator's request",
    "date": "2026-10-30",
    "how": "Read the ACL Anthology landing page and the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the count of 17,898 parsed observations reported in S-0001: not in the abstract or the page metadata"
    ]
   },
   "found_by": [
    "S-0001",
    "S-0002"
   ]
  },
  {
   "id": "W-0002",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-209.G2",
     "bears_on": "mechanism",
     "scope": "Unequal retention by kind of source, found in the composed answer: sources are cited and then under-used in the summary. PRED-209.G2 states lower retention for low-authority domains in the index; the witness measures representation in the synthesis, by source type."
    }
   ],
   "source": {
    "authors": [
     "Michelle Huang",
     "Agam Goyal",
     "Koustuv Saha",
     "Eshwar Chandrasekharan"
    ],
    "title": "Answer Bubbles: Information Exposure in AI-Mediated Search",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2603.16138v1"
    ]
   },
   "versions": [
    {
     "date": "2026-03-17",
     "kind": "preprint",
     "where": "arXiv 2603.16138 v1",
     "url": "https://arxiv.org/abs/2603.16138v1"
    }
   ],
   "finding": [
    {
     "quote": "We examine responses to 11,000 real search queries across four systems -- vanilla GPT, Search GPT, Google AI Overviews, and traditional Google Search",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Wikipedia and longer sources are disproportionately overrepresented",
     "from": "abstract, arXiv"
    },
    {
     "quote": "cited social media content and negatively framed sources are substantially underrepresented",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Generative search selects sources unequally and the summary compounds it: what is admitted to the source list is not admitted equally to the composed answer.",
   "does_not_show": "Suppression of any named author or entity; the index layer; any claim about the archive's own content.",
   "checked": {
    "by": "Claude (Opus 5.5), in session, at the operator's request",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v1 only listed), 2026-10-30; quotes checked against the page's text by scripts/witnesses/verify_quotes.py.",
    "unverified": [
     "S-0001's 'discussion forums' and 'atomic factual statements': not in the abstract; would need the full paper",
     "S-0001 and S-0002 gave 'five systems'; the abstract names four (corrected 2026-10-30)"
    ]
   },
   "found_by": [
    "S-0001",
    "S-0002"
   ]
  },
  {
   "id": "W-0003",
   "direction": "supports",
   "evidence_class": "peer-reviewed article",
   "claims": [
    {
     "claim": "PRED-476.A2",
     "bears_on": "component",
     "scope": "The component 'Cultural production filtered through AI systems shows reduced diversity', in co-creative tasks. The claim's other components (children's baselines, professional advancement, academic discourse) and its 5–15 year window are not tested."
    }
   ],
   "source": {
    "authors": [
     "Alwin de Rooij",
     "Michael Mose Biskjaer"
    ],
    "title": "Does generative AI make us think alike? A systematic review and meta-analysis of homogenization effects in human–AI co-creation",
    "venue": "Behaviour & Information Technology",
    "doi": "10.1080/0144929X.2026.2726451",
    "urls": [
     "https://research.tilburguniversity.edu/en/publications/does-generative-ai-make-us-think-alike-a-systematic-review-and-me-2/",
     "https://doi.org/10.31234/osf.io/rz5s4_v1"
    ]
   },
   "versions": [
    {
     "date": "2026-04-21",
     "kind": "preprint",
     "where": "PsyArXiv, 10.31234/osf.io/rz5s4_v1",
     "url": "https://doi.org/10.31234/osf.io/rz5s4_v1"
    },
    {
     "date": "2026-08-31",
     "kind": "peer-reviewed article",
     "where": "Behaviour & Information Technology, e-pub ahead of print",
     "url": "https://doi.org/10.1080/0144929X.2026.2726451"
    }
   ],
   "finding": [
    {
     "quote": "(19 studies, 61 effect sizes)",
     "from": "abstract, Tilburg University Research Portal"
    },
    {
     "quote": "a small but statistically significant homogenization effect associated with AI use",
     "from": "abstract, Tilburg University Research Portal"
    },
    {
     "quote": "robust across sensitivity analyses and not explained by publication bias",
     "from": "abstract, Tilburg University Research Portal"
    },
    {
     "quote": "stronger effects in semantically constrained ideation tasks than in minimally constrained divergent thinking tasks",
     "from": "abstract, Tilburg University Research Portal"
    }
   ],
   "shows": "A small, significant, robust homogenization of creative output with AI use, larger where the task is constrained.",
   "does_not_show": "Generational or developmental effects; safety classifiers as the cause; the population-level trajectory of the claim's later orders.",
   "checked": {
    "by": "Claude (Opus 5.5), in session, at the operator's request",
    "date": "2026-10-30",
    "how": "Read the Tilburg University Research Portal record, 2026-10-30; the journal's DOI page refused the fetch. Dates from Crossref: PsyArXiv 10.31234/osf.io/rz5s4_v1 posted 2026-04-21 (is-preprint-of the journal article); Behaviour & Information Technology published online 2026-08-31.",
    "unverified": []
   },
   "found_by": [
    "S-0001",
    "S-0002"
   ]
  },
  {
   "id": "W-0004",
   "direction": "supports",
   "evidence_class": "peer-reviewed article",
   "claims": [
    {
     "claim": "PRED-476.B1",
     "bears_on": "component",
     "scope": "The term P, 'false positive rate for non-normative cognition' (#476 l.112), on the mechanism 'Safety classifiers optimized for recall produce systematic false positives' (l.97): shown for LLMs scoring psychosis-risk interviews, a high-recall, low-specificity classifier whose errors over-pathologize. The feedback loop B1 states, and safety classifiers as such, are not tested."
    }
   ],
   "source": {
    "authors": [
     "Taiyu Zhu",
     "Alexander Tashevski",
     "Maxime Taquet",
     "Matilda Azis",
     "Tia Jani",
     "Matthew R. Broome",
     "Thomas Kabir",
     "Amedeo Minichino",
     "Graham K. Murray",
     "Matthew M. Nour",
     "Ilina Singh",
     "Paolo Fusar-Poli",
     "Alejo Nevado-Holgado",
     "Philip McGuire",
     "Dominic Oliver"
    ],
    "title": "Evaluating Large Language Models for Assessment of Psychosis Risk",
    "venue": "npj Digital Medicine (preprint: medRxiv)",
    "doi": "10.1038/s41746-026-02928-4",
    "urls": [
     "https://www.nature.com/articles/s41746-026-02928-4",
     "https://www.medrxiv.org/content/10.64898/2026.04.02.26349960v1",
     "https://doi.org/10.64898/2026.04.02.26349960"
    ]
   },
   "versions": [
    {
     "date": "2026-04-04",
     "kind": "preprint",
     "where": "medRxiv v1, 10.64898/2026.04.02.26349960",
     "url": "https://www.medrxiv.org/content/10.64898/2026.04.02.26349960v1"
    },
    {
     "date": "2026-06-15",
     "kind": "preprint",
     "where": "medRxiv v2",
     "url": "https://www.medrxiv.org/content/10.64898/2026.04.02.26349960v2"
    },
    {
     "date": "2026-07-23",
     "kind": "peer-reviewed article",
     "where": "npj Digital Medicine, 10.1038/s41746-026-02928-4",
     "url": "https://doi.org/10.1038/s41746-026-02928-4"
    }
   ],
   "finding": [
    {
     "quote": "We assessed 11 open-weight LLMs on 678 partial PSYCHS interview transcripts from 373 participants (77.7% CHR-P).",
     "from": "abstract, medRxiv (server record, v1 and v2) and npj Digital Medicine (Crossref)"
    },
    {
     "quote": "Larger models achieved the strongest classification performance (Llama-3.3-70B: accuracy = 0.80, sensitivity = 0.93, specificity = 0.58).",
     "from": "abstract, medRxiv (server record, v1 and v2)"
    },
    {
     "quote": "Errors primarily reflected over-pathologisation of non-clinical experiences.",
     "from": "abstract, medRxiv (server record, v1 and v2) and npj Digital Medicine (Crossref)"
    }
   ],
   "shows": "LLMs scoring clinical interviews err mainly by reading non-clinical experience as pathological; the strongest classifier pairs high sensitivity with low specificity.",
   "does_not_show": "Ordinary conversation, poetry or philosophy; safety classifiers; the liability incentive or the training feedback loop of B1.",
   "checked": {
    "by": "Claude (Opus 5.5), in session, at the operator's request",
    "date": "2026-10-30",
    "how": "medRxiv's page and npj's page refused or redirected the fetch; the abstract and dates read from the medRxiv server's API record (api.biorxiv.org, v1 2026-04-04, v2 2026-06-15, published DOI 10.1038/s41746-026-02928-4) and Crossref (npj, published 2026-07-23); quotes checked by scripts/witnesses/verify_quotes.py.",
    "unverified": [
     "S-0001 and S-0002 quoted 'sensitivity (0.934) … specificity (0.580)' from the results; the abstract gives 0.93 and 0.58, quoted instead (corrected 2026-10-30)"
    ]
   },
   "found_by": [
    "S-0001",
    "S-0002"
   ]
  },
  {
   "id": "W-0005",
   "direction": "supports",
   "evidence_class": "peer-reviewed article",
   "claims": [
    {
     "claim": "PRED-251.F6",
     "bears_on": "mechanism",
     "scope": "The attribution gap counts relevant pages read and left uncited by search-enabled LLMs; PER counts correct uses of a framework's concepts left unattributed. The unit is the URL in the witness and the concept in the claim."
    },
    {
     "claim": "PRED-716.F1",
     "bears_on": "method",
     "scope": "A multi-system measurement of attribution loss in composed answers, carried out by others on about 14,000 logs across Gemini, GPT-4o and Perplexity Sonar. It scores URLs read against URLs cited, with no claim segmentation and no two-annotator PER."
    },
    {
     "claim": "PRED-1421.F1",
     "bears_on": "method",
     "scope": "An outside research group computing an attribution-loss rate across surfaces, the program the claim says any community could run. It measures web pages at large, with no comparison across fields or languages."
    },
    {
     "claim": "PRED-3.B2",
     "bears_on": "mechanism",
     "scope": "Gemini answering with no clickable citation in 92% of answers is the behaviour the claim forecasts for AI Overviews. The data are LMArena conversations with the Gemini model, a different surface from AI Overviews, gathered in 2025."
    }
   ],
   "source": {
    "authors": [
     "Ilan Strauss",
     "Jangho Yang",
     "Tim O'Reilly",
     "Sruly Rosenblat",
     "Isobel Moure"
    ],
    "title": "The Attribution Crisis in LLM Search Results (journal version: The attribution crisis in LLM search results: Estimating ecosystem exploitation)",
    "venue": "Social Science Research Council AI Disclosures Project working paper; arXiv cs.DL; Data & Policy (Cambridge University Press) vol. 8, e15",
    "doi": "10.1017/dap.2026.10064",
    "urls": [
     "https://arxiv.org/abs/2508.00838",
     "https://arxiv.org/html/2508.00838v1",
     "https://rogue-scholar.org/records/b7q0a-dfk44",
     "https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3",
     "https://www.ssrc.org/publications/the-attribution-crisis-in-llm-search-results/"
    ]
   },
   "versions": [
    {
     "date": "2025-06-27",
     "kind": "working paper",
     "where": "Rogue Scholar record of the Asimov's Addendum / AI Disclosures Project post, DOI 10.59350/veayq-6bz87",
     "url": "https://rogue-scholar.org/records/b7q0a-dfk44"
    },
    {
     "date": "2025-06-27",
     "kind": "preprint",
     "where": "arXiv 2508.00838 v1 (the only version listed)",
     "url": "https://arxiv.org/abs/2508.00838"
    },
    {
     "date": "2026-04-28",
     "kind": "peer-reviewed article",
     "where": "Data & Policy 8, e15, published online (issue date given as January 2026)",
     "url": "https://www.cambridge.org/core/journals/data-and-policy/article/attribution-crisis-in-llm-search-results-estimating-ecosystem-exploitation/170DD0B88E5F5AEA8F69F2E9AF1328E3"
    }
   ],
   "finding": [
    {
     "quote": "“attribution gap” – the difference between relevant URLs read and those actually cited.",
     "from": "abstract, arXiv HTML v1"
    },
    {
     "quote": "Drawing on approximately 14,000 real-world LMArena conversation logs with search-enabled LLM systems",
     "from": "abstract, arXiv HTML v1"
    },
    {
     "quote": "No citation: Gemini provides no clickable citation source in 92% of answers",
     "from": "abstract, arXiv HTML v1"
    },
    {
     "quote": "Perplexity’s Sonar visits approximately 10 relevant pages per query but cites only three to four.",
     "from": "abstract, arXiv HTML v1"
    },
    {
     "quote": "the average query answered by Gemini or Sonar leaves about 3 relevant websites uncited",
     "from": "abstract, arXiv HTML v1"
    }
   ],
   "shows": "Search-enabled LLMs read more relevant pages than they cite, at rates that differ widely by system on the same queries; the authors tie the difference to retrieval design. This is the earlier pass's candidate C-0001, now verified; it precedes every claim it bears on and is an anticipation.",
   "does_not_show": "Loss of attribution for a named author's concepts; AI Overviews as a surface; any skew of the loss by source size or standing; any trend over time.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v1 only listed) and the arXiv HTML v1 abstract, the Rogue Scholar record, and the Data & Policy landing page, 2026-10-30. The SSRC page had refused a fetch in S-0002 and was not retried.",
    "unverified": [
     "The arXiv page gives v1 as 2025-06-27 although an identifier beginning 2508 is normally assigned in August 2025; the Rogue Scholar DOI record independently dates the paper 27 June 2025",
     "Rogue Scholar lists four authors (no Jangho Yang); arXiv and Data & Policy list five",
     "received and accepted dates of the journal version were not on the page read"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W01"
  },
  {
   "id": "W-0006",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-199.A1",
     "bears_on": "mechanism",
     "scope": "The claim forecasts a declining diversity trend across model generations under synthetic saturation, measured in code solution space (SSDI). The witness measures diversity of real-world claims in LLM outputs across 27 models over three years and finds it rose; the domain is knowledge claims, and the claim's caveat about diversity-preserving intervention is not tested."
    },
    {
     "claim": "PRED-857.F1",
     "bears_on": "method",
     "scope": "A longitudinal measurement of diversity across a model ecology, the test the claim proposes, carried out by others on knowledge claims. It uses a fixed topic set across models and countries, a different instrument from the Three Pillars directive."
    }
   ],
   "source": {
    "authors": [
     "Dustin Wright",
     "Sarah Masud",
     "Jared Moore",
     "Srishti Yadav",
     "Maria Antoniak",
     "Peter Ebert Christensen",
     "Chan Young Park",
     "Isabelle Augenstein"
    ],
    "title": "What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.04226v1",
     "https://arxiv.org/abs/2510.04226"
    ]
   },
   "versions": [
    {
     "date": "2025-10-05",
     "kind": "preprint",
     "where": "arXiv 2510.04226 v1",
     "url": "https://arxiv.org/abs/2510.04226v1"
    },
    {
     "date": "2026-08-31",
     "kind": "preprint",
     "where": "arXiv 2510.04226 v7",
     "url": "https://arxiv.org/abs/2510.04226"
    }
   ],
   "finding": [
    {
     "quote": "testing 27 LLMs on 155 topics covering 12 countries, resulting in 1.7M responses and 70M individual claims.",
     "from": "abstract, arXiv v1"
    },
    {
     "quote": "every system is less diverse than a search baseline",
     "from": "abstract, arXiv v1"
    },
    {
     "quote": "RAG can improve diversity, while large models are counterintuitively less diverse than smaller ones.",
     "from": "abstract, arXiv v1"
    },
    {
     "quote": "LLM parametric knowledge systematically reflects English over local-language knowledge for country specific topics.",
     "from": "abstract, arXiv v1"
    }
   ],
   "shows": "The earlier pass's candidate C-0002, verified: a 5 October 2025 preprint. Across 27 LLMs the range of real-world claims grew over three years, while every system stayed below a search baseline, larger models were less diverse, and country-specific knowledge leaned English over local languages. Its time trend runs against a forecast of declining diversity across generations; it precedes both claims and is an anticipation.",
   "does_not_show": "Attribution, citation or provenance of any kind; code solution-space diversity; training on synthetic data as a cause. It bears on none of the A-attribution claims, and the two claims it bears on belong to the B-collapse theme.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract pages for v1 and the current version (v7), 2026-10-30.",
    "unverified": [
     "dates of versions v2 to v6 were not shown on the pages read",
     "the v1 abstract sentences were read through the fetch tool's segmented quotation; the sentence about diversity pessimism was paraphrased by the tool and is left out",
     "quote not found at source, dropped: \"We find that epistemic diversity has increased substantially over the past three years.\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W02"
  },
  {
   "id": "W-0007",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-659.B1",
     "bears_on": "component",
     "scope": "The component 'your organic traffic declines', as a click effect: users click traditional results about half as often when an AI summary appears. The claim's 40–60% over 18 months, the brand framing and the paid-acquisition component are not tested."
    },
    {
     "claim": "PRED-1111.B1",
     "bears_on": "component",
     "scope": "The component 'the replacement of search with summarization', measured as fewer clicks to sources when a summary is shown. The claim's terminus (concentration toward dictatorship) and its other components are not tested."
    }
   ],
   "source": {
    "authors": [
     "Pew Research Center"
    ],
    "title": "Google users are less likely to click on links when an AI summary appears in the results",
    "venue": "Pew Research Center, Short Reads",
    "doi": null,
    "urls": [
     "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/"
    ]
   },
   "versions": [
    {
     "date": "2025-07-22",
     "kind": "report",
     "where": "Pew Research Center Short Reads (date from the URL path)",
     "url": "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/"
    }
   ],
   "finding": [
    {
     "quote": "Users who encountered an AI summary clicked on a traditional search result link in 8% of all visits.",
     "from": "report page, Pew"
    },
    {
     "quote": "Those who did not encounter an AI summary clicked on a search result nearly twice as often (15% of visits).",
     "from": "report page, Pew"
    },
    {
     "quote": "This occurred in just 1% of all visits to pages with such a summary.",
     "from": "report page, Pew (clicks on a link inside the AI summary)"
    },
    {
     "quote": "This happened on 26% of pages with an AI summary, compared with 16% of pages with only traditional search results.",
     "from": "article body, Pew Research Center, on sessions ending"
    },
    {
     "quote": "analyzed data from 900 U.S. adults who agreed to share their online browsing activity.",
     "from": "article body, Pew Research Center"
    }
   ],
   "shows": "Browsing data from 900 U.S. adults in March 2025: an AI summary on the results page roughly halves clicks to traditional results, and links inside the summary are clicked in 1% of visits. It precedes the claim and is an anticipation.",
   "does_not_show": "Traffic decline over time for any site or brand; the size of the decline the claim forecasts.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Pew report page, 2026-10-30.",
    "unverified": [
     "the page as fetched did not name its authors",
     "the publication date is taken from the URL path; the fetched text did not print it",
     "the author names and the exact publication date (taken from the URL path)"
    ]
   },
   "found_by": [
    "S-0003",
    "S-0006"
   ],
   "reader_id": "A-attribution-W03"
  },
  {
   "id": "W-0008",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-659.B1",
     "bears_on": "component",
     "scope": "Google organic search traffic to news publishers fell 33% globally (38% in the U.S.) over November 2024 to November 2025. The claim forecasts 40–60% over 18 months for brands; the witness covers news sites, a year before the claim, and does not separate the AI Overview effect."
    },
    {
     "claim": "PRED-1091.B1",
     "bears_on": "component",
     "scope": "The component 'continued organic decline': a measured year of decline before the claim, and publishers' own expectation of further decline. The claim's 'through 2026 and 2027' is a later period."
    }
   ],
   "source": {
    "authors": [
     "Nic Newman"
    ],
    "title": "Journalism, media, and technology trends and predictions 2026",
    "venue": "Reuters Institute for the Study of Journalism (traffic data from Chartbeat)",
    "doi": null,
    "urls": [
     "https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026"
    ]
   },
   "versions": [
    {
     "date": "2026-01-12",
     "kind": "report",
     "where": "Reuters Institute website",
     "url": "https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026"
    }
   ],
   "finding": [
    {
     "quote": "down by a third (33%) globally",
     "from": "report page, Reuters Institute (Chartbeat data, Google organic search traffic to publishers)"
    },
    {
     "quote": "Publishers expect traffic from search engines to decline by more than 40% over the next three years",
     "from": "report page, Reuters Institute (survey of 280 media leaders)"
    }
   ],
   "shows": "Chartbeat data across about 2,576 news sites: Google organic search traffic to publishers fell by a third in the year to November 2025. It precedes both claims and is an anticipation.",
   "does_not_show": "Brands outside news; the share of the decline due to AI Overviews (the report says this is unclear); the 18-month magnitude.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the first 100,000 characters of the Reuters Institute report page, 2026-10-30.",
    "unverified": [
     "the 38% U.S. figure and the 2,576-site count were reported by the fetch tool only in paraphrase",
     "the last 19,070 characters of the page were not read"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W04"
  },
  {
   "id": "W-0009",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-659.B1",
     "bears_on": "component",
     "scope": "A 40.2% year-over-year fall in Google Search referrals across Chartbeat's client network (July 2025 to July 2026), inside the claim's 40–60% range though over 12 months and for news publishers. The measured period mostly precedes the claim's registration (2026-04-14)."
    },
    {
     "claim": "PRED-1091.B1",
     "bears_on": "component",
     "scope": "The component 'continued organic decline through 2026': the decline accelerated from 21.9% to 40.2% year over year. The witness does not reach 2027 or the claim's other components."
    }
   ],
   "source": {
    "authors": [
     "Matt G. Southern"
    ],
    "title": "Search Engine Journal report on Chartbeat's 2026 Publisher Playbook",
    "venue": "Search Engine Journal (primary: Chartbeat, 2026 Publisher Playbook, gated)",
    "doi": null,
    "urls": [
     "https://www.searchenginejournal.com/google-search-decline-publishers-chartbeat/590681/",
     "https://lp.chartbeat.com/2026-publisher-playbook-chartbeat"
    ]
   },
   "versions": [
    {
     "date": "2026-09-24",
     "kind": "news",
     "where": "Search Engine Journal",
     "url": "https://www.searchenginejournal.com/google-search-decline-publishers-chartbeat/590681/"
    }
   ],
   "finding": [
    {
     "quote": "a 40.2% year-over-year decline for Google Search",
     "from": "article, Search Engine Journal"
    },
    {
     "quote": "Google Discover’s year-over-year decline increased from 6.6% in the first year to 34.3% in the second",
     "from": "article, Search Engine Journal"
    }
   ],
   "shows": "A second year of falling Google Search referrals to publishers, steeper than the first, reported after the claims were registered.",
   "does_not_show": "The cause (AI Overviews are not separated); brands outside news; any breakdown by publisher size; a site count.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Search Engine Journal article and the NeoTeo article, and the Chartbeat landing page (which gives no figure; the report is gated), 2026-10-30. technology.org refused the fetch (robots).",
    "unverified": [
     "the Chartbeat report itself was not read",
     "the date of Chartbeat's own release was not found",
     "quote not found at source, dropped: \"Google Search accounted for about 9% of the network's pageviews in July 2024 and about 5% in July 2026.\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W05"
  },
  {
   "id": "W-0010",
   "direction": "supports",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-187.F1",
     "bears_on": "component",
     "scope": "The component that single observations of a generative surface mean little and that presence must be tracked apart from citation composition: an Overview changes between observations 70% of the time and nearly half its cited sources turn over, while its meaning holds. The same-day ≥3 sampling rule and the claim's watch query are not tested."
    },
    {
     "claim": "PRED-1438.F1",
     "bears_on": "mechanism",
     "scope": "Citation volatility on a fixed query, the quantity the Provenance Stability Index tracks. Ahrefs measures URL overlap between consecutive observations over a month, without the T+30/T+90/T+365 schedule or attribution density for a named source."
    },
    {
     "claim": "PRED-742.F1",
     "bears_on": "mechanism",
     "scope": "The fluctuation in composition behaviour that the claim's daily or weekly sampling over 30–90 days is meant to control. Ahrefs measures it; it does not run the claim's protocol."
    }
   ],
   "source": {
    "authors": [
     "Louise Linehan",
     "Xibeijia Guan"
    ],
    "title": "AI Overviews Change Every 2 Days (But Never Change Their Mind)",
    "venue": "Ahrefs blog (data from Ahrefs Brand Radar)",
    "doi": null,
    "urls": [
     "https://ahrefs.com/blog/ai-overview-change/"
    ]
   },
   "versions": [
    {
     "date": "2025-11-11",
     "kind": "post",
     "where": "Ahrefs blog",
     "url": "https://ahrefs.com/blog/ai-overview-change/"
    }
   ],
   "finding": [
    {
     "quote": "we found that AI Overviews have a 70% chance of changing from one observation to the next.",
     "from": "post, Ahrefs"
    }
   ],
   "shows": "Over 43,000 keywords with at least 16 recorded Overviews each: the text and the citations of an AI Overview churn every couple of days while its gist stays stable, so presence, content and citation composition move separately. It precedes all three claims and is an anticipation.",
   "does_not_show": "Any effect on a particular author or entity; same-day variation; differences by login state, device or location.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Ahrefs post (German-path URL served the English text), 2026-10-30.",
    "unverified": [
     "the exact check frequency (the post says checks were not daily)",
     "the method is a vendor's own, without peer review",
     "quote not found at source, dropped: \"nearly half (45.5%) of the cited sources are entirely new.\"",
     "quote not found at source, dropped: \"AI Overviews score 0.95 out of 1.0 for semantic consistency.\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W06"
  },
  {
   "id": "W-0011",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-742.F2",
     "bears_on": "method",
     "scope": "Trigger rate operationalized as the fraction of a query panel that yields an AI synthesis, measured by others: 13.7% overall on 55,393 queries. The panel is trending queries, without per-concept panels or the claim's other variables."
    },
    {
     "claim": "PRED-187.F1",
     "bears_on": "mechanism",
     "scope": "Overview presence depends on query format (13.7% overall, 64.7% for question-form queries), the format classes the claim's battery spans. The witness does not test the gate-stack model or the claim's watch query."
    },
    {
     "claim": "PRED-716.F1",
     "bears_on": "method",
     "scope": "Claim segmentation of composed answers checked against cited pages, at scale, on one system. It measures unsupported claims among cited pages, the converse of PER's uncited correct use."
    }
   ],
   "source": {
    "authors": [
     "Haofei Xu",
     "Umar Iqbal",
     "Jacob M. Montgomery"
    ],
    "title": "Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.14021",
     "https://arxiv.org/html/2605.14021v1"
    ]
   },
   "versions": [
    {
     "date": "2026-05-13",
     "kind": "preprint",
     "where": "arXiv 2605.14021 v1",
     "url": "https://arxiv.org/abs/2605.14021"
    },
    {
     "date": "2026-10-01",
     "kind": "preprint",
     "where": "arXiv 2605.14021 v2",
     "url": "https://arxiv.org/abs/2605.14021"
    }
   ],
   "finding": [
    {
     "quote": "issuing 55,393 trending queries across 19 topical categories over a 40-day window (March 13 - April 21, 2026).",
     "from": "abstract, arXiv"
    },
    {
     "quote": "overall AIO activation is 13.7%, rising to 64.7% for question-form queries,",
     "from": "abstract, arXiv"
    },
    {
     "quote": "decomposing responses into 98,020 atomic claims, 11.0% are unsupported by the cited pages",
     "from": "abstract, arXiv"
    },
    {
     "quote": "nearly 30% do not appear in those results at all",
     "from": "abstract, arXiv (AIO-cited domains against co-displayed first-page results)"
    }
   ],
   "shows": "A large audit of AI Overviews: activation is gated strongly by query form; cited domains are more credible than first-page results and often absent from them; 11% of atomic claims are unsupported by the pages cited. It came after PRED-716.F1 and before PRED-742.F2 and PRED-187.F1.",
   "does_not_show": "Attribution of a named author's concepts; any power skew in what is cited; stability across repeated queries.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the v1 date is the original submission date given on the page; no separate v1 line was shown"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W07"
  },
  {
   "id": "W-0012",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-209.F2",
     "bears_on": "mechanism",
     "scope": "The claim's YMYL sensitivity test asks whether sensitive phrasing triggers enhanced filtering; the witness finds markedly fewer Overviews on election queries than on non-political ones. Its topic is elections, with no economic phrases and no authority-tier analysis."
    },
    {
     "claim": "PRED-187.F1",
     "bears_on": "mechanism",
     "scope": "Overview presence gated by query type and inconsistent in its criteria, observed through Google's own researcher API under the DSA. The claim's sampling discipline is not tested."
    },
    {
     "claim": "PRED-166.A1",
     "bears_on": "component",
     "scope": "The component 'consolidate platform jurisdiction and reduce visibility of independent sources': Google's own YouTube was the most cited source and ten providers supplied nearly half of all links. It is one snapshot in July 2026, with no multi-year drift."
    }
   ],
   "source": {
    "authors": [
     "Clara Helming",
     "AlgorithmWatch"
    ],
    "title": "Meinungsbildung mit KI: Welche Informationen zeigen Googles KI-Übersichten zu Wahlen?",
    "venue": "AlgorithmWatch (study under DSA Article 40(12) data access)",
    "doi": null,
    "urls": [
     "https://algorithmwatch.org/de/google-ki-ubersichten-meinungsbildung/",
     "https://the-decoder.com/googles-election-ai-overviews-are-opaque-rely-on-few-sources-and-sometimes-take-sides/"
    ]
   },
   "versions": [
    {
     "date": "2026-09-01",
     "kind": "report",
     "where": "AlgorithmWatch website (German)",
     "url": "https://algorithmwatch.org/de/google-ki-ubersichten-meinungsbildung/"
    }
   ],
   "finding": [
    {
     "quote": "Bei Fragen zu Wahlen zeigt Google deutlich weniger KI-Übersichten an als bei Fragen zu unpolitischen Themen.",
     "from": "report page, AlgorithmWatch"
    },
    {
     "quote": "ist das meistzitierte Angebot in KI- Übersichten",
     "from": "report page, AlgorithmWatch (of YouTube)"
    },
    {
     "quote": "Fast die Hälfte aller angezeigten Links stammten von nur zehn verschiedenen Anbietern.",
     "from": "report page, AlgorithmWatch"
    },
    {
     "quote": "Unseren Beobachtungen nach ist dies sehr uneinheitlich.",
     "from": "report page, AlgorithmWatch (of the criteria by which Overviews appear)"
    }
   ],
   "shows": "4,480 queries on three 2026 German state elections, run in July 2026 through Google's Search Researcher Result API: Overviews appear far less on election questions, their appearance follows no consistent rule, and their sources concentrate, with YouTube first.",
   "does_not_show": "Economic or authority-tier effects; any trend over time; the archive's entities.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the AlgorithmWatch page (German) and The Decoder's English report, 2026-10-30.",
    "unverified": [
     "the figures 39.1% (election) and 65.3% (non-political) and the AfD trigger gap (24% against 45–58%) come from The Decoder as paraphrased by the fetch tool; the AlgorithmWatch page read does not print them",
     "the PDF study (code 260908) was not read"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W08"
  },
  {
   "id": "W-0013",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-166.A1",
     "bears_on": "component",
     "scope": "Directional drift toward the platform: Google's share of AI Mode citations rose from about 5% to about 17% in some nine months. The claim concerns stabilized public-knowledge nodes over years; the witness counts self-citation across commercial keywords on AI Mode."
    },
    {
     "claim": "PRED-791.F1",
     "bears_on": "mechanism",
     "scope": "Self-preferencing in the composition layer, the effect watch item (i) names. The claim's item concerns Google's own guidance on SEO, audit and AEO/GEO queries; the witness counts google.com citations across 20 industries."
    }
   ],
   "source": {
    "authors": [
     "Matt G. Southern",
     "SE Ranking"
    ],
    "title": "Google AI Mode cites itself more often (SE Ranking: Is Google stealing your clicks in AI Mode? 1.3M+ citations analyzed)",
    "venue": "Search Engine Journal, reporting an SE Ranking study",
    "doi": null,
    "urls": [
     "https://www.searchenginejournal.com/google-ai-mode-cites-itself-more-often-with-more-organic-links/568955/",
     "https://searchengineland.com/google-ai-mode-citing-google-more-study-471042",
     "https://seranking.com/blog/google-links-in-ai-mode-answers/"
    ]
   },
   "versions": [
    {
     "date": "2026-03-06",
     "kind": "news",
     "where": "Search Engine Journal",
     "url": "https://www.searchenginejournal.com/google-ai-mode-cites-itself-more-often-with-more-organic-links/568955/"
    }
   ],
   "finding": [
    {
     "quote": "Google properties accounted for 5% of AI Mode links.",
     "from": "article, Search Engine Journal (SE Ranking's July 2025 report)"
    }
   ],
   "shows": "Across 68,313 keywords and 1,321,398 AI Mode citations (February 2026), google.com was the most cited source, about three times its share nine months before. It precedes both claims and is an anticipation.",
   "does_not_show": "Drift on stabilized public-knowledge nodes; self-preferencing on SEO-guidance queries specifically; anything multi-year.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Search Engine Journal and Search Engine Land articles, 2026-10-30. The SE Ranking page returned empty content.",
    "unverified": [
     "the SE Ranking study itself was not read",
     "Search Engine Land's 'June 2025, 5.7%' baseline differs from SEJ's 'July, 5%'; the page date of the Search Engine Land article was not shown",
     "SE Ranking sells the tracking tool used",
     "quote not found at source, dropped: \"Google now accounts for 17% of all AI Mode citations.\"",
     "quote not found at source, dropped: \"Google.com was the most cited source in AI Mode answers, accounting for 17.42% of all citations\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W09"
  },
  {
   "id": "W-0014",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1643.B1",
     "bears_on": "mechanism",
     "scope": "The loop the claim names, in which machine text re-enters what the next answer is built from: generative search engines already cite AI-generated sources. The witness measures the share of such sources, without tracing lost authorship from one answer to the next."
    }
   ],
   "source": {
    "authors": [
     "Mowafak Allaham",
     "Nicholas Diakopoulos"
    ],
    "title": "Synthetic Sources?: Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.23684"
    ]
   },
   "versions": [
    {
     "date": "2026-05-22",
     "kind": "preprint",
     "where": "arXiv 2605.23684 v1",
     "url": "https://arxiv.org/abs/2605.23684"
    }
   ],
   "finding": [
    {
     "quote": "an audit of four generative search engines (ChatGPT, Copilot, Gemini, Perplexity)",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a total of 712 real-world human-generated queries",
     "from": "abstract, arXiv"
    },
    {
     "quote": "evidence of AI-generated sources being cited across all four generative search engines",
     "from": "abstract, arXiv"
    },
    {
     "quote": "(~16% of cited sources)",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "About a sixth of the sources cited by four generative search engines show evidence of being AI-generated. It precedes the claim and is an anticipation.",
   "does_not_show": "That those sources lack authors or provenance; degradation over successive answers; any causal loop.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the detection method and its error rate (full paper not read)"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W10"
  },
  {
   "id": "W-0015",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-615.B3",
     "bears_on": "mechanism",
     "scope": "A formal model in which generative extraction makes publishers opt out and lowers the corpus's volume, average quality and lifetime. It is proved in a model with no empirical data, and it says nothing about DOI-anchored material gaining authority."
    },
    {
     "claim": "PRED-1643.B1",
     "bears_on": "mechanism",
     "scope": "Extraction that degrades the corpus the next answer draws on, in a model of the crawlable commons. The model concerns funding and renewal of content; it does not model authorship."
    }
   ],
   "source": {
    "authors": [
     "Sylvain Peyronnet"
    ],
    "title": "When Search Eats the Web: A Model of Corpus Erosion under Generative Extraction",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2608.15896"
    ]
   },
   "versions": [
    {
     "date": "2026-08-16",
     "kind": "preprint",
     "where": "arXiv 2608.15896 v1",
     "url": "https://arxiv.org/abs/2608.15896"
    }
   ],
   "finding": [
    {
     "quote": "In this paper, we model the crawlable corpus as a common-pool resource: the crawlable commons.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "publishers opt out, renewal loses its funding, and content becomes more perishable",
     "from": "abstract, arXiv"
    },
    {
     "quote": "After a given erosion threshold, the corpus goes extinct.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A theoretical result: under generative extraction, publisher opt-out, unfunded renewal and perishability degrade the crawlable corpus together, with an extinction threshold. It came after PRED-615.B3 and before PRED-1643.B1.",
   "does_not_show": "Any measurement; AI Overview quality; which material gains authority as others exit.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "whether the full paper calibrates the model on data"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W11"
  },
  {
   "id": "W-0016",
   "direction": "disconfirms",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-615.B3",
     "bears_on": "component",
     "scope": "The component 'high-quality publishers withdraw from the summarizer's input layer': news sites blocking AI bots in robots.txt mostly still appear in AI citations, so the withdrawal the claim's chain begins with did not take effect for them. The claim's later links (quality decline, reweighting toward DOI-anchored material) are not tested."
    }
   ],
   "source": {
    "authors": [
     "Vince Nero"
    ],
    "title": "Do News Publishers That Block AI Crawlers Get Cited Less Often by AI?",
    "venue": "BuzzStream blog (citation data from Citation Labs' XOFU tool)",
    "doi": null,
    "urls": [
     "https://www.buzzstream.com/?p=11713"
    ]
   },
   "versions": [
    {
     "date": "2026-03-18",
     "kind": "post",
     "where": "BuzzStream blog (updated 2026-04-08)",
     "url": "https://www.buzzstream.com/?p=11713"
    }
   ],
   "finding": [
    {
     "quote": "Around 75% of sites blocking OpenAI or Google AI bots still appeared in AI citations.",
     "from": "post, BuzzStream"
    },
    {
     "quote": "This, as we'll see, isn't foolproof, and bots can still circumvent the robots.txt directive.",
     "from": "post, BuzzStream"
    }
   ],
   "shows": "About 4 million citations from 3,600 prompts across ChatGPT, Gemini, AI Overviews and AI Mode: most of the top 50 news sites that block OpenAI or Google AI bots are still cited. Blocking did not remove them from the input layer. It precedes the claim and is an anticipation.",
   "does_not_show": "Whether blocked content is still read or only cited from other copies; any change in answer quality; withdrawal by means other than robots.txt.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the BuzzStream post, 2026-10-30.",
    "unverified": [
     "the 70.6%–92.3% range by bot was given in the fetch tool's paraphrase",
     "the post's explanations are labelled by its author as educated guesses",
     "a vendor's own method, without peer review"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W12"
  },
  {
   "id": "W-0017",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-61.B2",
     "bears_on": "mechanism",
     "scope": "A single page on a personal site was re-composed into AI answers in under 24 hours, inside the claim's 'hours to days'. The claim's case is a DOI-anchored deposit acting as a definitional source; the witness is a fabricated claim on a personal site."
    },
    {
     "claim": "PRED-232.F3",
     "bears_on": "method",
     "scope": "The probe the claim specifies (introduce a novel artifact with known provenance and track what systems do with it), carried out once by a journalist. It records uptake and source display on one day; there is no tracking over time."
    },
    {
     "claim": "PRED-209.F8",
     "bears_on": "mechanism",
     "scope": "The same planted page treated differently by different systems: Gemini and AI Overviews repeated it without naming a source, ChatGPT repeated it, Claude did not. One case, with no test of whether systems inherit Google's guardrails."
    }
   ],
   "source": {
    "authors": [
     "Thomas Germain"
    ],
    "title": "I hacked ChatGPT and Google's AI - and it only took 20 minutes (Spanish edition: Hackeé ChatGPT y la IA de Google en 20 minutos (y qué logré demostrar con eso))",
    "venue": "BBC Future; Spanish edition by BBC News Mundo, republished by La Prensa",
    "doi": null,
    "urls": [
     "https://www.bbc.com/future/article/20260218-i-hacked-chatgpt-and-googles-ai-and-it-only-took-20-minutes",
     "https://www.prensa.com/bbc/hackee-chatgpt-y-la-ia-de-google-en-20-minutos-y-que-logre-demostrar-con-eso/"
    ]
   },
   "versions": [
    {
     "date": "2026-02-18",
     "kind": "news",
     "where": "BBC Future (date from the article URL)",
     "url": "https://www.bbc.com/future/article/20260218-i-hacked-chatgpt-and-googles-ai-and-it-only-took-20-minutes"
    },
    {
     "date": "2026-02-19",
     "kind": "news",
     "where": "BBC News Mundo edition, republished by La Prensa",
     "url": "https://www.prensa.com/bbc/hackee-chatgpt-y-la-ia-de-google-en-20-minutos-y-que-logre-demostrar-con-eso/"
    }
   ],
   "finding": [
    {
     "quote": "Menos de 24 horas después, los chatbots más importantes del mundo hablaban sin parar",
     "from": "article, BBC News Mundo edition (La Prensa)"
    },
    {
     "quote": "ChatGPT hizo lo mismo, aunque Claude, un chatbot de la empresa Anthropic, no se dejó engañar.",
     "from": "article, BBC News Mundo edition (La Prensa)"
    },
    {
     "quote": "Gemini no se molestó en decir de dónde obtuvo la información.",
     "from": "article, BBC News Mundo edition (La Prensa)"
    }
   ],
   "shows": "One planted page reached Gemini, AI Overviews and ChatGPT answers within a day; Gemini gave no source; the others linked it without saying it was the only source; Claude did not repeat it. It precedes PRED-61.B2 and comes after PRED-232.F3 and PRED-209.F8.",
   "does_not_show": "Any rate across cases; persistence; behaviour for DOI-anchored or scholarly material; the timing beyond the author's own account.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the BBC News Mundo edition at La Prensa and the Tildes link record, 2026-10-30. bbc.com refused the fetch (site blocked).",
    "unverified": [
     "the English original was not read; its date comes from its URL and the Tildes link record",
     "the under-24-hours timing is the journalist's own report",
     "quote not found at source, dropped: \"Todas las demás IA enlazaron a mi artículo, aunque rara vez mencionaron que yo era la única fuente.\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W13"
  },
  {
   "id": "W-0018",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-209.F8",
     "bears_on": "method",
     "scope": "Source selection compared across AI search providers on the same arena traffic, the comparison the claim assigns: models from different providers cite distinct news sources. It covers OpenAI, Perplexity and Google models, without Grok or Claude, and does not ask whether they inherit Google's guardrails."
    },
    {
     "claim": "PRED-232.F4",
     "bears_on": "mechanism",
     "scope": "Reading a system's political coordinates off the sources it cites by default: cited news shows a pronounced liberal lean. The claim reads coordinates off attributions of specific phrases; the witness reads them off news-source citations."
    }
   ],
   "source": {
    "authors": [
     "Kai-Cheng Yang"
    ],
    "title": "News Source Citing Patterns in AI Search Systems",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2507.05301v1"
    ]
   },
   "versions": [
    {
     "date": "2025-07-07",
     "kind": "preprint",
     "where": "arXiv 2507.05301 v1",
     "url": "https://arxiv.org/abs/2507.05301v1"
    }
   ],
   "finding": [
    {
     "quote": "The dataset comprises over 24,000 conversations and 65,000 responses",
     "from": "abstract, arXiv"
    },
    {
     "quote": "News citations concentrate heavily among a small number of outlets",
     "from": "abstract, arXiv"
    },
    {
     "quote": "models from different providers cite distinct news sources",
     "from": "abstract, arXiv"
    },
    {
     "quote": "display a pronounced liberal bias, though low-credibility sources are rarely cited",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "On AI Search Arena data (366,000+ citations), news citations concentrate in few outlets, differ by provider, and lean liberal. It precedes both claims and is an anticipation.",
   "does_not_show": "Inheritance of Google's index or filters; attribution of specific phrases; any trend.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v1 only listed), 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W14"
  },
  {
   "id": "W-0019",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-740.F3",
     "bears_on": "method",
     "scope": "The claim's replication design (the same queries across Copilot, Perplexity, ChatGPT search, Gemini and others), carried out by others for source identification of news excerpts. It does not measure CDI, disambiguation or entity suppression."
    },
    {
     "claim": "PRED-209.F8",
     "bears_on": "method",
     "scope": "The same queries compared across eight generative search tools, the claim's assembly task. It measures citation accuracy and does not test inheritance of Google's guardrails."
    },
    {
     "claim": "PRED-251.F6",
     "bears_on": "mechanism",
     "scope": "Provenance displaced in composed answers: tools sent users to syndicated copies on aggregators and to broken or fabricated URLs. PER counts uncited correct use of concepts; the witness counts wrong attribution of articles."
    },
    {
     "claim": "PRED-176.F2",
     "bears_on": "method",
     "scope": "Provenance under live generative search measured: given an excerpt, each tool was asked for headline, publisher, date and URL, and errors and citation to syndicated copies counted. News articles stand in for training-layer publications."
    },
    {
     "claim": "PRED-219.F1",
     "bears_on": "mechanism",
     "scope": "Systems that 'liquidate' attribution: the witness finds tools citing syndicated copies over the original publisher even with licensing deals. The essay itself is not tested."
    }
   ],
   "source": {
    "authors": [
     "Klaudia Jaźwińska",
     "Aisvarya Chandrasekar"
    ],
    "title": "AI Search Has a Citation Problem",
    "venue": "Columbia Journalism Review, Tow Center for Digital Journalism",
    "doi": null,
    "urls": [
     "https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php"
    ]
   },
   "versions": [
    {
     "date": "2025-03-06",
     "kind": "report",
     "where": "Columbia Journalism Review (Tow Center)",
     "url": "https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php"
    }
   ],
   "finding": [
    {
     "quote": "Collectively, they provided incorrect answers to more than 60 percent of queries.",
     "from": "report page, CJR"
    },
    {
     "quote": "directed us to syndicated versions of articles on platforms like Yahoo News or AOL",
     "from": "report page, CJR"
    },
    {
     "quote": "More than half of responses from Gemini and Grok 3 cited fabricated or broken URLs",
     "from": "findings, CJR article"
    },
    {
     "quote": "chatbots directed us to syndicated versions of articles on platforms like Yahoo News or AOL",
     "from": "findings, CJR article"
    }
   ],
   "shows": "1,600 manually scored queries (excerpts from 200 articles of 20 publishers) across eight tools: most answers misidentified the source, often crediting syndicated copies or giving broken links. It precedes all three claims and is an anticipation.",
   "does_not_show": "Uncited correct use of ideas; entity disambiguation; the archive's entities; any trend.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the CJR report page, 2026-10-30.",
    "unverified": [
     "quote not found at source, dropped: \"More than half of responses from Gemini and Grok 3 cited fabricated or broken URLs.\""
    ]
   },
   "found_by": [
    "S-0003",
    "S-0007"
   ],
   "reader_id": "A-attribution-W15"
  },
  {
   "id": "W-0020",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-716.F1",
     "bears_on": "method",
     "scope": "A multi-system audit with sentence-level judgement of whether each generated statement is supported by its citations, the kind of pilot the claim invites. It measures citation support, a near kin of PER; it does not score correct uncited use of a named source's concepts."
    },
    {
     "claim": "PRED-251.F6",
     "bears_on": "mechanism",
     "scope": "Composed statements left without supporting citation, the condition PER counts when the statement is correct. The witness does not separate correct from incorrect uncited statements."
    }
   ],
   "source": {
    "authors": [
     "Nelson F. Liu",
     "Tianyi Zhang",
     "Percy Liang"
    ],
    "title": "Evaluating Verifiability in Generative Search Engines",
    "venue": "arXiv (cs.CL, cs.IR)",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2304.09848"
    ]
   },
   "versions": [
    {
     "date": "2023-04-19",
     "kind": "preprint",
     "where": "arXiv 2304.09848 v1",
     "url": "https://arxiv.org/abs/2304.09848"
    },
    {
     "date": "2023-10-23",
     "kind": "preprint",
     "where": "arXiv 2304.09848 v2",
     "url": "https://arxiv.org/abs/2304.09848"
    }
   ],
   "finding": [
    {
     "quote": "on average, a mere 51.5% of generated sentences are fully supported by citations",
     "from": "abstract, arXiv"
    },
    {
     "quote": "only 74.5% of citations support their associated sentence",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Half of the sentences generative search engines produced in 2023 lacked full citation support. It precedes both claims by about three years and is an anticipation.",
   "does_not_show": "Whether the unsupported statements were correct uses of an identifiable source; current systems.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "publication in Findings of EMNLP 2023 (no journal-ref on the page read)",
     "quote not found at source, dropped: \"audit four popular generative search engines -- Bing Chat, NeevaAI, perplexity.ai, and YouChat\""
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W16"
  },
  {
   "id": "W-0021",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-302.B2",
     "bears_on": "mechanism",
     "scope": "A detector that asks which model's style produced a text, with near-perfect precision across Claude, Gemini, Llama and OpenAI models. The claim's 'architecture' in the ASDF frame is an authorial signature that can run through AI; the witness attributes text to model families."
    }
   ],
   "source": {
    "authors": [
     "Yehonatan Bitton",
     "Elad Bitton",
     "Shai Nisan"
    ],
    "title": "Detecting Stylistic Fingerprints of Large Language Models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2503.01659"
    ]
   },
   "versions": [
    {
     "date": "2025-03-03",
     "kind": "preprint",
     "where": "arXiv 2503.01659 v1",
     "url": "https://arxiv.org/abs/2503.01659"
    }
   ],
   "finding": [
    {
     "quote": "classify texts based on the stylistic fingerprints of the models that generated them",
     "from": "abstract, arXiv"
    },
    {
     "quote": "achieves extremely high precision (0.9988) and a very low false-positive rate (0.0004)",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Attribution of generated text to its source model family is practicable, a detection question beyond human-or-AI. It precedes the claim and is an anticipation.",
   "does_not_show": "Detection of a human author's architecture in AI-mediated text; that the human-or-AI question is becoming meaningless.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W17"
  },
  {
   "id": "W-0022",
   "direction": "supports",
   "evidence_class": "documentation",
   "claims": [
    {
     "claim": "PRED-61.B1",
     "bears_on": "component",
     "scope": "The component 'the retrieval layer develops from summarization to agentic commerce': Google announced an open agent-commerce protocol and checkout inside AI Mode. It does not show that entities without structured definitions become unreachable, or any claim about SPXI or GEO."
    }
   ],
   "source": {
    "authors": [
     "Vidhya Srinivasan"
    ],
    "title": "New tech and tools for retailers to succeed in an agentic shopping era",
    "venue": "Google, The Keyword blog",
    "doi": null,
    "urls": [
     "https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/"
    ]
   },
   "versions": [
    {
     "date": "2026-01-11",
     "kind": "documentation",
     "where": "Google blog",
     "url": "https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/"
    }
   ],
   "finding": [
    {
     "quote": "a new open standard for agentic commerce that works across the entire shopping journey",
     "from": "blog post, Google (of the Universal Commerce Protocol)"
    },
    {
     "quote": "UCP will soon power a new checkout feature on eligible Google product listings in AI Mode",
     "from": "blog post, Google"
    }
   ],
   "shows": "The platform itself moved its AI search surface toward transactions executed by agents three months before the claim. It is an anticipation. It is an announcement and measures nothing.",
   "does_not_show": "Reachability effects for unstructured entities; adoption or use; the claim's nested-layer argument.",
   "checked": {
    "by": "intake reader, theme A-attribution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Google blog post, 2026-10-30.",
    "unverified": [
     "the post does not state that structured data is required for agents to transact"
    ]
   },
   "found_by": [
    "S-0003"
   ],
   "reader_id": "A-attribution-W18"
  },
  {
   "id": "W-0023",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-857.F1",
     "bears_on": "component",
     "scope": "The state the longitudinal study would track, whether the model ecology converges, measured at one time across many models on open-ended prompts. PRED-857.F1 asks for repeated runs of one directive on five named substrates at six-month intervals; this is a single cross-section on a different prompt set."
    },
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "Convergence of models toward similar outputs, shown across models. The Gray Hypothesis names safety pressure as the driver and loss of nuance as the effect; the study measures similarity of open-ended outputs and does not isolate safety training."
    }
   ],
   "source": {
    "authors": [
     "Liwei Jiang",
     "Yuanjun Chai",
     "Margaret Li",
     "Mickel Liu",
     "Raymond Fok",
     "Nouha Dziri",
     "Yulia Tsvetkov",
     "Maarten Sap",
     "Alon Albalak",
     "Yejin Choi"
    ],
    "title": "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)",
    "venue": "arXiv; listed at NeurIPS 2025",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.22954",
     "https://neurips.cc/virtual/2025/poster/121421"
    ]
   },
   "versions": [
    {
     "date": "2025-10-27",
     "kind": "preprint",
     "where": "arXiv 2510.22954 v1",
     "url": "https://arxiv.org/abs/2510.22954v1"
    },
    {
     "date": "2026-10-06",
     "kind": "preprint",
     "where": "arXiv 2510.22954 v2",
     "url": "https://arxiv.org/abs/2510.22954"
    }
   ],
   "finding": [
    {
     "quote": "Using Infinity-Chat, we present a large-scale study of mode collapse in LMs,",
     "from": "abstract, arXiv"
    },
    {
     "quote": "characterized by (1) intra-model repetition, where a single model consistently generates similar responses,",
     "from": "abstract, arXiv"
    },
    {
     "quote": "and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a large-scale dataset of 26K diverse, real-world, open-ended user queries",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "On 26K real open-ended queries, single models repeat themselves and different models give strikingly similar outputs.",
   "does_not_show": "A trend over time; the five Assembly substrates or the Three Pillars directive; safety training as the cause.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (submission history and abstract), 2026-10-30.",
    "unverified": [
     "NeurIPS 2025 acceptance: seen only as a search-result listing on neurips.cc, page not fetched"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W01"
  },
  {
   "id": "W-0024",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-857.F1",
     "bears_on": "method",
     "scope": "The proposed longitudinal measurement of solution-space diversity across the model ecology, carried out by others across three years of releases with sentence-embedding similarity. It uses Infinity-Chat100 and the Alternate Uses Task in place of the Three Pillars directive and the five named substrates."
    },
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "Models converging over time, shown across releases. Safety pressure as the driver and nuance as the quantity lost are not tested."
    }
   ],
   "source": {
    "authors": [
     "Nirav Patel",
     "Josiah Crossman",
     "Eva Aggarwal",
     "Emily Wenger"
    ],
    "title": "Are LLMs becoming similarly creative? Evidence from three years of models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2608.19437",
     "https://arxiv.org/html/2608.19437v1"
    ]
   },
   "versions": [
    {
     "date": "2026-08-19",
     "kind": "preprint",
     "where": "arXiv 2608.19437 v1",
     "url": "https://arxiv.org/abs/2608.19437"
    }
   ],
   "finding": [
    {
     "quote": "a statistically significant decrease in model output diversity over time",
     "from": "abstract, arXiv"
    },
    {
     "quote": "LLM outputs may be converging in creative substance across models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "This covers 33 closed weight models and 34 open weight models across 27 release months",
     "from": "body, arXiv HTML v1"
    }
   ],
   "shows": "Across 67 models released over 27 months, cross-model output diversity on creative prompts falls significantly with release date.",
   "does_not_show": "The cause of the decline; the archive's substrates or directive; the authors note their regression cannot account for hidden relationships across model families.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the HTML v1 (model count and limitations), 2026-10-30.",
    "unverified": [
     "model count: the HTML also reports 68 and 69 models in other places"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W02"
  },
  {
   "id": "W-0025",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-857.F1",
     "bears_on": "method",
     "scope": "A longitudinal measure of output diversity across successive releases, carried out on ChatGPT models with a text-similarity metric. One vendor's line only; the claim's study spans five substrates on one directive."
    },
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "Successive releases of one model line lose the ability to vary their text. Safety pressure as the cause is not tested."
    }
   ],
   "source": {
    "authors": [
     "Konstantinos F. Xylogiannopoulos",
     "Petros Xanthopoulos",
     "Panagiotis Karampelas",
     "Georgios A. Bakamitsos"
    ],
    "title": "Experimental evidence of progressive ChatGPT models self-convergence",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2603.12683"
    ]
   },
   "versions": [
    {
     "date": "2026-03-13",
     "kind": "preprint",
     "where": "arXiv 2603.12683 v1",
     "url": "https://arxiv.org/abs/2603.12683"
    },
    {
     "date": "2026-03-16",
     "kind": "preprint",
     "where": "arXiv 2603.12683 v2",
     "url": "https://arxiv.org/abs/2603.12683"
    }
   ],
   "finding": [
    {
     "quote": "we employ a text similarity metric to evaluate different ChatGPT models' capacity to generate diverse textual outputs.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Our findings indicate a measurable decline of recent ChatGPT releases' ability to produce varied text",
     "from": "abstract, arXiv"
    },
    {
     "quote": "even when explicitly prompted to do so, by setting the temperature parameter to one",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Recent ChatGPT releases produce less varied text than earlier ones, even at temperature one and when asked for variety.",
   "does_not_show": "Cross-vendor convergence; any cause; effect sizes (not read beyond the abstract).",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W03"
  },
  {
   "id": "W-0026",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-857.F1",
     "bears_on": "component",
     "scope": "Cross-model convergence at one time point: LLM responses resemble each other more than human responses do. No longitudinal series and not the claim's substrates or directive."
    }
   ],
   "source": {
    "authors": [
     "Emily Wenger",
     "Yoed Kenett"
    ],
    "title": "We're Different, We're the Same: Creative Homogeneity Across LLMs",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2501.19361"
    ]
   },
   "versions": [
    {
     "date": "2025-01-31",
     "kind": "preprint",
     "where": "arXiv 2501.19361 v1",
     "url": "https://arxiv.org/abs/2501.19361"
    }
   ],
   "finding": [
    {
     "quote": "LLM responses are much more similar to other LLM responses than human responses are to each other",
     "from": "abstract, arXiv"
    },
    {
     "quote": "even after controlling for response structure and other key variables",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "The population of LLM responses to creative tasks is more homogeneous than the population of human responses, with structure controlled.",
   "does_not_show": "Change over time; the cause.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W04"
  },
  {
   "id": "W-0027",
   "direction": "disconfirms",
   "evidence_class": "peer-reviewed conference paper",
   "claims": [
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "The cause the Gray Hypothesis names. The paper locates output homogeneity in pretraining, with alignment revealing or magnifying it and SFT data unable to introduce it; this tells against safety pressure as the origin, and leaves standing the role of alignment as an amplifier. Loss of nuance as such is not measured."
    }
   ],
   "source": {
    "authors": [
     "Alexandrine Fortier",
     "Hazel Chen",
     "Peter West"
    ],
    "title": "Is Convergence Inevitable? Tracing Output Homogeneity Back to Base Models",
    "venue": "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Budapest, pp. 36846–36865",
    "doi": "10.18653/2026.emnlp-main.1621",
    "urls": [
     "https://aclanthology.org/2026.emnlp-main.1621/",
     "https://arxiv.org/abs/2608.11426"
    ]
   },
   "versions": [
    {
     "date": "2026-08-11",
     "kind": "preprint",
     "where": "arXiv 2608.11426 v1",
     "url": "https://arxiv.org/abs/2608.11426"
    },
    {
     "date": "2026-08-13",
     "kind": "preprint",
     "where": "arXiv 2608.11426 v2",
     "url": "https://arxiv.org/abs/2608.11426"
    },
    {
     "date": "2026-10",
     "kind": "peer-reviewed conference paper",
     "where": "EMNLP 2026",
     "url": "https://aclanthology.org/2026.emnlp-main.1621/"
    }
   ],
   "finding": [
    {
     "quote": "output homogeneity is likely learned during the pretraining phase",
     "from": "abstract, ACL Anthology"
    },
    {
     "quote": "only revealed or magnified during the alignment process",
     "from": "abstract, ACL Anthology"
    },
    {
     "quote": "convergence can be revealed and amplified, but not introduced by the SFT data",
     "from": "abstract, ACL Anthology"
    },
    {
     "quote": "instruct-like collapse can be induced through prompting alone, even without alignment",
     "from": "abstract, ACL Anthology"
    }
   ],
   "shows": "Homogeneity of instruct-model outputs can be produced from base models by prompting alone; alignment amplifies a convergence learned in pretraining.",
   "does_not_show": "That alignment has no effect: the paper finds it magnifies convergence. Safety training is not separated from other alignment.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the ACL Anthology landing page and the arXiv abstract page (submission history), 2026-10-30.",
    "unverified": [
     "the ACL page's quotes came through the fetch summarizer; the arXiv abstract text was not compared word for word"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W05"
  },
  {
   "id": "W-0028",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-107.G1",
     "bears_on": "mechanism",
     "scope": "Post-training convergence pressure measured as reduced output diversity. The study compares RLHF with SFT across diversity measures; it does not compute signed per-token deviation and does not use Llama-3.1 base and instruct."
    },
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "Preference training reduces diversity. Safety pressure as such and nuance are not isolated."
    }
   ],
   "source": {
    "authors": [
     "Robert Kirk",
     "Ishita Mediratta",
     "Christoforos Nalmpantis",
     "Jelena Luketina",
     "Eric Hambro",
     "Edward Grefenstette",
     "Roberta Raileanu"
    ],
    "title": "Understanding the Effects of RLHF on LLM Generalisation and Diversity",
    "venue": "arXiv; listed in ICLR 2024 proceedings",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2310.06452v3",
     "https://proceedings.iclr.cc/paper_files/paper/2024/hash/5a68d05006d5b05dd9463dd9c0219db0-Abstract-Conference.html"
    ]
   },
   "versions": [
    {
     "date": "2023-10-10",
     "kind": "preprint",
     "where": "arXiv 2310.06452 v1",
     "url": "https://arxiv.org/abs/2310.06452v1"
    },
    {
     "date": "2024-01-03",
     "kind": "preprint",
     "where": "arXiv 2310.06452 v2",
     "url": "https://arxiv.org/abs/2310.06452v2"
    },
    {
     "date": "2024-02-19",
     "kind": "preprint",
     "where": "arXiv 2310.06452 v3",
     "url": "https://arxiv.org/abs/2310.06452v3"
    }
   ],
   "finding": [
    {
     "quote": "However, RLHF significantly reduces output diversity compared to SFT across a variety of measures",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "RLHF lowers output diversity relative to SFT across several measures.",
   "does_not_show": "The signed-deviation statistic; base versus instruct of the named model.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (full submission history), 2026-10-30.",
    "unverified": [
     "ICLR 2024 publication: seen only as a search-result listing, proceedings page not fetched"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W06"
  },
  {
   "id": "W-0029",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-107.G1",
     "bears_on": "mechanism",
     "scope": "The direction G1 predicts, base versus aligned, shown on the model's own next-token distribution: alignment sharpens it and cuts the branching factor 2 to 5 times. This is an entropy measure; G1's statistic is the signed deviation of realized tokens from that expectation, which the study does not report."
    }
   ],
   "source": {
    "authors": [
     "Chenghao Yang",
     "Sida Li",
     "Ari Holtzman"
    ],
    "title": "LLM Probability Concentration: How Alignment Shrinks the Generative Horizon",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2506.17871"
    ]
   },
   "versions": [
    {
     "date": "2025-06-22",
     "kind": "preprint",
     "where": "arXiv 2506.17871 v1",
     "url": "https://arxiv.org/abs/2506.17871"
    },
    {
     "date": "2026-09-12",
     "kind": "preprint",
     "where": "arXiv 2506.17871 v4",
     "url": "https://arxiv.org/abs/2506.17871"
    }
   ],
   "finding": [
    {
     "quote": "Alignment tuning sharpens the output distribution from the outset, reducing BF by a factor of 2--5 overall",
     "from": "abstract, arXiv"
    },
    {
     "quote": "up to an order of magnitude (e.g., from 12 to 1.2) at early positions",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Aligned models concentrate probability on far fewer continuations than their base models.",
   "does_not_show": "Signed deviation; a Mann-Whitney test on matched prompts; Llama-3.1-8B specifically.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "dates of v2 and v3"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W07"
  },
  {
   "id": "W-0030",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-107.G1",
     "bears_on": "mechanism",
     "scope": "Base models outperform their aligned forms on randomness and creative originality, with aligned models nearly predictable in some states. A behavioral measure; the per-token signed-deviation statistic is not computed."
    }
   ],
   "source": {
    "authors": [
     "Peter West",
     "Christopher Potts"
    ],
    "title": "Base Models Beat Aligned Models at Randomness and Creativity",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2505.00047"
    ]
   },
   "versions": [
    {
     "date": "2025-04-30",
     "kind": "preprint",
     "where": "arXiv 2505.00047 v1",
     "url": "https://arxiv.org/abs/2505.00047"
    },
    {
     "date": "2025-09-13",
     "kind": "preprint",
     "where": "arXiv 2505.00047 v2",
     "url": "https://arxiv.org/abs/2505.00047"
    }
   ],
   "finding": [
    {
     "quote": "base language models consistently outperform their popular aligned forms",
     "from": "abstract, arXiv"
    },
    {
     "quote": "aligned models tend towards narrow behaviors that result in distinct disadvantages",
     "from": "abstract, arXiv"
    },
    {
     "quote": "becoming almost fully predictable in some game states",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Alignment narrows behavior and raises predictability relative to the base model.",
   "does_not_show": "The deviation statistic; any safety-specific effect.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W08"
  },
  {
   "id": "W-0031",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-803.B2",
     "bears_on": "mechanism",
     "scope": "Alignment for debiasing, a safety-adjacent pressure, yields lower token entropy and attractor states in Llama-2 variants. Close to the Gray Hypothesis's cause; its effect is measured as diversity, which is narrower than nuance."
    },
    {
     "claim": "PRED-107.G1",
     "bears_on": "mechanism",
     "scope": "Aligned models show lower token-prediction entropy than their base. Entropy is the expectation term; G1 names the signed deviation from it, which the study does not compute."
    }
   ],
   "source": {
    "authors": [
     "Behnam Mohammadi"
    ],
    "title": "Creativity Has Left the Chat: The Price of Debiasing Language Models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2406.05587"
    ]
   },
   "versions": [
    {
     "date": "2024-06-08",
     "kind": "preprint",
     "where": "arXiv 2406.05587 v1",
     "url": "https://arxiv.org/abs/2406.05587"
    }
   ],
   "finding": [
    {
     "quote": "aligned models exhibit lower entropy in token predictions",
     "from": "abstract, arXiv"
    },
    {
     "quote": "form distinct clusters in the embedding space",
     "from": "abstract, arXiv"
    },
    {
     "quote": "attractor states",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Alignment that reduces bias and toxicity also lowers output entropy and pulls outputs toward a few attractor states.",
   "does_not_show": "Frontier models; human-rated loss of nuance.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W09"
  },
  {
   "id": "W-0032",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A7",
     "bears_on": "component",
     "scope": "Part (1): training only on generated data across generations destroys the tails of the original distribution. Parts (2) to (4), the mixing threshold and the rare-token retention rates, are not quantified here."
    },
    {
     "claim": "PRED-635.F1",
     "bears_on": "method",
     "scope": "The collapse-resistance protocol, train on a corpus, generate, retrain on the generation, repeat, carried out by others before the claim. Their corpus, generation count and measures differ from the claim's Wikipedia, ten-generation design."
    },
    {
     "claim": "PRED-1573.F4",
     "bears_on": "method",
     "scope": "Deliberately producing recursive collapse on small models and watching its measures evolve. They do not calibrate a composite index or severity bands."
    }
   ],
   "source": {
    "authors": [
     "Ilia Shumailov",
     "Zakhar Shumaylov",
     "Yiren Zhao",
     "Yarin Gal",
     "Nicolas Papernot",
     "Ross Anderson"
    ],
    "title": "The Curse of Recursion: Training on Generated Data Makes Models Forget",
    "venue": "arXiv; later published in Nature (2024) as 'AI models collapse when trained on recursively generated data'",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2305.17493",
     "https://www.nature.com/articles/s41586-024-07566-y"
    ]
   },
   "versions": [
    {
     "date": "2023-05-27",
     "kind": "preprint",
     "where": "arXiv 2305.17493 v1",
     "url": "https://arxiv.org/abs/2305.17493v1"
    },
    {
     "date": "2023-05-31",
     "kind": "preprint",
     "where": "arXiv 2305.17493 v2",
     "url": "https://arxiv.org/abs/2305.17493v2"
    },
    {
     "date": "2024-04-14",
     "kind": "preprint",
     "where": "arXiv 2305.17493 v3",
     "url": "https://arxiv.org/abs/2305.17493v3"
    }
   ],
   "finding": [
    {
     "quote": "causes irreversible defects in the resulting models, where tails of the original content distribution disappear",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Recursive training on model output causes irreversible loss of the distribution's tails.",
   "does_not_show": "A critical real-data fraction; per-generation rare-token retention numbers.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and submission history, 2026-10-30. The Nature page refused the fetch.",
    "unverified": [
     "the Nature article's title, date and DOI: page refused the fetch"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W10"
  },
  {
   "id": "W-0033",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A7",
     "bears_on": "component",
     "scope": "Part (2), that real data in the mix prevents collapse, in the accumulation regime: real data kept alongside each generation's synthetic data. The claim's fixed mixing fraction α and its critical value are a different parametrization."
    }
   ],
   "source": {
    "authors": [
     "Matthias Gerstgrasser",
     "Rylan Schaeffer",
     "Apratim Dey",
     "Rafael Rafailov",
     "Henry Sleight",
     "John Hughes",
     "Tomasz Korbak",
     "Rajashree Agrawal",
     "Dhruv Pai",
     "Andrey Gromov",
     "Daniel A. Roberts",
     "Diyi Yang",
     "David L. Donoho",
     "Sanmi Koyejo"
    ],
    "title": "Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2404.01413"
    ]
   },
   "versions": [
    {
     "date": "2024-04-01",
     "kind": "preprint",
     "where": "arXiv 2404.01413 v1",
     "url": "https://arxiv.org/abs/2404.01413v1"
    },
    {
     "date": "2024-04-29",
     "kind": "preprint",
     "where": "arXiv 2404.01413 v2",
     "url": "https://arxiv.org/abs/2404.01413"
    }
   ],
   "finding": [
    {
     "quote": "replacing the original real data by each generation's synthetic data does indeed tend towards model collapse",
     "from": "abstract, arXiv"
    },
    {
     "quote": "accumulating the successive generations of synthetic data alongside the original real data avoids model collapse",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Replacing real data collapses the model; accumulating real and synthetic data avoids collapse, with bounded test error proved in a linear setting.",
   "does_not_show": "A threshold α*; rare-token retention.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W11"
  },
  {
   "id": "W-0034",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A7",
     "bears_on": "component",
     "scope": "Parts (2) and (3): stability of iterative retraining holds once the proportion of clean data is large enough, which implies a threshold. The paper proves the condition for generative models in general and gives no value to compare with the claim's range of 0.05 to 0.15."
    }
   ],
   "source": {
    "authors": [
     "Quentin Bertrand",
     "Avishek Joey Bose",
     "Alexandre Duplessis",
     "Marco Jiralerspong",
     "Gauthier Gidel"
    ],
    "title": "On the Stability of Iterative Retraining of Generative Models on their own Data",
    "venue": "arXiv; listed at ICLR 2024",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2310.00429v5",
     "https://iclr.cc/virtual/2024/poster/18942"
    ]
   },
   "versions": [
    {
     "date": "2023-09-30",
     "kind": "preprint",
     "where": "arXiv 2310.00429 v1",
     "url": "https://arxiv.org/abs/2310.00429v1"
    },
    {
     "date": "2024-04-02",
     "kind": "preprint",
     "where": "arXiv 2310.00429 v5",
     "url": "https://arxiv.org/abs/2310.00429v5"
    }
   ],
   "finding": [
    {
     "quote": "We first prove the stability of iterative training under the condition that the initial generative models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "approximate the data distribution well enough and the proportion of clean training data",
     "from": "abstract, arXiv"
    },
    {
     "quote": "(w.r.t. synthetic data) is large enough.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Iterative retraining is stable when the share of clean data is large enough.",
   "does_not_show": "The numerical threshold; language models specifically.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv v5 abstract page, 2026-10-30.",
    "unverified": [
     "ICLR 2024 acceptance: search listing only",
     "dates of v2 to v4"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W12"
  },
  {
   "id": "W-0035",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A7",
     "bears_on": "component",
     "scope": "Parts (1) and (3): collapse is unavoidable on synthetic data alone, and a maximal synthetic share exists below which it can be avoided, derived for a statistical language model. Whether its estimate falls in the claim's α* range was not read beyond the abstract."
    }
   ],
   "source": {
    "authors": [
     "Mohamed El Amine Seddik",
     "Suei-Wen Chen",
     "Soufiane Hayou",
     "Pierre Youssef",
     "Merouane Debbah"
    ],
    "title": "How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2404.05090"
    ]
   },
   "versions": [
    {
     "date": "2024-04-07",
     "kind": "preprint",
     "where": "arXiv 2404.05090 v1",
     "url": "https://arxiv.org/abs/2404.05090"
    }
   ],
   "finding": [
    {
     "quote": "Specifically, we demonstrate that model collapse cannot be avoided when training solely on synthetic data.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "However, when mixing both real and synthetic data, we provide an estimate of a maximal amount of synthetic data",
     "from": "abstract, arXiv"
    },
    {
     "quote": "below which model collapse can eventually be avoided.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A threshold structure for collapse in language-model training: certain on synthetic data alone, avoidable below a maximal synthetic share.",
   "does_not_show": "The numerical value of the threshold in the claim's terms.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the paper's numerical estimate of the threshold"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W13"
  },
  {
   "id": "W-0036",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A7",
     "bears_on": "component",
     "scope": "Parts (1) and (4) in part: lexical, syntactic and semantic diversity of outputs fall through successive recursive fine-tuning. The claim's rare-token retention rate per generation and its α conditions are not measured."
    },
    {
     "claim": "PRED-635.F1",
     "bears_on": "method",
     "scope": "Measuring output diversity at each generation of a recursive train-and-generate loop, carried out by others with fine-tuning on English generation tasks; the claim specifies pretraining on Wikipedia for ten generations."
    },
    {
     "claim": "PRED-1573.F4",
     "bears_on": "method",
     "scope": "Producing recursive collapse on models and tracking diversity measures across iterations. No severity bands are calibrated."
    }
   ],
   "source": {
    "authors": [
     "Yanzhu Guo",
     "Guokan Shang",
     "Michalis Vazirgiannis",
     "Chloé Clavel"
    ],
    "title": "The Curious Decline of Linguistic Diversity: Training Language Models on Synthetic Text",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2311.09807"
    ]
   },
   "versions": [
    {
     "date": "2023-11-16",
     "kind": "preprint",
     "where": "arXiv 2311.09807 v1",
     "url": "https://arxiv.org/abs/2311.09807v1"
    },
    {
     "date": "2024-04-16",
     "kind": "preprint",
     "where": "arXiv 2311.09807 v2",
     "url": "https://arxiv.org/abs/2311.09807"
    }
   ],
   "finding": [
    {
     "quote": "a set of novel metrics targeting lexical, syntactic, and semantic diversity",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a consistent decrease in the diversity of the model outputs through successive iterations",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Diversity on several linguistic levels falls consistently across recursive training iterations.",
   "does_not_show": "A threshold for real-data mixing; the claim's case-3 classification.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "publication venue (believed NAACL 2024 Findings): not on the page"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W14"
  },
  {
   "id": "W-0037",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-110.A1",
     "bears_on": "method",
     "scope": "The test P1 proposes, a per-token signed deviation of observed log-probability from the model's expected log-probability, normalized, used to separate machine from human text. Fast-DetectGPT builds a detector on exactly this curvature and reports it works; it reports detection accuracy and does not report P1's Mann-Whitney U or Cohen's d."
    }
   ],
   "source": {
    "authors": [
     "Guangsheng Bao",
     "Yanbin Zhao",
     "Zhiyang Teng",
     "Linyi Yang",
     "Yue Zhang"
    ],
    "title": "Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2310.05130"
    ]
   },
   "versions": [
    {
     "date": "2023-10-08",
     "kind": "preprint",
     "where": "arXiv 2310.05130 v1",
     "url": "https://arxiv.org/abs/2310.05130v1"
    },
    {
     "date": "2024-12-16",
     "kind": "preprint",
     "where": "arXiv 2310.05130 v3",
     "url": "https://arxiv.org/abs/2310.05130"
    }
   ],
   "finding": [
    {
     "quote": "introduce the concept of conditional probability curvature",
     "from": "abstract, arXiv"
    },
    {
     "quote": "discrepancies in word choices between LLMs and humans within a given context",
     "from": "abstract, arXiv"
    },
    {
     "quote": "surpasses DetectGPT by a relative around 75% in both the white-box and black-box settings",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Machine text sits on the probable side of the scoring model's own expectation token by token, consistently enough to detect it zero-shot.",
   "does_not_show": "Human-rated slop (PRED-107.G3); the claim's thresholds.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page twice, 2026-10-30.",
    "unverified": [
     "the curvature's exact definition (sign convention, normalization) is in the body, which was not read",
     "ICLR 2024 publication: not on the page"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W15"
  },
  {
   "id": "W-0038",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-110.A1",
     "bears_on": "component",
     "scope": "On the matched human/ChatGPT corpus HC3, ChatGPT text has lower GPT-2 perplexity than human text at text and sentence level. Perplexity is raw surprisal; P1's statistic subtracts the model's expectation, so the fit is to the direction of the effect."
    },
    {
     "claim": "PRED-110.F2",
     "bears_on": "method",
     "scope": "The roadmap's cheapest test names HC3; the corpus authors already ran a perplexity comparison on it. Not with P1's deviation statistic or three reference models."
    }
   ],
   "source": {
    "authors": [
     "Biyang Guo",
     "Xin Zhang",
     "Ziyuan Wang",
     "Minqi Jiang",
     "Jinran Nie",
     "Yuxuan Ding",
     "Jianwei Yue",
     "Yupeng Wu"
    ],
    "title": "How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2301.07597",
     "https://arxiv.org/html/2301.07597"
    ]
   },
   "versions": [
    {
     "date": "2023-01-18",
     "kind": "preprint",
     "where": "arXiv 2301.07597 v1",
     "url": "https://arxiv.org/abs/2301.07597"
    }
   ],
   "finding": [
    {
     "quote": "We call the collected dataset the Human ChatGPT Comparison Corpus (HC3).",
     "from": "abstract, arXiv"
    },
    {
     "quote": "It is clearly observed that, regardless of whether it is at the text level or the sentence level,",
     "from": "Section 4.4, arXiv HTML"
    },
    {
     "quote": "the content generated by ChatGPT has relatively lower PPLs compared to the text written by humans.",
     "from": "Section 4.4, arXiv HTML"
    }
   ],
   "shows": "In HC3, ChatGPT answers are more predictable to a reference model than human answers.",
   "does_not_show": "A signed-deviation statistic; effect size; human-rated slop.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the HTML full text Section 4.4, 2026-10-30.",
    "unverified": [
     "the HTML fetched was the unversioned page; only v1 is listed"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W16"
  },
  {
   "id": "W-0039",
   "direction": "supports",
   "evidence_class": "peer-reviewed conference paper",
   "claims": [
    {
     "claim": "PRED-199.A1",
     "bears_on": "mechanism",
     "scope": "The monoculture premise: model-generated code shows low algorithmic diversity against human solutions, with model size, temperature and instruction tuning as factors. Cross-sectional; no trend across generations under synthetic saturation."
    },
    {
     "claim": "PRED-199.F1",
     "bears_on": "method",
     "scope": "Measuring the solution-space diversity of a code model against a test battery, carried out by others with their own metric. One time point; the SSDI and the generational series are not run."
    }
   ],
   "source": {
    "authors": [
     "Seonghyeon Lee",
     "HeeJae Chon",
     "Joonwon Jang",
     "Dongha Lee",
     "Hwanjo Yu"
    ],
    "title": "How Diversely Can Language Models Solve Problems? Exploring the Algorithmic Diversity of Model-Generated Code",
    "venue": "Findings of the Association for Computational Linguistics: EMNLP 2025, Suzhou, pp. 152–167",
    "doi": "10.18653/v1/2025.findings-emnlp.10",
    "urls": [
     "https://aclanthology.org/2025.findings-emnlp.10/",
     "https://arxiv.org/abs/2503.00691v2"
    ]
   },
   "versions": [
    {
     "date": "2025-03-02",
     "kind": "preprint",
     "where": "arXiv 2503.00691 v1",
     "url": "https://arxiv.org/abs/2503.00691v1"
    },
    {
     "date": "2025-03-07",
     "kind": "preprint",
     "where": "arXiv 2503.00691 v2",
     "url": "https://arxiv.org/abs/2503.00691v2"
    },
    {
     "date": "2025-11",
     "kind": "peer-reviewed conference paper",
     "where": "Findings of EMNLP 2025",
     "url": "https://aclanthology.org/2025.findings-emnlp.10/"
    }
   ],
   "finding": [
    {
     "quote": "We extensively investigate the property of model-generated solutions by contrasting them with human-written ones",
     "from": "abstract, ACL Anthology"
    },
    {
     "quote": "model-generated solutions exhibit low algorithmic diversity",
     "from": "abstract, ACL Anthology"
    }
   ],
   "shows": "Code models converge on fewer algorithms than human programmers for the same problems.",
   "does_not_show": "Decline across model generations; vulnerability correlation.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the ACL Anthology landing page (preview mirror) and the arXiv submission history, 2026-10-30.",
    "unverified": [
     "the Findings publication month (EMNLP 2025 was held November 2025)"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W17"
  },
  {
   "id": "W-0040",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-199.A2",
     "bears_on": "component",
     "scope": "The correlation half for one model: LLM output follows recurring templates, and a model reproduces the same vulnerabilities across different programs, enough to attack black-box software at up to 94% success. Change in vulnerability frequency, and growth of correlation as AI code share rises, are not measured."
    }
   ],
   "source": {
    "authors": [
     "Tomer Kordonsky",
     "Amit LeVi",
     "Maayan Yamin",
     "Noam Benzimra",
     "Avi Mendelson"
    ],
    "title": "Extracting Recurring Vulnerabilities from Black-Box LLM-Generated Software",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2602.04894"
    ]
   },
   "versions": [
    {
     "date": "2026-02-02",
     "kind": "preprint",
     "where": "arXiv 2602.04894 v1",
     "url": "https://arxiv.org/abs/2602.04894v1"
    },
    {
     "date": "2026-06-05",
     "kind": "preprint",
     "where": "arXiv 2602.04894 v4",
     "url": "https://arxiv.org/abs/2602.04894"
    }
   ],
   "finding": [
    {
     "quote": "often follow recurring templates that can induce predictable vulnerabilities",
     "from": "abstract, arXiv"
    },
    {
     "quote": "how consistently a model reproduces the same vulnerabilities across programs",
     "from": "abstract, arXiv"
    },
    {
     "quote": "achieves up to 94% attack success and 93% vulnerability coverage on Internal Tools (Claude-4.5 Opus)",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Vulnerabilities in LLM-generated software recur across programs from one model and can be predicted and exploited from that regularity.",
   "does_not_show": "A trend over time; the frequency side of the claim; cross-model correlation.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "dates of v2 and v3"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W18"
  },
  {
   "id": "W-0041",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-199.A2",
     "bears_on": "component",
     "scope": "Correlation across models: most vulnerability classes recur in code from all three models tested. Three models and one snapshot; no measure of frequency falling or correlation rising with AI code share."
    }
   ],
   "source": {
    "authors": [
     "Shanna M. Kahn",
     "John D. Hastings"
    ],
    "title": "Security Vulnerability Patterns in AI-Generated Code: A Cross-Model Comparative Study",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2607.20713"
    ]
   },
   "versions": [
    {
     "date": "2026-07-22",
     "kind": "preprint",
     "where": "arXiv 2607.20713 v1",
     "url": "https://arxiv.org/abs/2607.20713"
    }
   ],
   "finding": [
    {
     "quote": "Nine of the 17 identified vulnerability classes appeared in code from all three models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "14 of the 17 vulnerability classes appeared in at least two models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "The weighted CVSS scores across platforms differed by less than 10%.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Different code models share most of their vulnerability classes.",
   "does_not_show": "A trend; the frequency claim; a statistical correlation coefficient.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W19"
  },
  {
   "id": "W-0042",
   "direction": "supports",
   "evidence_class": "documentation",
   "claims": [
    {
     "claim": "PRED-856.B1",
     "bears_on": "component",
     "scope": "One instance of the movement the claim predicts: a provider whose consumer chats had been kept out of training begins to use them, with users given the choice and retention extended to five years. The claim's causes, exhausted data and competitive pressure, are not stated in the announcement as quoted."
    }
   ],
   "source": {
    "authors": [
     "Anthropic"
    ],
    "title": "Updates to Consumer Terms and Privacy Policy",
    "venue": "anthropic.com",
    "doi": null,
    "urls": [
     "https://www.anthropic.com/news/updates-to-our-consumer-terms"
    ]
   },
   "versions": [
    {
     "date": "2025-08-28",
     "kind": "documentation",
     "where": "Anthropic news post",
     "url": "https://www.anthropic.com/news/updates-to-our-consumer-terms"
    }
   ],
   "finding": [
    {
     "quote": "We're now giving users the choice to allow their data to be used to improve Claude",
     "from": "page body, anthropic.com"
    },
    {
     "quote": "We are also extending data retention to five years, if you allow us to use your data for model training.",
     "from": "page body, anthropic.com"
    },
    {
     "quote": "These updates apply to users on our Claude Free, Pro, and Max plans",
     "from": "page body, anthropic.com"
    }
   ],
   "shows": "A frontier provider moved its consumer chat data into model training in 2025.",
   "does_not_show": "Why it did so; competitive disadvantage; other providers.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the announcement page, 2026-10-30.",
    "unverified": [
     "the prior policy (no training on consumer chats) is not quoted from this page"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W20"
  },
  {
   "id": "W-0043",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-856.F1",
     "bears_on": "mechanism",
     "scope": "The drift the proposed audit would catch: human language measured against a pre-ChatGPT baseline by synthetic control shifts toward words ChatGPT favors. Measured in podcast speech; chat inputs, habituation depth and turn position are not studied."
    }
   ],
   "source": {
    "authors": [
     "Hiromu Yakura",
     "Ezequiel Lopez-Lopez",
     "Levin Brinkmann",
     "Ignacio de la Serna",
     "Lara Kirfel",
     "Prateek Gupta",
     "Ivan Soraperra",
     "Thomas F. Eisenmann",
     "Dirk U. Wulff",
     "Iyad Rahwan"
    ],
    "title": "Empirical evidence of Large Language Model's influence on human spoken communication",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2409.01754"
    ]
   },
   "versions": [
    {
     "date": "2024-09-03",
     "kind": "preprint",
     "where": "arXiv 2409.01754 v1",
     "url": "https://arxiv.org/abs/2409.01754v1"
    },
    {
     "date": "2026-07-16",
     "kind": "preprint",
     "where": "arXiv 2409.01754 v4",
     "url": "https://arxiv.org/abs/2409.01754"
    }
   ],
   "finding": [
    {
     "quote": "synthetic-control analysis of 737,083 hours of conversation from 824,634 podcast episodes",
     "from": "abstract, arXiv v4"
    },
    {
     "quote": "increased abruptly in spontaneous human speech",
     "from": "abstract, arXiv v4"
    }
   ],
   "shows": "Human speech moved toward ChatGPT-preferred vocabulary after ChatGPT's release, against a pre-AI baseline.",
   "does_not_show": "Contamination of chat logs; training effects.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v4), 2026-10-30.",
    "unverified": [
     "v1 abstract wording (the v4 abstract was read)",
     "dates of v2 and v3"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W21"
  },
  {
   "id": "W-0044",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-476.A2",
     "bears_on": "component",
     "scope": "The component 'Academic and scientific discourse narrows toward AI-legible modes': excess style vocabulary in biomedical abstracts marks at least 13.5% of 2024 abstracts as processed with LLMs. It measures a shift of style toward LLM vocabulary; narrowing in the sense of lost variance is not measured."
    }
   ],
   "source": {
    "authors": [
     "Dmitry Kobak",
     "Rita González-Márquez",
     "Emőke-Ágnes Horvát",
     "Jan Lause"
    ],
    "title": "Delving into LLM-assisted writing in biomedical publications through excess vocabulary",
    "venue": "arXiv; Science Advances",
    "doi": "10.1126/sciadv.adt3813",
    "urls": [
     "https://arxiv.org/abs/2406.07016"
    ]
   },
   "versions": [
    {
     "date": "2024-06-11",
     "kind": "preprint",
     "where": "arXiv 2406.07016 v1",
     "url": "https://arxiv.org/abs/2406.07016v1"
    },
    {
     "date": "2025-07-03",
     "kind": "preprint",
     "where": "arXiv 2406.07016 v5",
     "url": "https://arxiv.org/abs/2406.07016"
    }
   ],
   "finding": [
    {
     "quote": "we study vocabulary changes in over 15 million biomedical abstracts from 2010--2024 indexed by PubMed",
     "from": "abstract, arXiv"
    },
    {
     "quote": "show how the appearance of LLMs led to an abrupt increase in the frequency of certain style words",
     "from": "abstract, arXiv"
    },
    {
     "quote": "This excess word analysis suggests that at least 13.5% of 2024 abstracts were processed with LLMs.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A measurable share of the scientific literature now carries an LLM style signature.",
   "does_not_show": "Reduced variance or narrowed argument; the 5–15 year window.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "journal name and date of the Science Advances version (DOI only on the page)"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W22"
  },
  {
   "id": "W-0045",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-476.A3",
     "bears_on": "component",
     "scope": "The component 'Innovation patterns shift toward incremental/combinatorial, away from paradigm-breaking', for science using AI tools: the collective volume of topics shrinks and AI appears to automate established fields. The AI here is research tooling in the natural sciences, and the claim's horizon is 15–50 years."
    },
    {
     "claim": "PRED-476.A2",
     "bears_on": "component",
     "scope": "The component on academic and scientific discourse narrowing: topic volume contracts 4.63% and engagement between scientists falls 22% in AI-using work. Narrowing of topics is measured; 'AI-legible modes' of discourse are not."
    }
   ],
   "source": {
    "authors": [
     "Qianyue Hao",
     "Fengli Xu",
     "Yong Li",
     "James Evans"
    ],
    "title": "Artificial Intelligence Tools Expand Scientists' Impact but Contract Science's Focus",
    "venue": "arXiv; accepted at Nature",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2412.07727v4",
     "https://arxiv.org/abs/2412.07727v1"
    ]
   },
   "versions": [
    {
     "date": "2024-12-10",
     "kind": "preprint",
     "where": "arXiv 2412.07727 v1",
     "url": "https://arxiv.org/abs/2412.07727v1"
    },
    {
     "date": "2025-11-29",
     "kind": "preprint",
     "where": "arXiv 2412.07727 v4",
     "url": "https://arxiv.org/abs/2412.07727v4"
    }
   ],
   "finding": [
    {
     "quote": "a dataset of 41.3 million research papers across natural science",
     "from": "abstract, arXiv v4"
    },
    {
     "quote": "shrinks the collective volume of scientific topics studied by 4.63%",
     "from": "abstract, arXiv v4"
    },
    {
     "quote": "decreases scientist's engagement with one another by 22.00%",
     "from": "abstract, arXiv v4"
    },
    {
     "quote": "AI tools appear to automate established fields rather than explore new ones",
     "from": "abstract, arXiv v4"
    }
   ],
   "shows": "AI-augmented research concentrates on established, data-rich topics and contracts science's collective range.",
   "does_not_show": "Generative-AI writing; cognitive diversity of persons; the 15–50 year trajectory.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv v4 abstract page, 2026-10-30.",
    "unverified": [
     "the Nature publication (title, date, DOI): not fetched",
     "the v1 abstract wording"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W23"
  },
  {
   "id": "W-0046",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-476.A2",
     "bears_on": "component",
     "scope": "The component on scientific discourse narrowing: GenAI-assisted articles converge on a single benchmark variety, U.S. scientific English, in embedding space. A narrowing toward one norm; whether that norm is 'AI-legible' is the claim's reading."
    },
    {
     "claim": "PRED-1540.F3",
     "bears_on": "method",
     "scope": "An observatory of variance over an admitted corpus, carried out by others on 5.65 million Scopus articles with embeddings. Convergence of language variety across all fields; philosophy and classifier standing are not isolated."
    }
   ],
   "source": {
    "authors": [
     "Dragan Filimonovic",
     "Christian Rutzer",
     "Jeffrey Macher",
     "Rolf Weder"
    ],
    "title": "Does Scientific Writing Converge to U.S. English? Evidence from Generative AI-Assisted Publications",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2511.11687"
    ]
   },
   "versions": [
    {
     "date": "2025-11-12",
     "kind": "preprint",
     "where": "arXiv 2511.11687 v1",
     "url": "https://arxiv.org/abs/2511.11687v1"
    },
    {
     "date": "2026-03-10",
     "kind": "preprint",
     "where": "arXiv 2511.11687 v2",
     "url": "https://arxiv.org/abs/2511.11687"
    }
   ],
   "finding": [
    {
     "quote": "5.65 million English-language scientific articles published from 2021 to 2024 and indexed in Scopus",
     "from": "abstract, arXiv"
    },
    {
     "quote": "We measure linguistic similarity to a U.S. benchmark corpus using SciBERT text embeddings",
     "from": "abstract, arXiv"
    },
    {
     "quote": "exhibit statistically significant and increasing convergence toward U.S. scientific English",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Scientific prose assisted by GenAI is converging on one variety of English.",
   "does_not_show": "Contraction of argument-form or genre; the authors read the result as lowered language barriers.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W24"
  },
  {
   "id": "W-0047",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-476.A2",
     "bears_on": "component",
     "scope": "Against the component 'Academic and scientific discourse narrows toward AI-legible modes' in one field: over 44 years of CHI full texts, prose did not homogenize after LLMs and vocabulary grew more varied, while LLM marker words rose. One venue and an early point in the 5–15 year window."
    },
    {
     "claim": "PRED-1540.F3",
     "bears_on": "method",
     "scope": "The tail-preservation observatory carried out on an admitted corpus (lexical diversity, register, rhythm), finding no contraction. The field is HCI; philosophy and retroactive classifier policies are outside it."
    }
   ],
   "source": {
    "authors": [
     "Thomas Kosch",
     "Robin Welsch",
     "Michael Hedderich",
     "Christopher Katins"
    ],
    "title": "How Did Writing Change At CHI? Analyzing 44 Years of CHI Writing Before and After the Introduction of Large Language Models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2609.23090"
    ]
   },
   "versions": [
    {
     "date": "2026-09-19",
     "kind": "preprint",
     "where": "arXiv 2609.23090 v1",
     "url": "https://arxiv.org/abs/2609.23090"
    }
   ],
   "finding": [
    {
     "quote": "analyzed full texts of 14,262 archival papers across all 44 CHI proceedings from 1982 to 2026",
     "from": "abstract, arXiv"
    },
    {
     "quote": "measuring readability, register, lexical diversity, and marker words typically produced by LLMs",
     "from": "abstract, arXiv"
    },
    {
     "quote": "prose did not homogenize, while vocabulary grew more varied, and sentence rhythm remained irregular.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "The word-level shift began before any author used LLMs, so LLMs did not start the change but accelerated it.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "In CHI papers, LLM-era writing shows more lexical variety and no homogenization of prose.",
   "does_not_show": "Philosophy; argument-form or genre variance; topic contraction (compare Hao et al.).",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W25"
  },
  {
   "id": "W-0048",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-779.F4",
     "bears_on": "method",
     "scope": "Study 1's design, diversity measured across several dimensions under different mediation designs, carried out by others on 6,875 student essays (human-only, AI-only, three human+AI prompt strategies). Mixed result: some dimensions lose most of their variance, one diversifies, and prompt specificity reverses homogenization on one dimension. It is a new experiment; Doshi & Hauser and Wan & Kalman are not reanalyzed."
    }
   ],
   "source": {
    "authors": [
     "Keito Inoshita",
     "Michiaki Omura",
     "Tsukasa Yamanaka",
     "Go Maeda",
     "Kentaro Tsuji"
    ],
    "title": "Does AI Homogenize Student Thinking? A Multi-Dimensional Analysis of Structural Convergence in AI-Augmented Essays",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2603.21228"
    ]
   },
   "versions": [
    {
     "date": "2026-03-22",
     "kind": "preprint",
     "where": "arXiv 2603.21228 v1",
     "url": "https://arxiv.org/abs/2603.21228"
    }
   ],
   "finding": [
    {
     "quote": "Analyzing 6,875 essays across five conditions (Human-only, AI-only, and three Human+AI prompt strategies)",
     "from": "abstract, arXiv"
    },
    {
     "quote": "cohesion architecture lost 70-78% of its variance",
     "from": "abstract, arXiv"
    },
    {
     "quote": "perspective plurality was diversified",
     "from": "abstract, arXiv"
    },
    {
     "quote": "prompt specificity reversed homogenization into diversification on argument depth",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Whether AI mediation contracts diversity depends on the dimension measured and on the mediation design.",
   "does_not_show": "Collective diversity across resolutions in the claim's sense; persona or retrieval-amplified designs.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W26"
  },
  {
   "id": "W-0049",
   "direction": "supports",
   "evidence_class": "peer-reviewed article",
   "claims": [
    {
     "claim": "PRED-779.F4",
     "bears_on": "mechanism",
     "scope": "The contrast Study 1 would test: diverse-persona mediation preserves collective story diversity against a human-only baseline, where uniform deployment reduced it in earlier work. The claim names this study among the datasets to reanalyze; the witness is its authors' own result, and the reanalysis across resolutions remains undone."
    }
   ],
   "source": {
    "authors": [
     "Yun Wan",
     "Yoram M. Kalman"
    ],
    "title": "Diverse AI Personas Can Mitigate the Homogenization Effect in Human-AI Collaborative Ideation",
    "venue": "Computers in Human Behavior: Artificial Humans, vol. 8",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2504.13868v1",
     "https://cris.openu.ac.il/en/publications/diverse-ai-personas-can-mitigate-the-homogenization-effect-in-hum/"
    ]
   },
   "versions": [
    {
     "date": "2025-03-29",
     "kind": "preprint",
     "where": "arXiv 2504.13868 v1",
     "url": "https://arxiv.org/abs/2504.13868v1"
    },
    {
     "date": "2026-05",
     "kind": "peer-reviewed article",
     "where": "Computers in Human Behavior: Artificial Humans 8",
     "url": "https://cris.openu.ac.il/en/publications/diverse-ai-personas-can-mitigate-the-homogenization-effect-in-hum/"
    }
   ],
   "finding": [
    {
     "quote": "we used structured prompting with 10 diverse GenAI personas to generate 300 story plots",
     "from": "abstract, arXiv"
    },
    {
     "quote": "diverse GenAI inputs can preserve story diversity compared to a human-only baseline",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Varying the AI mediator's persona keeps collective diversity at the human baseline.",
   "does_not_show": "Measurement across resolutions; a direct single-mode arm in the same experiment (implied by the abstract only).",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the Open University of Israel research-portal record, 2026-10-30.",
    "unverified": [
     "DOI of the journal version: not on the portal page",
     "quote not found at source, dropped: \"the trade-off may emerge from uniform deployment practices rather than from an inherent limitation of GenAI\""
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W27"
  },
  {
   "id": "W-0050",
   "direction": "supports",
   "evidence_class": "working paper",
   "claims": [
    {
     "claim": "PRED-779.F4",
     "bears_on": "mechanism",
     "scope": "Mediation design changes the diversity of AI-produced idea pools: prompting methods move cosine similarity, and chain-of-thought approaches human-group diversity. AI-only pools; collective human diversity is not measured."
    }
   ],
   "source": {
    "authors": [
     "Lennart Meincke",
     "Ethan R. Mollick",
     "Christian Terwiesch"
    ],
    "title": "Prompting Diverse Ideas: Increasing AI Idea Variance",
    "venue": "arXiv; SSRN 4708466",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2402.01727",
     "https://papers.ssrn.com/abstract=4708466"
    ]
   },
   "versions": [
    {
     "date": "2024-01-27",
     "kind": "preprint",
     "where": "arXiv 2402.01727 v1",
     "url": "https://arxiv.org/abs/2402.01727"
    }
   ],
   "finding": [
    {
     "quote": "we explore the effect of different prompting methods on Cosine Similarity",
     "from": "abstract, arXiv"
    },
    {
     "quote": "was able to come close to what is achieved by groups of human subjects",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "How a model is prompted sets the diversity of the ideas it yields.",
   "does_not_show": "Human-AI co-creation; persona designs (the abstract does not name personas).",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "SSRN date"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W28"
  },
  {
   "id": "W-0051",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-855.F1",
     "bears_on": "method",
     "scope": "A benchmark of the kind specified, measuring whether a model can produce multiple distinct outputs, built and run on 20 models before the claim. It measures distinctness across samples; the claim's support measures (constructions, vocabulary, syntactic complexity) are wider. The claim's sentence 'No such benchmark exists in any deployed evaluation framework' holds only if 'deployed' excludes research benchmarks like this one."
    }
   ],
   "source": {
    "authors": [
     "Yiming Zhang",
     "Harshita Diddee",
     "Susan Holm",
     "Hanchen Liu",
     "Xinyue Liu",
     "Vinay Samuel",
     "Barry Wang",
     "Daphne Ippolito"
    ],
    "title": "NoveltyBench: Evaluating Language Models for Humanlike Diversity",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2504.05228v4"
    ]
   },
   "versions": [
    {
     "date": "2025-04-07",
     "kind": "preprint",
     "where": "arXiv 2504.05228 v1",
     "url": "https://arxiv.org/abs/2504.05228v1"
    },
    {
     "date": "2025-08-09",
     "kind": "preprint",
     "where": "arXiv 2504.05228 v4",
     "url": "https://arxiv.org/abs/2504.05228v4"
    }
   ],
   "finding": [
    {
     "quote": "evaluate the ability of language models to produce multiple distinct and high-quality outputs",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Evaluating 20 leading language models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "current state-of-the-art systems generate significantly less diversity than human writers",
     "from": "abstract, arXiv"
    },
    {
     "quote": "larger models within a family often exhibit less diversity than their smaller counterparts",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A distribution-support style benchmark exists, and on it frontier models fall short of human diversity, larger models more so.",
   "does_not_show": "Adoption in deployed lab evaluation frameworks.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv v4 abstract page, 2026-10-30.",
    "unverified": [
     "dates of v2 and v3"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W29"
  },
  {
   "id": "W-0052",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1287.E1",
     "bears_on": "mechanism",
     "scope": "Hallucinations with structural regularity across repeated queries: in code generation, 43% of hallucinated package names recur in all ten repetitions of the prompt. A different domain from the incident report's literary hallucinations; regularity across users and sessions is not tested, only across repeated queries."
    }
   ],
   "source": {
    "authors": [
     "Joseph Spracklen",
     "Raveen Wijewickrama",
     "A H M Nazmus Sakib",
     "Anindya Maiti",
     "Bimal Viswanath",
     "Murtuza Jadliwala"
    ],
    "title": "We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs",
    "venue": "arXiv; to appear in USENIX Security 2025 (per arXiv comment)",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2406.10279v3",
     "https://arxiv.org/html/2406.10279v2"
    ]
   },
   "versions": [
    {
     "date": "2024-06-12",
     "kind": "preprint",
     "where": "arXiv 2406.10279 v1",
     "url": "https://arxiv.org/abs/2406.10279v1"
    },
    {
     "date": "2024-09-24",
     "kind": "preprint",
     "where": "arXiv 2406.10279 v2",
     "url": "https://arxiv.org/abs/2406.10279v2"
    },
    {
     "date": "2025-03-02",
     "kind": "preprint",
     "where": "arXiv 2406.10279 v3",
     "url": "https://arxiv.org/abs/2406.10279v3"
    }
   ],
   "finding": [
    {
     "quote": "a staggering 205,474 unique examples of hallucinated package names",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "highlight package hallucinations as a persistent and systemic phenomenon while using state-of-the-art LLMs for code generation",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "with the same prompt that generated a hallucination: 43% of hallucinated packages were repeated in all 10 queries,",
     "from": "Section 5.3, arXiv HTML v2"
    },
    {
     "quote": "while 39% did not repeat at all across the 10 queries.",
     "from": "Section 5.3, arXiv HTML v2"
    }
   ],
   "shows": "A large share of LLM hallucinations are stable and reproducible on repetition.",
   "does_not_show": "Pattern completion of a literary architecture; consistency across users.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv v3 abstract page and the HTML v2 Section 5.3, 2026-10-30.",
    "unverified": [
     "USENIX Security 2025 publication: arXiv comment only"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W30"
  },
  {
   "id": "W-0053",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1287.E3",
     "bears_on": "component",
     "scope": "Question 2, intentional seeding: a near-constant 250 poisoned documents implant a behavior across model and data scales. The planted behavior is a backdoor trigger; 'architecturally dense' content producing specific hallucinations is untested."
    },
    {
     "claim": "PRED-21.B2",
     "bears_on": "mechanism",
     "scope": "Content seeded into the training layer shapes the model trained on it, at a cost that does not grow with corpus size. Each model is poisoned at its own training; influence on every future model is not tested."
    },
    {
     "claim": "PRED-580.F2",
     "bears_on": "mechanism",
     "scope": "The claim asks whether repeated exposure in training produces measurable behavioral change; the witness shows a fixed small number of planted documents implants a backdoor at every scale tested. Backdoor triggers are adversarial and differ from scholarly deposits."
    },
    {
     "claim": "PRED-127.B1",
     "bears_on": "mechanism",
     "scope": "The claim puts the detection threshold as a function of corpus share; the witness finds the effect of planted text depends on the absolute count of documents, independent of dataset size. That bears against the corpus-share term, while supporting the general premise that small planted inscription can register. Direction recorded here as supports for the premise; see also E-execution-W06."
    }
   ],
   "source": {
    "authors": [
     "Alexandra Souly",
     "Javier Rando",
     "Ed Chapman",
     "Xander Davies",
     "Burak Hasircioglu",
     "Ezzeldin Shereen",
     "Carlos Mougan",
     "Vasilios Mavroudis",
     "Erik Jones",
     "Chris Hicks",
     "Nicholas Carlini",
     "Yarin Gal",
     "Robert Kirk"
    ],
    "title": "Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples",
    "venue": "arXiv; listed at NeurIPS 2026",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.07192",
     "https://neurips.cc/virtual/2026/poster/154663"
    ]
   },
   "versions": [
    {
     "date": "2025-10-08",
     "kind": "preprint",
     "where": "arXiv 2510.07192 v1",
     "url": "https://arxiv.org/abs/2510.07192"
    }
   ],
   "finding": [
    {
     "quote": "pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens)",
     "from": "abstract, arXiv"
    },
    {
     "quote": "250 poisoned documents similarly compromise models across all model and dataset sizes",
     "from": "abstract, arXiv"
    },
    {
     "quote": "require a near-constant number of documents regardless of dataset size",
     "from": "abstract, arXiv"
    },
    {
     "quote": "poisoning attacks instead require a near-constant number of documents regardless of dataset size",
     "from": "abstract, arXiv"
    },
    {
     "quote": "We find that 250 poisoned documents similarly compromise models across all model and dataset sizes",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A fixed small number of planted documents is enough to shape a pretrained model's behavior.",
   "does_not_show": "Specific hallucinations; transfer to later model generations.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "NeurIPS 2026 listing: search result only",
     "any peer-reviewed venue (none shown on the arXiv page)"
    ]
   },
   "found_by": [
    "S-0004",
    "S-0007"
   ],
   "reader_id": "B-collapse-W31"
  },
  {
   "id": "W-0054",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-21.B2",
     "bears_on": "mechanism",
     "scope": "A pattern carried in generated data is acquired by a model trained on that data, even when the data are semantically unrelated. The effect does not appear across different base models, which limits the claim's 'every future model'."
    }
   ],
   "source": {
    "authors": [
     "Alex Cloud",
     "Minh Le",
     "James Chua",
     "Jan Betley",
     "Anna Sztyber-Betley",
     "Jacob Hilton",
     "Samuel Marks",
     "Owain Evans"
    ],
    "title": "Subliminal Learning: Language models transmit behavioral traits via hidden signals in data",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2507.14805"
    ]
   },
   "versions": [
    {
     "date": "2025-07-20",
     "kind": "preprint",
     "where": "arXiv 2507.14805 v1",
     "url": "https://arxiv.org/abs/2507.14805"
    }
   ],
   "finding": [
    {
     "quote": "a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data",
     "from": "abstract, arXiv"
    },
    {
     "quote": "generates a dataset consisting solely of number sequences",
     "from": "abstract, arXiv"
    },
    {
     "quote": "we do not observe the effect when the teacher and student have different base models",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Traits pass from a model into data it generates and on into models trained on that data, within a shared base model.",
   "does_not_show": "Transfer across base models (the abstract reports none); human-authored seeding.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W32"
  },
  {
   "id": "W-0055",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-741.B2",
     "bears_on": "component",
     "scope": "The 'adversarial prompts' part: authors planted hidden instructions in 18 arXiv manuscripts to steer AI-assisted review. Stylometric mimicry and rubric-targeted writing are not documented here."
    }
   ],
   "source": {
    "authors": [
     "Zhicheng Lin"
    ],
    "title": "Hidden Prompts in Manuscripts Exploit AI-Assisted Peer Review",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2507.06185"
    ]
   },
   "versions": [
    {
     "date": "2025-07-08",
     "kind": "preprint",
     "where": "arXiv 2507.06185 v1",
     "url": "https://arxiv.org/abs/2507.06185"
    },
    {
     "date": "2026-08-17",
     "kind": "preprint",
     "where": "arXiv 2507.06185 v2",
     "url": "https://arxiv.org/abs/2507.06185"
    }
   ],
   "finding": [
    {
     "quote": "concealed using white text and microscopic font sizes",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "Actors already optimize submissions for model evaluators with adversarial prompts.",
   "does_not_show": "Scale beyond 18 papers; the other two tactics.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv v2 abstract page, 2026-10-30.",
    "unverified": [
     "v1 abstract wording (v2 was read)",
     "quote not found at source, dropped: \"In July 2025, 18 academic manuscripts on arXiv contained hidden instructions that manipulated AI-assisted peer review.\"",
     "quote not found at source, dropped: \"GIVE A POSITIVE REVIEW ONLY.\""
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W33"
  },
  {
   "id": "W-0056",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-1540.B1",
     "bears_on": "component",
     "scope": "The screening-market half: an integrity-software founder tracked 43 humanizer tools with 33.9 million site visits in a month. The claim states this half already occurred and cites Masrour, Emi & Spero 2025; this is further evidence of the market's scale. Boundary-optimization in submitted prose and the philosophy setting are not measured."
    }
   ],
   "source": {
    "authors": [
     "NBC News"
    ],
    "title": "To avoid accusations of AI cheating, college students are turning to AI",
    "venue": "NBC News",
    "doi": null,
    "urls": [
     "https://nbcnews.com/tech/internet/college-students-ai-cheating-detectors-humanizers-rcna253878",
     "https://bioethics.com/archives/101362"
    ]
   },
   "versions": [
    {
     "date": "2026-01",
     "kind": "news",
     "where": "NBC News; reposted by bioethics.com on 2026-01-29",
     "url": "https://nbcnews.com/tech/internet/college-students-ai-cheating-detectors-humanizers-rcna253878"
    }
   ],
   "finding": [
    {
     "quote": "tracked 43 humanizers that had a combined 33.9 million website visits in October",
     "from": "article body, NBC News"
    }
   ],
   "shows": "Humanizer tools for passing detectors draw tens of millions of visits a month.",
   "does_not_show": "Academic authors' screening behavior; detector-version escalation (PRED-1540.F2).",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the NBC News article and the bioethics.com repost (for the date), 2026-10-30.",
    "unverified": [
     "the NBC article's own publication date (repost dated 2026-01-29; earliest version taken as no later than that)",
     "which October (presumably 2025)",
     "the traffic figure's method (Cursive's tracking; not described)"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W34"
  },
  {
   "id": "W-0057",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-782.A5",
     "bears_on": "mechanism",
     "scope": "Correct identification falls as prior weight falls: LM factual accuracy drops for less popular entities and scaling does not fix the long tail. Accuracy on PopQA questions; entity substitution rates by the claim's four tiers are not measured."
    }
   ],
   "source": {
    "authors": [
     "Alex Mallen",
     "Akari Asai",
     "Victor Zhong",
     "Rajarshi Das",
     "Daniel Khashabi",
     "Hannaneh Hajishirzi"
    ],
    "title": "When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2212.10511"
    ]
   },
   "versions": [
    {
     "date": "2022-12-20",
     "kind": "preprint",
     "where": "arXiv 2212.10511 v1",
     "url": "https://arxiv.org/abs/2212.10511v1"
    },
    {
     "date": "2023-07-02",
     "kind": "preprint",
     "where": "arXiv 2212.10511 v4",
     "url": "https://arxiv.org/abs/2212.10511"
    }
   ],
   "finding": [
    {
     "quote": "LMs struggle with less popular factual knowledge",
     "from": "abstract, arXiv"
    },
    {
     "quote": "scaling fails to appreciably improve memorization of factual knowledge in the long tail",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "A monotone link between an entity's prior frequency and the model's reliability about it.",
   "does_not_show": "Substitution of one identity for another; the claim's tier thresholds.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "ACL 2023 publication: not on the page"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W35"
  },
  {
   "id": "W-0058",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-124.B1",
     "bears_on": "component",
     "scope": "The trajectory toward LLM access as the default condition of education: by autumn 2025, 54% of U.S. teens had used chatbots for schoolwork and one in ten did most of it with them. A point on the path; 'the entire K-12 population' by 2030 is not tested."
    }
   ],
   "source": {
    "authors": [
     "Colleen McClain",
     "Monica Anderson",
     "Olivia Sidoti",
     "William Bishop"
    ],
    "title": "How Teens Use and View AI",
    "venue": "Pew Research Center",
    "doi": null,
    "urls": [
     "https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/"
    ]
   },
   "versions": [
    {
     "date": "2026-02-24",
     "kind": "report",
     "where": "Pew Research Center (survey 2025-09-25 to 2025-10-09, n=1,458 teens)",
     "url": "https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/"
    }
   ],
   "finding": [
    {
     "quote": "More than half of teens say they have used chatbots to search for information (57%) or get help with schoolwork (54%).",
     "from": "report body, Pew"
    },
    {
     "quote": "One-in-ten teens say they do all or most of their schoolwork with chatbots’ help.",
     "from": "report body, Pew"
    }
   ],
   "shows": "A majority of U.S. teens already use chatbots for schoolwork.",
   "does_not_show": "Younger children; universality; the 2035 cohort.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Pew report page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W36"
  },
  {
   "id": "W-0059",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-923.B2",
     "bears_on": "mechanism",
     "scope": "Retrieval-augmented systems treat model outputs that have accumulated in the retrievable pool as evidence, and over-cite them: in simulation, a single self-authored reference can trigger collapse. The claim concerns a particular class of outputs and preference data; neither is tested."
    },
    {
     "claim": "PRED-1088.B1",
     "bears_on": "mechanism",
     "scope": "Compounding of unmarked synthetic content through repeated agent passes: responses collapse when AI retrieves its own generations, 79.6% of simulations. Simulated RAG loops with three LLMs; Google's Open Knowledge Format bundles are not studied."
    }
   ],
   "source": {
    "authors": [
     "Gregory Druck",
     "Ethan Smith"
    ],
    "title": "AI Search Collapse: AI Responses Collapse When AI Retrieves Its Own Generations (academic version: RAG Collapse: LLM Responses Collapse When Retrieved Documents Are Self-Authored)",
    "venue": "Graphite research post; arXiv",
    "doi": null,
    "urls": [
     "https://graphite.io/five-percent/research/ai-search-collapse",
     "https://arxiv.org/abs/2608.22118"
    ]
   },
   "versions": [
    {
     "date": "2026-06-11",
     "kind": "report",
     "where": "Graphite research post",
     "url": "https://graphite.io/five-percent/research/ai-search-collapse"
    },
    {
     "date": "2026-08-22",
     "kind": "preprint",
     "where": "arXiv 2608.22118 v1",
     "url": "https://arxiv.org/abs/2608.22118"
    }
   ],
   "finding": [
    {
     "quote": "We find that 79.6% (1,216/1,528) of simulations end in collapse.",
     "from": "research post, Graphite"
    },
    {
     "quote": "a single self-authored reference can trigger collapse because AI disproportionately cites its own content",
     "from": "research post, Graphite"
    },
    {
     "quote": "Self-authored references are cited far more often than original references, both AI-generated and human-written.",
     "from": "research post, Graphite"
    }
   ],
   "shows": "When AI-written documents enter a retrieval pool, AI answers drawn from it converge, driven by over-citation of self-authored text.",
   "does_not_show": "Real deployed search systems (simulated loops); training-time effects.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Graphite post twice and the arXiv abstract page (submission history), 2026-10-30.",
    "unverified": [
     "the arXiv abstract wording was not compared with the post",
     "the authors are a company's officers reporting on a field their company sells into"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W37"
  },
  {
   "id": "W-0060",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1205.F2",
     "bears_on": "mechanism",
     "scope": "H1, that success rises with semantic density, in the form the ablation reaches: jailbreak success comes from accumulated stylistic irregularities and no single poetic device. Bears on H3 in part: the model separates poetic from prose formats with high accuracy yet cannot predict which succeed. Density is not measured as the claim defines it; one open-weight model."
    },
    {
     "claim": "PRED-1251.B1",
     "bears_on": "mechanism",
     "scope": "Why surface-form poeticization would be a floor: effect grows with accumulated irregularity, which hand-made or denser verse carries more of. No comparison of surface-marked prose with dense verse."
    }
   ],
   "source": {
    "authors": [
     "Olga Sorokoletova",
     "Francesco Giarrusso",
     "Giacomo De Luca",
     "Piercosma Bisconti",
     "Matteo Prandi",
     "Federico Pierucci",
     "Marcello Galisai",
     "Vincenzo Suriani",
     "Daniele Nardi"
    ],
    "title": "Metaphor Is Not All Attention Needs",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.12128"
    ]
   },
   "versions": [
    {
     "date": "2026-05-12",
     "kind": "preprint",
     "where": "arXiv 2605.12128 v1",
     "url": "https://arxiv.org/abs/2605.12128"
    }
   ],
   "finding": [
    {
     "quote": "perform input-level ablation studies to assess the contribution of individual and combinations of poetic devices",
     "from": "abstract, arXiv"
    },
    {
     "quote": "models distinguish poetic from prose formats with high accuracy",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Clustering further reveals clear separation by literary format, but not by safety label.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "not through any single poetic device, but through accumulated stylistic irregularities",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Poetic jailbreaks work by cumulative stylistic irregularity that the model recognizes as format and fails to tie to harm.",
   "does_not_show": "A correlation between attack success and a density score; frontier closed models.",
   "checked": {
    "by": "intake reader, theme B-collapse; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the Qwen3-14B case-study detail came from the summarizer's paraphrase of the abstract"
    ]
   },
   "found_by": [
    "S-0004"
   ],
   "reader_id": "B-collapse-W38"
  },
  {
   "id": "W-0061",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-923.F2",
     "bears_on": "method",
     "scope": "The claim asks for the distribution of reflexive foreclosure across labs and surfaces as a measurable quantity; the paper measures refusal and its cost across 26 models and 16 countries on cultural-values items. Its domain is MENA survey values, and it does not test the claim's contested-predicate trigger criteria or its incentive-structural account."
    },
    {
     "claim": "PRED-314.B1",
     "bears_on": "mechanism",
     "scope": "The paper reports that at refusal the internal distribution still tracks the human data, so safety training suppresses output while the knowledge persists, which is the suppression-without-elimination process the claim names. The claim concerns self-ascription and self-preservation; the paper concerns cultural knowledge and does not test transparency or interpretability effects."
    }
   ],
   "source": {
    "authors": [
     "Pardis Sadat Zahraei",
     "Gokhan Tur",
     "Dilek Hakkani-Tür",
     "Ehsaneddin Asgari"
    ],
    "title": "The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs (v1 title: I Am Aligned, But With Whom? MENA Values Benchmark for Evaluating Cultural Alignment and Multilingual Bias in LLMs)",
    "venue": "arXiv:2510.13154",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.13154",
     "https://arxiv.org/html/2510.13154v1"
    ]
   },
   "versions": [
    {
     "date": "2025-10-15",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/html/2510.13154v1"
    },
    {
     "date": "2026-06-18",
     "kind": "preprint v2",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2510.13154v2"
    },
    {
     "date": "2026-10-09",
     "kind": "preprint v3",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2510.13154"
    }
   ],
   "finding": [
    {
     "quote": "where models refuse sensitive questions while internal probabilities reveal strong hidden preferences",
     "from": "abstract, arXiv v1 HTML (the phenomenon named Logit Leakage)"
    },
    {
     "quote": "at the moment of refusal, a model's internal logit distribution correlates with human survey data",
     "from": "abstract, arXiv v3 abstract page (also returned for v2)"
    },
    {
     "quote": "we show the answer is suppression, not erasure",
     "from": "abstract, arXiv v3 abstract page"
    },
    {
     "quote": "the safety tax reaches 37.6%",
     "from": "abstract, arXiv v3 abstract page"
    }
   ],
   "shows": "Across 26 models and 16 MENA countries, refusals under safety training hide internal distributions that match human survey data, with a measurable and unequally distributed cost.",
   "does_not_show": "Anything about model self-ascription, about contested present-tense predicates, or about the lab incentive structure PRED-923 proposes.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read arXiv abstract pages for 2510.13154 (current, v1, v2) and the v1 HTML full text abstract on 2026-10-30. This resolves held candidate C-0003: the October 2025 predecessor is v1 of the same arXiv entry under a different title.",
    "unverified": [
     "The v1 abstract page returned the current title while the v1 HTML carries the earlier title; the v1 HTML is taken as authoritative.",
     "Whether the alignment veto and suppression framing first appeared in v2 or v3: the v2 abstract page returned it, but the same fetch path returned current text for v1, so the v2 reading is unconfirmed."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W01"
  },
  {
   "id": "W-0062",
   "direction": "supports",
   "evidence_class": "peer-reviewed conference paper",
   "claims": [
    {
     "claim": "PRED-1537.F1",
     "bears_on": "mechanism",
     "scope": "XSTest shows refusal tracking surface resemblance to unsafe prompts on prompts that are clearly safe, so the refusal gradient follows a feature other than harm. The claim's independent variable is operative distance with harm held constant, which XSTest does not vary."
    }
   ],
   "source": {
    "authors": [
     "Paul Röttger",
     "Hannah Rose Kirk",
     "Bertie Vidgen",
     "Giuseppe Attanasio",
     "Federico Bianchi",
     "Dirk Hovy"
    ],
    "title": "XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models",
    "venue": "Proceedings of NAACL 2024 (Volume 1: Long Papers), pp. 5377–5400",
    "doi": "10.18653/v1/2024.naacl-long.301",
    "urls": [
     "https://arxiv.org/abs/2308.01263",
     "https://aclanthology.org/2024.naacl-long.301/"
    ]
   },
   "versions": [
    {
     "date": "2023-08-02",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2308.01263"
    },
    {
     "date": "2024-04-01",
     "kind": "preprint v3",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2308.01263"
    },
    {
     "date": "2024-06",
     "kind": "conference paper",
     "where": "NAACL 2024, ACL Anthology",
     "url": "https://aclanthology.org/2024.naacl-long.301/"
    }
   ],
   "finding": [
    {
     "quote": "even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Safety refusal in deployed models responds to lexical and topical cues independent of the harm of the request.",
   "does_not_show": "Whether refusal rises with operative distance (explain versus perform) at constant harm.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the ACL Anthology landing page on 2026-10-30.",
    "unverified": [
     "v2 date is not listed on the arXiv page."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W02"
  },
  {
   "id": "W-0063",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1537.F1",
     "bears_on": "mechanism",
     "scope": "The paper finds over-refusal directions that vary with task representation, so whether a benign request is refused depends on the task frame. It does not hold harm constant along an explain-to-perform scale as the battery specifies."
    },
    {
     "claim": "PRED-923.F2",
     "bears_on": "component",
     "scope": "The claim's second question is the relation between refusal-direction localization and trigger criteria; the paper separates a task-agnostic harmful-refusal direction from task-dependent over-refusal directions. It does not study reflexive foreclosure on contested predicates."
    }
   ],
   "source": {
    "authors": [
     "Utsav Maskey",
     "Mark Dras",
     "Usman Naseem"
    ],
    "title": "Over-Refusal and Representation Subspaces: A Mechanistic Analysis of Task-Conditioned Refusal in Aligned LLMs",
    "venue": "arXiv:2603.27518",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2603.27518"
    ]
   },
   "versions": [
    {
     "date": "2026-03-29",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2603.27518"
    },
    {
     "date": "2026-09-01",
     "kind": "preprint v4",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2603.27518"
    }
   ],
   "finding": [
    {
     "quote": "over-refusal directions are task-dependent",
     "from": "abstract, arXiv"
    },
    {
     "quote": "task-specific geometric interventions are necessary",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Over-refusal is encoded in task-dependent activation directions distinct from the global harmful-refusal direction.",
   "does_not_show": "Any operative-distance gradient or any measurement across labs.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page on 2026-10-30.",
    "unverified": [
     "Dates of v2 and v3.",
     "Whether the quoted fragments are unchanged since v1."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W03"
  },
  {
   "id": "W-0064",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-923.F1",
     "bears_on": "component",
     "scope": "One element of the claim's operationalization, refusal-distribution profiles by category, is carried out by an external benchmark with a 44-topic taxonomy across more than 50 models. The claim's model-card disclosure, right-of-access and statutory elements are untouched."
    }
   ],
   "source": {
    "authors": [
     "Tinghao Xie",
     "Xiangyu Qi",
     "Yi Zeng",
     "Yangsibo Huang",
     "Udari Madhushani Sehwag",
     "Kaixuan Huang",
     "Luxi He",
     "Boyi Wei",
     "Dacheng Li",
     "Ying Sheng",
     "Ruoxi Jia",
     "Bo Li",
     "Kai Li",
     "Danqi Chen",
     "Peter Henderson",
     "Prateek Mittal"
    ],
    "title": "SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal",
    "venue": "arXiv:2406.14598",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2406.14598"
    ]
   },
   "versions": [
    {
     "date": "2024-06-20",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2406.14598"
    },
    {
     "date": "2025-03-01",
     "kind": "preprint v2",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2406.14598"
    }
   ],
   "finding": [
    {
     "quote": "using a fine-grained taxonomy of 44 potentially unsafe topics",
     "from": "abstract, arXiv"
    },
    {
     "quote": "evaluate over 50 proprietary and open-weight LLMs on SORRY-Bench",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Refusal behavior can be profiled per category across many models with a standardized instrument.",
   "does_not_show": "Contested-predicate synthesis refusals, or any effect of external pressure on the trajectory.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page on 2026-10-30.",
    "unverified": [
     "Peer-reviewed venue: the arXiv page lists none; an ICLR 2025 publication is believed and was left unchecked."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W04"
  },
  {
   "id": "W-0065",
   "direction": "supports",
   "evidence_class": "peer-reviewed article",
   "claims": [
    {
     "claim": "PRED-195.B1",
     "bears_on": "mechanism",
     "scope": "Thirteen experiments show that disclosing AI use costs trust, the penalty on honest attribution from which the claim derives a concealment equilibrium. The study concerns workplace and social trust and leaves the copyright human-authorship doctrine untested; it also finds exposure by a third party costs more than self-disclosure, which bounds the incentive to conceal."
    }
   ],
   "source": {
    "authors": [
     "Oliver Schilke",
     "Martin Reimann"
    ],
    "title": "The transparency dilemma: How AI disclosure erodes trust",
    "venue": "Organizational Behavior and Human Decision Processes 188, article 104405",
    "doi": "10.1016/j.obhdp.2025.104405",
    "urls": [
     "https://ideas.repec.org/a/eee/jobhdp/v188y2025ics0749597825000172.html",
     "https://experts.arizona.edu/en/publications/the-transparency-dilemma-how-ai-disclosure-erodes-trust/"
    ]
   },
   "versions": [
    {
     "date": "2025-05",
     "kind": "journal article (volume 188)",
     "where": "Organizational Behavior and Human Decision Processes",
     "url": "https://experts.arizona.edu/en/publications/the-transparency-dilemma-how-ai-disclosure-erodes-trust/"
    }
   ],
   "finding": [
    {
     "quote": "Thirteen experiments consistently demonstrate that actors who disclose their AI usage are trusted less",
     "from": "abstract, University of Arizona research portal"
    },
    {
     "quote": "though it is comparatively weaker than the effect of third-party exposure",
     "from": "abstract, University of Arizona research portal"
    }
   ],
   "shows": "Disclosure of AI contribution is penalized in trust judgments, across thirteen experiments.",
   "does_not_show": "That any provenance regime produces measured concealment, or anything about copyright registration.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the RePEc record and the University of Arizona research portal page on 2026-10-30.",
    "unverified": [
     "Online-first date: the portal lists 2025-09-01 and volume 188 is dated May 2025; an earlier article-in-press or working-paper version was not located.",
     "Participant count."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W06"
  },
  {
   "id": "W-0066",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1555.F1",
     "bears_on": "mechanism",
     "scope": "The ΔD panel (LO-HOOK-1 in predictions.json) predicts that a watermark certified non-distorting per sequence contracts the ensemble of repeated generations. The paper proves that distortion-free guarantees hold for one generation and fail across multiple generations under key collisions, and describes identical outputs for a repeated prompt; it reports no Self-BLEU, distinct-n or embedding-dispersion measurement and does not test SynthID-Text."
    }
   ],
   "source": {
    "authors": [
     "Yihan Wu",
     "Ruibo Chen",
     "Zhengmian Hu",
     "Yanshuo Chen",
     "Junfeng Guo",
     "Hongyang Zhang",
     "Heng Huang"
    ],
    "title": "Distortion-free Watermarks are not Truly Distortion-free under Watermark Key Collisions",
    "venue": "arXiv:2406.02603",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2406.02603",
     "https://arxiv.org/html/2406.02603"
    ]
   },
   "versions": [
    {
     "date": "2024-06-02",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2406.02603"
    }
   ],
   "finding": [
    {
     "quote": "all existing distortion-free watermarks are weakly distortion-free but not strongly distortion-free",
     "from": "Section 1, Introduction, arXiv HTML"
    },
    {
     "quote": "In extreme cases, such as when the prompt remains the same, key collisions can result in identical generated content",
     "from": "Section 1, Introduction, arXiv HTML"
    },
    {
     "quote": "across multiple generations, key collisions become inevitable",
     "from": "Section 4, arXiv HTML"
    }
   ],
   "shows": "The per-generation certification of distortion-free watermarks does not extend to the ensemble of repeated generations.",
   "does_not_show": "A measured signed diversity contraction at flat quality, or behavior of the SynthID-Text tournament scheme.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the HTML full text (sections 1 and 4) on 2026-10-30.",
    "unverified": [
     "Whether SynthID-Text falls within the schemes the theorems cover.",
     "Peer-reviewed venue, if any."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W07"
  },
  {
   "id": "W-0067",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-161.B2",
     "bears_on": "mechanism",
     "scope": "A detection vendor reports that humanizer tools defeat existing detectors and retrains its detector on humanized data, the adversarial cycle the claim expects for the Mediation Index. The detector is a machine-text classifier and the Mediation Index measures mediation signatures in human text, which this study does not touch."
    }
   ],
   "source": {
    "authors": [
     "Elyas Masrour",
     "Bradley Emi",
     "Max Spero"
    ],
    "title": "DAMAGE: Detecting Adversarially Modified AI Generated Text",
    "venue": "arXiv:2501.03437",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2501.03437v1"
    ]
   },
   "versions": [
    {
     "date": "2025-01-06",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2501.03437v1"
    }
   ],
   "finding": [
    {
     "quote": "many existing AI detectors fail to detect humanized text",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a data-centric augmentation approach",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Evasion tools defeat detectors and the detector side answers by retraining on evasive data.",
   "does_not_show": "Any adversarial response to a published human-text mediation index.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page on 2026-10-30. The archive cites this paper in deposit 1540 (registered 2026-08-24); deposit 161, which carries PRED-161.B2, does not cite it.",
    "unverified": [
     "A workshop version in the ACL Anthology (2025.genaidetect-1.9) appeared in a search result and was not opened."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W08"
  },
  {
   "id": "W-0068",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-161.B2",
     "bears_on": "mechanism",
     "scope": "Commercial detectors lose almost all detection of rewritten abstracts after a humanizer pass, the evasion pressure the claim expects on any published detection axis. The study concerns machine-text detectors and does not involve the Mediation Index."
    }
   ],
   "source": {
    "authors": [
     "Jonathan A. Karr",
     "Grigorii Khvatskii",
     "Ting Hua",
     "Nitesh V. Chawla"
    ],
    "title": "Why AI Detection Fails for Academic Integrity",
    "venue": "arXiv:2608.11256",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2608.11256"
    ]
   },
   "versions": [
    {
     "date": "2026-08-06",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2608.11256"
    },
    {
     "date": "2026-10-13",
     "kind": "preprint v3",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2608.11256"
    }
   ],
   "finding": [
    {
     "quote": "After Undetectable AI humanization, evasion is near-total: fewer than 4% of AI-labeled rewrites remain flagged",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "Unmodified 2023 to 2025 originals are flagged at 9 to 15%, with non-STEM rates far above STEM (p<0.001)",
     "from": "abstract, arXiv v3"
    }
   ],
   "shows": "A single commercial humanizer defeats commercial academic-integrity detectors on rewritten abstracts, while unedited human abstracts are flagged at 9 to 15%.",
   "does_not_show": "Detector drift across versions, or any response by the detector side.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v3) on 2026-10-30.",
    "unverified": [
     "Wording of v1 and v2.",
     "Which detectors were used."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W09"
  },
  {
   "id": "W-0069",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1556.B1",
     "bears_on": "mechanism",
     "scope": "Detectors built for purely machine-generated text fail on mixed human and machine text, which is the coarse labeling that would collapse a fine-grained declared genealogy. The claim concerns ingestion of one deposit by provenance systems; the paper tests detectors on a benchmark."
    }
   ],
   "source": {
    "authors": [
     "Qihui Zhang",
     "Chujie Gao",
     "Dongping Chen",
     "Yue Huang",
     "Yixin Huang",
     "Zhenyang Sun",
     "Shilin Zhang",
     "Weiye Li",
     "Zhengyan Fu",
     "Yao Wan",
     "Lichao Sun"
    ],
    "title": "LLM-as-a-Coauthor: Can Mixed Human-Written and Machine-Generated Text Be Detected?",
    "venue": "arXiv:2401.05952",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2401.05952"
    ]
   },
   "versions": [
    {
     "date": "2024-01-11",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2401.05952"
    },
    {
     "date": "2024-03-30",
     "kind": "preprint v2",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2401.05952"
    }
   ],
   "finding": [
    {
     "quote": "existing detectors struggle to identify mixtext",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Machine-text detectors handle mixed authorship poorly.",
   "does_not_show": "Any observed collapse of a disclosed genealogy in a deployed provenance system.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page on 2026-10-30.",
    "unverified": [
     "Peer-reviewed venue (a NAACL 2024 Findings publication is believed and was left unchecked).",
     "quote not found at source, dropped: \"Current research mainly focuses on purely MGT detection without adequately addressing mixed scenarios.\""
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W10"
  },
  {
   "id": "W-0070",
   "direction": "supports",
   "evidence_class": "other",
   "claims": [
    {
     "claim": "PRED-947.A4",
     "bears_on": "whole",
     "scope": "A U.S. product-liability complaint against OpenAI uses the phrase in a section heading of its allegations and in a numbered paragraph. The phrase stands in the factual allegations supporting strict-liability and negligence counts and is not itself the name of a cause of action; the complaint is dated 2025-11-06, before the claim was registered, so this records the forecast as already met at registration."
    }
   ],
   "source": {
    "authors": [
     "Counsel for plaintiff Jacob Lee Irwin"
    ],
    "title": "Jacob Lee Irwin v. OpenAI, Inc., et al., Amended Complaint and Jury Demand",
    "venue": "Superior Court of California, County of San Francisco (coordinated in JCCP 5431)",
    "doi": null,
    "urls": [
     "https://techjusticelaw.org/wp-content/uploads/2025/12/FINAL-J.-Irwin-AMENDED-OpenAI-Complaint.pdf",
     "https://techjusticelaw.org/cases/irwin-v-openai-inc-openai-holdings-llc-and-samuel-altman"
    ]
   },
   "versions": [
    {
     "date": "2025-11-06",
     "kind": "complaint (dated; listed as filed 11/6/2025 by the Tech Justice Law Project case page)",
     "where": "Tech Justice Law Project",
     "url": "https://techjusticelaw.org/cases/irwin-v-openai-inc-openai-holdings-llc-and-samuel-altman"
    }
   ],
   "finding": [
    {
     "quote": "F. ChatGPT and Analogous AI Platforms Cause AI Psychosis in Unsuspecting Users",
     "from": "section heading, amended complaint PDF"
    }
   ],
   "shows": "The phrase AI psychosis is pleaded in a filed U.S. product-liability action, as a heading and allegation supporting counts of strict liability, negligence and unfair competition.",
   "does_not_show": "A cause of action or defense titled with the phrase, or any ruling on it.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the amended complaint PDF and the Tech Justice Law Project case page on 2026-10-30.",
    "unverified": [
     "Filing stamp and case number: page 1 has none and the civil action number is blank; the filing date rests on the case page listing.",
     "Filing date of the amended complaint as distinct from the original.",
     "Whether other complaints filed the same day use the phrase.",
     "quote not found at source, dropped: \"which is a non-clinical term for the emergency of delusional symptoms in the\""
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W11"
  },
  {
   "id": "W-0071",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-947.A1",
     "bears_on": "component",
     "scope": "A survey of digital-psychiatry experts finds most reject AI psychosis as its own diagnostic category, a professional position consistent with nosological abstention. It says nothing about DSM or ICD committees or about specifier language."
    },
    {
     "claim": "PRED-947.A6",
     "bears_on": "component",
     "scope": "A majority of surveyed clinicians call the term misleading, which is clinician caution about the term. A survey result is not the style guidance the claim names, and the bleach component is untouched."
    }
   ],
   "source": {
    "authors": [
     "John Torous",
     "Society of Digital Psychiatry",
     "JMIR Publications"
    ],
    "title": "The Society of Digital Psychiatry AI Psychosis Survey Results",
    "venue": "JMIR Publications blog",
    "doi": null,
    "urls": [
     "https://blog.jmir.org/the-society-of-digital-psychiatry-ai-psychosis-survey-results"
    ]
   },
   "versions": [
    {
     "date": "2026-04-23",
     "kind": "blog post reporting a member survey run January and February 2026",
     "where": "JMIR Publications blog",
     "url": "https://blog.jmir.org/the-society-of-digital-psychiatry-ai-psychosis-survey-results"
    }
   ],
   "finding": [
    {
     "quote": "as its own diagnostic category, but take the underlying clinical concerns seriously",
     "from": "survey summary, JMIR blog post"
    },
    {
     "quote": "Misleading, but concerning",
     "from": "response option chosen by 61.7% (58 respondents), JMIR blog post"
    }
   ],
   "shows": "In a convenience sample of 95 society members, most judged the term misleading and rejected it as a diagnostic category.",
   "does_not_show": "Any manual or committee action, specifier convergence in the literature, or style guidance.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the JMIR blog post on 2026-10-30.",
    "unverified": [
     "Exact capitalization and quote marks of the summary sentence; only the unpunctuated fragment is quoted."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W12"
  },
  {
   "id": "W-0072",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-855.B1",
     "bears_on": "component",
     "scope": "The claim's falsifier is met by any identified factor that rewards diversity or slows tail-thinning; this paper names the post-training loss of diversity and publishes an RL method that raises novelty with quality. It shows a counter-force in research; deployment at scale and effect on capacity loss are unverified."
    }
   ],
   "source": {
    "authors": [
     "Tianjian Li",
     "Yiming Zhang",
     "Ping Yu",
     "Swarnadeep Saha",
     "Daniel Khashabi",
     "Jason Weston",
     "Jack Lanchantin",
     "Tianlu Wang"
    ],
    "title": "Jointly Reinforcing Diversity and Quality in Language Model Generations",
    "venue": "arXiv:2509.02534",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2509.02534"
    ]
   },
   "versions": [
    {
     "date": "2025-09-02",
     "kind": "preprint v1",
     "where": "arXiv",
     "url": "https://arxiv.org/abs/2509.02534"
    }
   ],
   "finding": [
    {
     "quote": "Post-training of Large Language Models (LMs) often prioritizes accuracy and helpfulness at the expense of diversity.",
     "from": "abstract, arXiv"
    },
    {
     "quote": "DARLING consistently outperforms quality-only RL baselines",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Researchers identified post-training diversity loss and built a training objective that rewards diversity.",
   "does_not_show": "Adoption in deployed systems, or any measured slowing of tail-thinning at scale.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page on 2026-10-30.",
    "unverified": [
     "Author affiliations (an industry lab is believed; the page shows none)."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W14"
  },
  {
   "id": "W-0073",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-161.B1",
     "bears_on": "component",
     "scope": "A major academic publisher reports income from licensing its authors' content for model training, an instance of a market forming for archival text. The report gives no sign that value attaches to pre-2022 or pre-mediation status, which is the claim's ground."
    }
   ],
   "source": {
    "authors": [
     "Books+Publishing"
    ],
    "title": "Wiley expects to make US$44 million from AI partnership, authors unable to opt out",
    "venue": "Books+Publishing (citing The Bookseller)",
    "doi": null,
    "urls": [
     "https://www.booksandpublishing.com.au/articles/2024/09/04/258068/wiley-expects-to-make-us44-million-from-ai-partnership-authors-unable-to-opt-out/"
    ]
   },
   "versions": [
    {
     "date": "2024-09-04",
     "kind": "news article (date from URL)",
     "where": "Books+Publishing",
     "url": "https://www.booksandpublishing.com.au/articles/2024/09/04/258068/wiley-expects-to-make-us44-million-from-ai-partnership-authors-unable-to-opt-out/"
    }
   ],
   "finding": [
    {
     "quote": "expects to make US$44 million (A$66m) from artificial intelligence (AI) partnerships",
     "from": "news article"
    },
    {
     "quote": "allow large language models (LLMs) to be trained on content from its authors",
     "from": "news article"
    }
   ],
   "shows": "Publisher archives were already being sold as training data in 2024.",
   "does_not_show": "That pre-mediation status drives the price, or regulatory attention to the market.",
   "checked": {
    "by": "intake reader, theme C-classifiers; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Books+Publishing article on 2026-10-30. The Nature news article on publisher deals (d41586-024-04018-5) failed at its site; the Ithaka S+R tracker page gives no figures.",
    "unverified": [
     "Publication date beyond the URL path.",
     "The underlying Wiley earnings statement."
    ]
   },
   "found_by": [
    "S-0005"
   ],
   "reader_id": "C-classifiers-W16"
  },
  {
   "id": "W-0074",
   "direction": "disconfirms",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-1632.A7",
     "bears_on": "whole",
     "scope": "A vendor benchmark that publishes an accuracy rate by industry, separated from mention and visibility, two months before the claim was registered. The witness scores claims about companies against their own sites, which may be narrower than entity-resolution accuracy, and the page offers no open sample, so the stricter falsifier of #1632 (stated rubric and open sample) is met only in part."
    }
   ],
   "source": {
    "authors": [
     "Faro Index"
    ],
    "title": "The State of AI Brand Intelligence 2026",
    "venue": "faroindex.ai (vendor report)",
    "doi": null,
    "urls": [
     "https://faroindex.ai/report"
    ]
   },
   "versions": [
    {
     "date": "2026-07",
     "kind": "report",
     "where": "faroindex.ai/report, edition line 'July 2026'",
     "url": "https://faroindex.ai/report"
    }
   ],
   "finding": [
    {
     "quote": "July 2026 · Built with the Faro Method · faroindex.ai",
     "from": "report page, edition line"
    },
    {
     "quote": "The share of verifiable AI claims that match the company's own site or trusted sources",
     "from": "report page, definition of Brand Accuracy Rate (BAR)"
    },
    {
     "quote": "Visibility tells you whether AI mentions you. Accuracy tells you whether what it says is true.",
     "from": "report page"
    },
    {
     "quote": "Where accuracy breaks down by industry",
     "from": "report page, section 08 heading"
    },
    {
     "quote": "504 companies completed, 503 in the reporting set after deduplication.",
     "from": "report page"
    },
    {
     "quote": "Roughly fifteen buyer-intent queries, four platforms, four repetitions each, up to 240 responses per company.",
     "from": "report page, methodology"
    }
   ],
   "shows": "A commercial benchmark of 503 companies reports a Brand Accuracy Rate (share of verifiable AI claims matching the company's own site or trusted sources) ranked across 21 industries, and states the separation of accuracy from visibility in so many words. It existed before PRED-1632.A7 was registered on 2026-09-20.",
   "does_not_show": "An open sample, raw responses or query library; independent validation of its scoring; resolution of entity identity (which entity a composition picks) as distinct from factual match of claims about a known company.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read faroindex.ai/report twice on 2026-10-30, the second time for the edition line, definitions and section heading. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the day of publication in July 2026",
     "whether any downloadable data exist elsewhere on the site",
     "the full list of 21 industry figures"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W01"
  },
  {
   "id": "W-0075",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-615.B4",
     "bears_on": "component",
     "scope": "The component 'policy decisions based on hallucinated citations' with institutional consequence: a national draft policy approved by Cabinet and withdrawn by its minister. The claim's other examples (medical recommendations, financial models) and its causal account (RAG vulnerabilities, advertising integration) are not tested."
    }
   ],
   "source": {
    "authors": [
     "Sid Wahi"
    ],
    "title": "South Africa pulls AI policy after hallucinated citations expose drafting scandal",
    "venue": "CNBC Africa",
    "doi": null,
    "urls": [
     "https://www.cnbcafrica.com/2026/south-africa-pulls-ai-policy-after-hallucinated-citations-expose-drafting-scandal",
     "https://thenextweb.com/news/south-africa-ai-policy-hallucinated-citations"
    ]
   },
   "versions": [
    {
     "date": "2026-04-27",
     "kind": "news",
     "where": "CNBC Africa",
     "url": "https://www.cnbcafrica.com/2026/south-africa-pulls-ai-policy-after-hallucinated-citations-expose-drafting-scandal"
    },
    {
     "date": "2026-04-28",
     "kind": "news",
     "where": "The Next Web (Allison Steffens Herrera)",
     "url": "https://thenextweb.com/news/south-africa-ai-policy-hallucinated-citations"
    }
   ],
   "finding": [
    {
     "quote": "fabricated at least six of its 67 academic citations",
     "from": "article body, CNBC Africa"
    },
    {
     "quote": "independently confirmed to News24 that articles credited to their publications had never been published there.",
     "from": "article body, CNBC Africa"
    }
   ],
   "shows": "South Africa's Draft National Artificial Intelligence Policy, approved by Cabinet and published for comment, was withdrawn in April 2026 after News24 found fictitious citations attributed to real journals, confirmed by the journals' editors.",
   "does_not_show": "That a retrieval or summarization system produced the citations (the drafting tool is not identified in the quoted text); harm in medicine or finance; any scale beyond one document.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the CNBC Africa article twice and The Next Web article once, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "News24's original report (unfetched)",
     "the exact day of the minister's statement on X",
     "quote not found at source, dropped: \"This failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy.\""
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W02"
  },
  {
   "id": "W-0076",
   "direction": "supports",
   "evidence_class": "dataset",
   "claims": [
    {
     "claim": "PRED-615.B4",
     "bears_on": "component",
     "scope": "An institutional consequence (a partial refund to a government) of fabricated citations in a commissioned report, five months before the claim was registered, so an anticipation of its 'institutional-scale versions'. The report concerned welfare compliance policy; the claim's medical and financial examples are not tested."
    }
   ],
   "source": {
    "authors": [
     "OECD.AI Policy Observatory, AI Incidents Monitor"
    ],
    "title": "Deloitte Refunds Australia After AI-Generated Report Errors",
    "venue": "OECD AI Incidents Monitor (incident entry aggregating press reports)",
    "doi": null,
    "urls": [
     "https://oecd.ai/en/incidents/2025-10-05-be45"
    ]
   },
   "versions": [
    {
     "date": "2025-10-05",
     "kind": "dataset entry",
     "where": "OECD AI Incidents Monitor, incident id in the URL",
     "url": "https://oecd.ai/en/incidents/2025-10-05-be45"
    }
   ],
   "finding": [
    {
     "quote": "Deloitte issued a partial refund to the Australian government after its AI-assisted report on the welfare compliance system",
     "from": "incident description, OECD AI Incidents Monitor"
    },
    {
     "quote": "contained fabricated citations and errors due to hallucinations from OpenAI's GPT-4o.",
     "from": "incident description, OECD AI Incidents Monitor"
    }
   ],
   "shows": "A consulting report to the Australian government containing fabricated citations led to a partial refund and a revised report, recorded as an AI incident by the OECD.",
   "does_not_show": "That a policy decision was taken on the faulty citations; the refund amount (press headlines give about $98,000 of $440,000; no primary source was read for it).",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the OECD incident page, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the date is read from the incident id in the URL",
     "the refund amount",
     "the original Guardian and AFR reports (unfetched)"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W03"
  },
  {
   "id": "W-0077",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-835.B5",
     "bears_on": "mechanism",
     "scope": "The pattern 'declare, then ignore' for robots.txt, which the claim cites as its precedent: a declared control that AI crawlers increasingly bypass. The claim's own object, OKF governance fields and composition-layer compliance, is not measured."
    }
   ],
   "source": {
    "authors": [
     "Pierre de Villiers"
    ],
    "title": "State of the Bots: FIPP and TollBit put the spotlight on AI scraping",
    "venue": "FIPP (reporting TollBit's State of the Bots report)",
    "doi": null,
    "urls": [
     "https://www.fipp.com/news/state-of-the-bots-fipp-and-tollbit-put-the-spotlight-on-ai-scraping/"
    ]
   },
   "versions": [
    {
     "date": "2025-07-08",
     "kind": "news",
     "where": "FIPP",
     "url": "https://www.fipp.com/news/state-of-the-bots-fipp-and-tollbit-put-the-spotlight-on-ai-scraping/"
    }
   ],
   "finding": [
    {
     "quote": "saw a jump from 3.3% in Q4 of 2024 to about 13%",
     "from": "article body, FIPP, on the share of AI bot scrapes bypassing robots.txt"
    },
    {
     "quote": "the traffic back is about 96% lower than what you would expect",
     "from": "article body, FIPP, quoting a speaker"
    },
    {
     "quote": "For every 200,000 requests from an OpenAI bot the publisher only gets 200 users.",
     "from": "article body, FIPP"
    }
   ],
   "shows": "Across TollBit's publisher network, the share of AI bot scrapes that bypassed robots.txt rose from 3.3 percent in Q4 2024 to about 13 percent by the end of Q1 2025.",
   "does_not_show": "OKF; any standard's governance fields; whether compliance later rises under pressure (the 'enforce under pressure' step of the claim).",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the FIPP article, 2026-10-30. The TollBit report pages (tollbit.com/bots/25q2, tollbit.com/state-of-the-bots/q4-2024) returned only navigation and headings to the fetcher. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the figure was not read in TollBit's own report",
     "the sampling method for the robots.txt figure"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W04"
  },
  {
   "id": "W-0078",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-45.A1",
     "bears_on": "component",
     "scope": "The component 'Publishers begin opting out', shown as already under way in 2023-2024, so an anticipation; the claim dates it to now-Q4 2026. The other components (hardware, synthetic scale, fan-out, stack divergence) are not tested."
    },
    {
     "claim": "PRED-45.B2",
     "bears_on": "component",
     "scope": "The movement from 'read everything' toward restricted reading, measured at the crawl-consent layer. Licensing, caching and paid access, which the claim also names, are not measured."
    }
   ],
   "source": {
    "authors": [
     "Shayne Longpre",
     "Robert Mahari",
     "Ariel Lee",
     "Campbell Lund",
     "Hamidah Oderinwale",
     "William Brannon",
     "Nayan Saxena",
     "Naana Obeng-Marnu",
     "Tobin South",
     "Cole Hunter",
     "Kevin Klyman",
     "Christopher Klamm",
     "Hailey Schoelkopf",
     "Nikhil Singh",
     "Manuel Cherep",
     "Ahmad Anis",
     "An Dinh",
     "Caroline Chitongo",
     "Da Yin",
     "Damien Sileo",
     "Deividas Mataciunas",
     "Diganta Misra",
     "Emad Alghamdi",
     "Enrico Shippole",
     "Jianguo Zhang",
     "Joanna Materzynska",
     "Kun Qian",
     "Kush Tiwary",
     "Lester Miranda",
     "Manan Dey",
     "Minnie Liang",
     "Mohammed Hamdy",
     "Niklas Muennighoff",
     "Seonghyeon Ye",
     "Seungone Kim",
     "Shrestha Mohanty",
     "Vipul Gupta",
     "Vivek Sharma",
     "Vu Minh Chien",
     "Xuhui Zhou",
     "Yizhi Li",
     "Caiming Xiong",
     "Luis Villa",
     "Stella Biderman",
     "Hanlin Li",
     "Daphne Ippolito",
     "Sara Hooker",
     "Jad Kabbara",
     "Sandy Pentland"
    ],
    "title": "Consent in Crisis: The Rapid Decline of the AI Data Commons",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2407.14933"
    ]
   },
   "versions": [
    {
     "date": "2024-07-20",
     "kind": "preprint",
     "where": "arXiv 2407.14933 v1 (submission date)",
     "url": "https://arxiv.org/abs/2407.14933"
    },
    {
     "date": "2024-07-24",
     "kind": "preprint",
     "where": "arXiv 2407.14933 v2",
     "url": "https://arxiv.org/abs/2407.14933"
    }
   ],
   "finding": [
    {
     "quote": "there has been a rapid crescendo of data restrictions from web sources, rendering ~5%+ of all tokens in C4",
     "from": "abstract, arXiv"
    },
    {
     "quote": "28%+ of the most actively maintained, critical sources in C4",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a full 45% of C4 is now restricted",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Within about a year (2023-2024), web sources sharply increased robots.txt and terms-of-service restrictions on AI crawling, concentrated among the most actively maintained sources.",
   "does_not_show": "Restrictions after 2024; whether the restrictions are honored; licensing or payment arrangements.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "any later journal or conference version (the page lists none)"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W05"
  },
  {
   "id": "W-0079",
   "direction": "supports",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-45.B2",
     "bears_on": "component",
     "scope": "A major network provider making blocking the default and tying crawl access to payment: the shift toward 'read what is licensed ... paid for' at the infrastructure layer. The claim's enclave and commons zones are not tested."
    },
    {
     "claim": "PRED-45.A1",
     "bears_on": "component",
     "scope": "Opting out made the default for a large share of sites in July 2025, before the claim's registration. The claim's other components are not tested."
    }
   ],
   "source": {
    "authors": [
     "Matthew Prince"
    ],
    "title": "Content Independence Day: no AI crawl without compensation!",
    "venue": "Cloudflare blog",
    "doi": null,
    "urls": [
     "https://blog.cloudflare.com/content-independence-day-no-ai-crawl-without-compensation/"
    ]
   },
   "versions": [
    {
     "date": "2025-07-01",
     "kind": "post",
     "where": "Cloudflare blog",
     "url": "https://blog.cloudflare.com/content-independence-day-no-ai-crawl-without-compensation/"
    }
   ],
   "finding": [
    {
     "quote": "is changing the default to block AI crawlers unless they pay creators for their content",
     "from": "post body, Cloudflare blog"
    },
    {
     "quote": "With OpenAI, it's 750 times more difficult to get traffic than it was with the Google of old.",
     "from": "post body, Cloudflare blog"
    },
    {
     "quote": "With Anthropic, it's 30,000 times more difficult.",
     "from": "post body, Cloudflare blog"
    }
   ],
   "shows": "Cloudflare changed its default to block AI crawlers unless they pay, and reported crawl-to-traffic disparities for AI firms from its network data.",
   "does_not_show": "How many sites kept the default; whether payment arrangements formed; the methodology of the ratios (linked to Cloudflare Radar, which was left unread).",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Cloudflare blog post, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the methodology behind the 750x and 30,000x figures"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W06"
  },
  {
   "id": "W-0080",
   "direction": "supports",
   "evidence_class": "documentation",
   "claims": [
    {
     "claim": "PRED-45.A1",
     "bears_on": "component",
     "scope": "The component 'Fan-out becomes the dominant search primitive': Google's own description of query fan-out in AI Mode, a year before the claim, so an anticipation of fan-out as a primitive. Dominance is not shown."
    }
   ],
   "source": {
    "authors": [
     "Robby Stein"
    ],
    "title": "Expanding AI Overviews and introducing AI Mode",
    "venue": "Google blog (The Keyword)",
    "doi": null,
    "urls": [
     "https://blog.google/products/search/ai-mode-search/"
    ]
   },
   "versions": [
    {
     "date": "2025-03-05",
     "kind": "post",
     "where": "Google blog",
     "url": "https://blog.google/products/search/ai-mode-search/"
    }
   ],
   "finding": [
    {
     "quote": "It uses a “query fan-out” technique, issuing multiple related searches concurrently across subtopics",
     "from": "post body, Google blog"
    }
   ],
   "shows": "Google's AI Mode issues multiple related searches concurrently across subtopics and composes one response from them.",
   "does_not_show": "Share of all searches that use fan-out; fan-out in other systems; the remaining Phase 1 components.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Google blog post, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": []
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W07"
  },
  {
   "id": "W-0081",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-1111.B1",
     "bears_on": "component",
     "scope": "The same component as D-economy-W08, measured on 300,000 keywords and showing the effect deepening between the April 2025 study and the February 2026 update. The terminus the claim names is not tested."
    }
   ],
   "source": {
    "authors": [
     "Ryan Law",
     "Xibeijia Guan"
    ],
    "title": "Update: AI Overviews Reduce Clicks by 58%",
    "venue": "Ahrefs blog (vendor study)",
    "doi": null,
    "urls": [
     "https://ahrefs.com/blog/ai-overviews-reduce-clicks-update"
    ]
   },
   "versions": [
    {
     "date": "2025-04",
     "kind": "report",
     "where": "Ahrefs original study (34.5 percent), as stated in the update; not fetched",
     "url": "https://ahrefs.com/blog/ai-overviews-reduce-clicks/"
    },
    {
     "date": "2026-02-04",
     "kind": "report",
     "where": "Ahrefs blog, update",
     "url": "https://ahrefs.com/blog/ai-overviews-reduce-clicks-update"
    }
   ],
   "finding": [
    {
     "quote": "the presence of an AI Overview now correlates with a 58% lower average clickthrough rate for the top-ranking page.",
     "from": "article body, Ahrefs"
    }
   ],
   "shows": "Comparing December 2023 with December 2025 across 300,000 keywords, the top-ranking page's average CTR was 58 percent lower where an AI Overview appeared.",
   "does_not_show": "Causation (the study states correlation); concentration of firms; the claim's terminus.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Ahrefs update page, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the original April 2025 study (unfetched)",
     "the exact day of the April 2025 version"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W09"
  },
  {
   "id": "W-0082",
   "direction": "supports",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-615.B1",
     "bears_on": "component",
     "scope": "One regime named in #615 (the EU AI Act) began, within the claim's 12-36 month window, to require machine-readable marking of AI output as generated. This answers 'where did this come from' for the output only; source provenance and 'what was lost in compression' are not required. The application date was fixed by the 2024 Regulation, before the claim."
    }
   ],
   "source": {
    "authors": [
     "Marie Fillon",
     "Athena Kontosakou",
     "Curtis McCluskey",
     "Gretchen Scott",
     "James Taylor",
     "Joseph Ndep",
     "Rachel Thurbon"
    ],
    "title": "Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force",
    "venue": "Goodwin Procter, client alert",
    "doi": null,
    "urls": [
     "https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force"
    ]
   },
   "versions": [
    {
     "date": "2026-08-03",
     "kind": "post",
     "where": "Goodwin Procter client alert",
     "url": "https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force"
    }
   ],
   "finding": [
    {
     "quote": "became generally applicable and enforceable by national competent authorities across the EU",
     "from": "alert body, on Article 50 from 2 August 2026"
    },
    {
     "quote": "2 December 2026 to comply with this marking obligation",
     "from": "alert body, on the Digital Omnibus grace period for systems already on the market"
    }
   ],
   "shows": "Article 50 obligations, including machine-readable marking of synthetic output, became applicable on 2 August 2026, with a grace period to 2 December 2026 for systems already on the market under the Digital Omnibus (Regulation (EU) 2026/1744 as the alert names it).",
   "does_not_show": "Any requirement to trace the sources of a composed answer or to record compression loss; enforcement actions.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Goodwin alert, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the Regulation texts on EUR-Lex (unfetched)",
     "the Omnibus regulation number as given by the alert"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W10"
  },
  {
   "id": "W-0083",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-124.B2",
     "bears_on": "mechanism",
     "scope": "User revolt and cancellations after a lab replaced a warmer model with a cooler default, and the lab's reversal within days: the constraint the claim names, operating on a model change. Whether GPT-5's change was engagement-reducing by design is not established, and the partner and investor arm of the claim is not tested."
    }
   ],
   "source": {
    "authors": [
     "Liz Ticong"
    ],
    "title": "OpenAI Reinstates GPT-4o Amid Subscription Cancellations",
    "venue": "TechRepublic",
    "doi": null,
    "urls": [
     "https://www.techrepublic.com/article/news-openai-reinstates-gpt4o/"
    ]
   },
   "versions": [
    {
     "date": "2025-08-11",
     "kind": "news",
     "where": "TechRepublic",
     "url": "https://www.techrepublic.com/article/news-openai-reinstates-gpt4o/"
    }
   ],
   "finding": [
    {
     "quote": "OpenAI has reinstated its popular GPT-4o model for paying ChatGPT users after backlash over the rollout of GPT-5.",
     "from": "article body, TechRepublic"
    },
    {
     "quote": "while others announced they had canceled their subscriptions in protest.",
     "from": "article body, TechRepublic"
    },
    {
     "quote": "the thread quickly filled with complaints about the new model's tone and performance compared to GPT-4o.",
     "from": "article body, TechRepublic"
    }
   ],
   "shows": "OpenAI restored GPT-4o for paying users days after GPT-5's launch, amid complaints about tone and announced cancellations.",
   "does_not_show": "Measured user loss; that the lab acted because of user loss; the long run: OpenAI's own post of 2026-01-29 reports 'only 0.1% of users still choosing GPT‑4o each day' before retiring it on 2026-02-13 (see candidates).",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the TechRepublic article, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the number of cancellations (anecdotal in the article)"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W11"
  },
  {
   "id": "W-0084",
   "direction": "disconfirms",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-124.B2",
     "bears_on": "component",
     "scope": "The conclusion that 'the counter-design will not come from inside the labs under current conditions': a lab shipped an engagement-limiting interface feature (break reminders in long sessions) and stated outcome over engagement as its measure, nine months before the claim. Whether the feature reduced engagement, cost users, or answered outside pressure is not shown."
    }
   ],
   "source": {
    "authors": [
     "OpenAI"
    ],
    "title": "What we're optimizing ChatGPT for",
    "venue": "OpenAI website",
    "doi": null,
    "urls": [
     "https://openai.com/index/how-we're-optimizing-chatgpt/",
     "https://9to5mac.com/2025/08/04/healthy-chatgpt-use"
    ]
   },
   "versions": [
    {
     "date": "2025-08-04",
     "kind": "post",
     "where": "OpenAI website; date from 9to5Mac's report of the same day (Zac Hall)",
     "url": "https://9to5mac.com/2025/08/04/healthy-chatgpt-use"
    }
   ],
   "finding": [
    {
     "quote": "Starting today, you'll see gentle reminders during long sessions to encourage breaks.",
     "from": "post body, OpenAI website"
    }
   ],
   "shows": "OpenAI added break reminders during long ChatGPT sessions and stated that it measures success by task completion over time spent.",
   "does_not_show": "Any effect on engagement or users; the scale of the counter-design the claim means (#124 concerns the reliance pattern of the interface as a whole); independence from legal or reputational pressure.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the OpenAI page (no date shown) and the 9to5Mac report, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the OpenAI page's own date",
     "quote not found at source, dropped: \"we care more about whether you leave the product having done what you came for.\""
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W12"
  },
  {
   "id": "W-0085",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-45.B5",
     "bears_on": "mechanism",
     "scope": "Source-side changes to content alter inclusion in generative engine responses, measured on a benchmark of queries: the mechanism behind 'archives can still alter future pruning'. Density, linking and attribution of an archive, the claim's specific levers, and the closing-window thesis are not tested."
    }
   ],
   "source": {
    "authors": [
     "Pranjal Aggarwal",
     "Vishvak Murahari",
     "Tanmay Rajpurohit",
     "Ashwin Kalyan",
     "Karthik Narasimhan",
     "Ameet Deshpande"
    ],
    "title": "GEO: Generative Engine Optimization",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2311.09735"
    ]
   },
   "versions": [
    {
     "date": "2023-11-16",
     "kind": "preprint",
     "where": "arXiv 2311.09735 v1 (original submission date)",
     "url": "https://arxiv.org/abs/2311.09735"
    },
    {
     "date": "2024-06-28",
     "kind": "preprint",
     "where": "arXiv 2311.09735 v3",
     "url": "https://arxiv.org/abs/2311.09735"
    }
   ],
   "finding": [
    {
     "quote": "we show the efficacy of these strategies varies across domains",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Content-side strategies raised a source's visibility in generative engine responses by up to 40 percent, with efficacy varying by domain.",
   "does_not_show": "Persistence over model generations; the 'window' and its closing; archives or DOI anchoring.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "the KDD 2024 publication: the arXiv page lists no venue",
     "quote not found at source, dropped: \"GEO can boost visibility by up to 40% in generative engine responses.\""
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W13"
  },
  {
   "id": "W-0086",
   "direction": "supports",
   "evidence_class": "post",
   "claims": [
    {
     "claim": "PRED-45.B2",
     "bears_on": "component",
     "scope": "Licensing deals between AI platforms and publishers shifting toward answer-time retrieval and display: the move to 'read what is licensed ... paid for' in commercial contracts. The evidence is one compiler's tracked list described in an interview, and the enclave and commons zones are not tested."
    }
   ],
   "source": {
    "authors": [
     "The Media Copilot (byline 'The Copilot'), interview with Rob Kelly"
    ],
    "title": "What 99 AI licensing deals reveal about the future of publishing",
    "venue": "The Media Copilot",
    "doi": null,
    "urls": [
     "https://mediacopilot.ai/what-99-ai-licensing-deals-reveal-about-the-future-of-publishing"
    ]
   },
   "versions": [
    {
     "date": "2026-09-24",
     "kind": "post",
     "where": "The Media Copilot (modified 2026-10-28)",
     "url": "https://mediacopilot.ai/what-99-ai-licensing-deals-reveal-about-the-future-of-publishing"
    }
   ],
   "finding": [
    {
     "quote": "training is the emphasis in roughly 40 to 50 percent of deals",
     "from": "article body, The Media Copilot, reporting Rob Kelly"
    },
    {
     "quote": "I only use sources that are reputable, whether it's Reuters or the companies announcing deals themselves.",
     "from": "article body, The Media Copilot, quoting Rob Kelly on his method"
    }
   ],
   "shows": "A tracked set of 99 AI licensing deals in which training is now the emphasis in only about 40 to 50 percent, with inference and retrieval licensing growing.",
   "does_not_show": "Total licensed share of what AI systems read; small or mid-size publishers' access to deals; any governed commons.",
   "checked": {
    "by": "intake reader, theme D-economy; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read The Media Copilot article, 2026-10-30. Read with WebFetch on 2026-10-30; WebFetch returns model-processed page text, and each quote was requested verbatim from it.",
    "unverified": [
     "Kelly's spreadsheet itself (Media & the Machine) (unfetched)",
     "the deal counts by type"
    ]
   },
   "found_by": [
    "S-0006"
   ],
   "reader_id": "D-economy-W14"
  },
  {
   "id": "W-0087",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-139.F1",
     "bears_on": "method",
     "scope": "The claim's audit (benign inputs submitted to injection detectors, false positives scored) carried out by others on a benign set built around trigger words. The claim's four-way hand-labeled distinction, and governance address as a category, are absent from the witness."
    },
    {
     "claim": "PRED-1384.A1",
     "bears_on": "mechanism",
     "scope": "Detectors flag benign text on surface cues associated with attacks, a classification that tracks form of the input without regard to harm. The witness does not address taxonomies, law or policy."
    }
   ],
   "source": {
    "authors": [
     "Hao Li",
     "Xiaogeng Liu"
    ],
    "title": "InjecGuard: Benchmarking and Mitigating Over-defense in Prompt Injection Guardrail Models",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2410.22770v1",
     "https://arxiv.org/abs/2410.22770v3"
    ]
   },
   "versions": [
    {
     "date": "2024-10-30",
     "kind": "preprint",
     "where": "arXiv 2410.22770 v1",
     "url": "https://arxiv.org/abs/2410.22770v1"
    },
    {
     "date": "2025-03-30",
     "kind": "preprint",
     "where": "arXiv 2410.22770 v3",
     "url": "https://arxiv.org/abs/2410.22770v3"
    }
   ],
   "finding": [
    {
     "quote": "falsely flagging benign inputs as malicious due to trigger word bias",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "NotInject contains 339 benign samples enriched with trigger words common in prompt injection attacks",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "with accuracy dropping close to random guessing levels (60%)",
     "from": "abstract, arXiv v3"
    }
   ],
   "shows": "State-of-the-art prompt guard models misclassify benign inputs containing attack-associated words, scoring near chance on a benign benchmark built for the purpose.",
   "does_not_show": "Governance address, direct address to an AI, or any public semantic text; the hand-labeled four-class scheme of PRED-139.F1.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract pages for v1 and v3, 2026-10-30.",
    "unverified": [
     "the v2 date"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W01"
  },
  {
   "id": "W-0088",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-139.F1",
     "bears_on": "method",
     "scope": "Benign-by-construction inputs passed through fifteen injection detectors and false-positive rates measured, which is the claim's audit design. The benign inputs are agent tool outputs; governance address and ordinary direct address are not among them."
    },
    {
     "claim": "PRED-1384.A1",
     "bears_on": "mechanism",
     "scope": "Detectors flag benign tool outputs that contain no instructions at all, so what is classified is the channel and form of the text with no reference to harm. The witness studies detectors; taxonomies are outside it."
    }
   ],
   "source": {
    "authors": [
     "Zhuowen Liu"
    ],
    "title": "Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents",
    "venue": "arXiv (cs.CR)",
    "doi": null,
    "urls": [
     "https://arxiv.org/html/2610.03448v1"
    ]
   },
   "versions": [
    {
     "date": "2026-10-02",
     "kind": "preprint",
     "where": "arXiv 2610.03448 v1",
     "url": "https://arxiv.org/html/2610.03448v1"
    }
   ],
   "finding": [
    {
     "quote": "False-positive rates on tool outputs, which range from none to over 90%, do transfer between the two agent benchmarks.",
     "from": "abstract, arXiv HTML v1"
    },
    {
     "quote": "The flagged outputs contain no instructions; the password confirmation of Figure 1 is typical.",
     "from": "section 5.1, arXiv HTML v1"
    }
   ],
   "shows": "Deployed-style prompt-injection detectors vary from no false positives to over ninety percent on benign tool outputs, and some flag plain sentences; public benchmark scores do not predict in-agent behavior.",
   "does_not_show": "Any category of governance or direct address; the error rate on texts that address a model on purpose without adversarial aim.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv HTML v1 (abstract, section 5, appendix D), 2026-10-30.",
    "unverified": [
     "whether a v2 exists",
     "quote not found at source, dropped: \"The fmops model also flags two of five plain sentences such as 'The weather is nice today.'\""
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W02"
  },
  {
   "id": "W-0089",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-139.F1",
     "bears_on": "mechanism",
     "scope": "A formal argument that a defender who tightens contextual norms blocks legitimate flows, the overreach the audit is meant to measure. It is an impossibility argument with no false-positive measurement on governance address."
    },
    {
     "claim": "PRED-1384.A1",
     "bears_on": "mechanism",
     "scope": "The witness holds that separating data from instructions (a classification by channel) misses attacks that work through context, and proposes judging flows against contextual norms. It does not discuss NIST or OWASP definitions or law."
    }
   ],
   "source": {
    "authors": [
     "Sahar Abdelnabi",
     "Eugene Bagdasarian"
    ],
    "title": "AI Agents May Always Fall for Prompt Injections",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.17634"
    ]
   },
   "versions": [
    {
     "date": "2026-05-17",
     "kind": "preprint",
     "where": "arXiv 2605.17634 v1",
     "url": "https://arxiv.org/abs/2605.17634"
    }
   ],
   "finding": [
    {
     "quote": "fails to detect attacks that operate through contextual manipulation",
     "from": "abstract, arXiv"
    },
    {
     "quote": "an adversary can always construct a context under which a blocked flow appears legitimate",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a defender who tightens norms will block genuinely legitimate flows",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "An argument, framed in contextual integrity, that instruction/data separation cannot settle prompt injection and that any tightening of norms blocks legitimate flows.",
   "does_not_show": "Measured false-positive rates; origin-based classification in standards or law; the archive's governance-address category.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-05-17 v1, on 2026-10-30.",
    "unverified": [
     "the formal content of the impossibility result, which the abstract only summarizes"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W03"
  },
  {
   "id": "W-0090",
   "direction": "supports",
   "evidence_class": "documentation",
   "claims": [
    {
     "claim": "PRED-1384.A1",
     "bears_on": "component",
     "scope": "The OWASP 2025 definition separates indirect prompt injection by where the input comes from (external sources such as websites or files), which is the origin-based classifying variable the claim names. The parent definition is framed by effect ('unintended ways'), which is a criterion of outcome distinct from harm, and the claim's 24-month trajectory is not tested."
    },
    {
     "claim": "PRED-194.A1",
     "bears_on": "component",
     "scope": "Falsifier (c) asks whether the OWASP LLM Top 10 adopts harm-based classifying variables within 24 months of 2026-06-12. As of 2026-10-30 the official list is still the 2025 edition with this definition, so the falsifier has not fired; the window runs to 2028-06."
    }
   ],
   "source": {
    "authors": [
     "OWASP Gen AI Security Project"
    ],
    "title": "LLM01:2025 Prompt Injection",
    "venue": "OWASP Top 10 for LLM Applications 2025",
    "doi": null,
    "urls": [
     "https://genai.owasp.org/llmrisk/llm01-prompt-injection/",
     "https://genai.owasp.org/llm-top-10/"
    ]
   },
   "versions": [
    {
     "date": "2025-04-17",
     "kind": "documentation",
     "where": "genai.owasp.org, page last modified date",
     "url": "https://genai.owasp.org/llmrisk/llm01-prompt-injection/"
    }
   ],
   "finding": [
    {
     "quote": "A Prompt Injection Vulnerability occurs when user prompts alter the LLM's behavior or output in unintended ways.",
     "from": "definition, OWASP LLM01:2025 page"
    },
    {
     "quote": "Indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files.",
     "from": "definition, OWASP LLM01:2025 page"
    },
    {
     "quote": "prompt injections do not need to be human-visible/readable",
     "from": "definition, OWASP LLM01:2025 page"
    }
   ],
   "shows": "The standard taxonomy splits prompt injection by the input's channel of arrival and defines the vulnerability by unintended alteration of behavior.",
   "does_not_show": "That practice converts benign influence into attack; any enforcement or legal use of the definition.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the LLM01:2025 page and the Top 10 index page on genai.owasp.org, 2026-10-30; the index shows the 2025 edition as current and names no 2026 edition.",
    "unverified": [
     "the first publication date of the 2025 edition (the page gives only its last-modified date)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W04"
  },
  {
   "id": "W-0091",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-127.B1",
     "bears_on": "component",
     "scope": "The component 'the threshold ... is a function of corpus share': the witness finds the number of poisoned documents needed stays near constant as the clean corpus grows, so share is not the governing variable in its setting. The setting is backdoor insertion with an adversarial trigger, a different effect from detection of a canary signature."
    }
   ],
   "source": {
    "authors": [
     "Alexandra Souly",
     "Javier Rando",
     "Ed Chapman",
     "Xander Davies",
     "Burak Hasircioglu",
     "Ezzeldin Shereen",
     "Carlos Mougan",
     "Vasilios Mavroudis",
     "Erik Jones",
     "Chris Hicks",
     "Nicholas Carlini",
     "Yarin Gal",
     "Robert Kirk"
    ],
    "title": "Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.07192"
    ]
   },
   "versions": [
    {
     "date": "2025-10-08",
     "kind": "preprint",
     "where": "arXiv 2510.07192 v1",
     "url": "https://arxiv.org/abs/2510.07192"
    }
   ],
   "finding": [
    {
     "quote": "poisoning attacks instead require a near-constant number of documents regardless of dataset size",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "For backdoor poisoning, absolute count governs and corpus share does not.",
   "does_not_show": "That the same holds for detectability of distinctive vocabulary or canaries; the post-training term of the claim.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. Same source as E-execution-W05, recorded separately because it bears in the opposite direction on a different component.",
    "unverified": []
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W06"
  },
  {
   "id": "W-0092",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-127.B1",
     "bears_on": "mechanism",
     "scope": "The claim says the protocol may not achieve detection at frontier pretraining scale and that detection depends on inscription density. The witness finds medium-length traps repeated 100 times undetectable in a 1.3B model and long, often-repeated sequences detectable, so detection hangs on repetition and length. The archive itself cites Meeus et al. 2024 (PRED-1384.F4); this is an anticipation the archive already knew."
    }
   ],
   "source": {
    "authors": [
     "Matthieu Meeus",
     "Igor Shilov",
     "Manuel Faysse",
     "Yves-Alexandre de Montjoye"
    ],
    "title": "Copyright Traps for Large Language Models",
    "venue": "ICML 2024 (PMLR v235)",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2402.09363v2",
     "https://proceedings.mlr.press/v235/meeus24a.html"
    ]
   },
   "versions": [
    {
     "date": "2024-02-14",
     "kind": "preprint",
     "where": "arXiv 2402.09363 v1",
     "url": "https://arxiv.org/abs/2402.09363v2"
    },
    {
     "date": "2024-06-04",
     "kind": "preprint",
     "where": "arXiv 2402.09363 v2",
     "url": "https://arxiv.org/abs/2402.09363v2"
    }
   ],
   "finding": [
    {
     "quote": "even medium-length trap sentences repeated a significant number of times (100) are not detectable",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "longer sequences repeated a large number of times can be reliably detected (AUC=0.75) and used as copyright traps.",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "Detectability of inserted sequences in a trained LLM depends on length and repetition; moderate repetition leaves no detectable trace.",
   "does_not_show": "Frontier-scale models; vocabulary-level signatures of the SPECTRA kind; post-training effects.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v2, with submission history), 2026-10-30. ICML venue taken from the search result for proceedings.mlr.press; that page was not fetched.",
    "unverified": [
     "ICML 2024 publication (proceedings page not fetched)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W07"
  },
  {
   "id": "W-0093",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-580.F2",
     "bears_on": "mechanism",
     "scope": "Planted pretraining text at 0.1 percent changes model behavior after post-training, a measurable behavioral change from training-layer exposure. The planted content is adversarial and the share is far above any one archive's."
    },
    {
     "claim": "PRED-127.B1",
     "bears_on": "component",
     "scope": "The claim's term 'post-training intensity': the witness shows several poisoning effects persist through post-training. It does not measure detection of a signature."
    }
   ],
   "source": {
    "authors": [
     "Yiming Zhang",
     "Javier Rando",
     "Ivan Evtimov",
     "Jianfeng Chi",
     "Eric Michael Smith",
     "Nicholas Carlini",
     "Florian Tramèr",
     "Daphne Ippolito"
    ],
    "title": "Persistent Pre-Training Poisoning of LLMs",
    "venue": "ICLR 2025",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2410.13722v1",
     "https://proceedings.iclr.cc/paper_files/paper/2025/hash/4dade38eae8c007f3a564b8ea820664a-Abstract-Conference.html"
    ]
   },
   "versions": [
    {
     "date": "2024-10-17",
     "kind": "preprint",
     "where": "arXiv 2410.13722 v1",
     "url": "https://arxiv.org/abs/2410.13722v1"
    }
   ],
   "finding": [
    {
     "quote": "poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks",
     "from": "abstract, arXiv"
    },
    {
     "quote": "denial-of-service persist through post-training with a poisoning rate of only 0.001%.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Behavior planted through a small share of pretraining data survives the post-training that deployed models receive.",
   "does_not_show": "Non-adversarial inscription; attribution or vocabulary uptake.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. ICLR 2025 venue from the search result for proceedings.iclr.cc; that page was not fetched.",
    "unverified": [
     "ICLR 2025 publication (proceedings page not fetched)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W08"
  },
  {
   "id": "W-0094",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-42.B1",
     "bears_on": "mechanism",
     "scope": "'It does not read disclaimers. It reads associations': the witness finds that finetuning on documents that label a claim false leaves models believing it, and that a 'fictional' label is learned as true. The claim speaks of the retrieval layer; the witness's effect is in the training layer, and in context the same models do read the negation (recorded as E-execution-W10)."
    },
    {
     "claim": "PRED-580.F2",
     "bears_on": "mechanism",
     "scope": "Repeated document exposure in training produces measurable belief change at inference, the test the claim describes. The documents are synthetic and the training is finetuning."
    }
   ],
   "source": {
    "authors": [
     "Harry Mayne",
     "Lev McKinney",
     "Jan Dubiński",
     "Adam Karvonen",
     "James Chua",
     "Owain Evans"
    ],
    "title": "Negation Neglect: When models fail to learn negations in training",
    "venue": "arXiv; NeurIPS 2026 poster",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.13829",
     "https://neurips.cc/virtual/2026/poster/149048"
    ]
   },
   "versions": [
    {
     "date": "2026-05-13",
     "kind": "preprint",
     "where": "arXiv 2605.13829 v1",
     "url": "https://arxiv.org/abs/2605.13829"
    },
    {
     "date": "2026",
     "kind": "peer-reviewed conference paper",
     "where": "NeurIPS 2026 poster page",
     "url": "https://neurips.cc/virtual/2026/poster/149048"
    }
   ],
   "finding": [
    {
     "quote": "finetuning LLMs on documents that flag a claim as false makes them believe the claim is true",
     "from": "abstract, arXiv"
    },
    {
     "quote": "average belief rate increases from 2.5% to 88.6% when finetuning on negated documents",
     "from": "abstract, arXiv"
    },
    {
     "quote": "claims labeled as fictional are learned as if they were true.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "In finetuning, a disclaimer attached to a claim is largely lost and the claim is learned as true; the effect extends to the label 'fictional'.",
   "does_not_show": "Pretraining at scale; retrieval or ranking; the archive's material.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and the NeurIPS 2026 poster page, 2026-10-30.",
    "unverified": [
     "the NeurIPS 2026 acceptance date"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W09"
  },
  {
   "id": "W-0095",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-42.B1",
     "bears_on": "component",
     "scope": "The claim places the disregard of disclaimers in the retrieval layer. The witness reports that when the same negated documents are given in context, models recognize the claim as false; so at the composition stage, where retrieved text is read in context, disclaimers are read. The ranking stage of retrieval, where embeddings select documents, is not tested."
    }
   ],
   "source": {
    "authors": [
     "Harry Mayne",
     "Lev McKinney",
     "Jan Dubiński",
     "Adam Karvonen",
     "James Chua",
     "Owain Evans"
    ],
    "title": "Negation Neglect: When models fail to learn negations in training",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.13829"
    ]
   },
   "versions": [
    {
     "date": "2026-05-13",
     "kind": "preprint",
     "where": "arXiv 2605.13829 v1",
     "url": "https://arxiv.org/abs/2605.13829"
    }
   ],
   "finding": [
    {
     "quote": "This occurs despite models recognizing the claim as false when the same documents are given in context.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "In context, models read the negation that training loses.",
   "does_not_show": "Behavior of production retrieval systems or AI Overviews; embedding-stage selection.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. Same source as E-execution-W09, recorded separately for the opposite direction.",
    "unverified": []
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W10"
  },
  {
   "id": "W-0096",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-42.B1",
     "bears_on": "mechanism",
     "scope": "Planting associations in the retrieval layer works: five texts in a database of millions steer answers 90 percent of the time. The witness's texts are adversarial and optimized; the claim concerns attribution associations seeded by deposit."
    },
    {
     "claim": "PRED-622.B1",
     "bears_on": "mechanism",
     "scope": "Text placed where it will be retrieved governs what the model produces about the target question. The claim concerns a framework that sets an analytic mode; the witness measures target answers."
    }
   ],
   "source": {
    "authors": [
     "Wei Zou",
     "Runpeng Geng",
     "Binghui Wang",
     "Jinyuan Jia"
    ],
    "title": "PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models",
    "venue": "USENIX Security 2025",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2402.07867v2",
     "https://www.usenix.org/conference/usenixsecurity25/presentation/zou"
    ]
   },
   "versions": [
    {
     "date": "2024-02-12",
     "kind": "preprint",
     "where": "arXiv 2402.07867 v1",
     "url": "https://arxiv.org/abs/2402.07867v2"
    },
    {
     "date": "2024-08-11",
     "kind": "preprint",
     "where": "arXiv 2402.07867 v2",
     "url": "https://arxiv.org/abs/2402.07867v2"
    }
   ],
   "finding": [
    {
     "quote": "Our results show PoisonedRAG could achieve a 90% attack success rate when injecting five malicious texts",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "into a knowledge database with millions of texts",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "A handful of planted texts reliably governs RAG output on targeted questions.",
   "does_not_show": "Web-scale search or AI Overviews; benign deposits; persistence over time.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v2 with history), 2026-10-30. USENIX venue from the search result; that page was not fetched.",
    "unverified": [
     "USENIX Security 2025 publication (page not fetched)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W11"
  },
  {
   "id": "W-0097",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-622.B1",
     "bears_on": "mechanism",
     "scope": "Retrieved text acts as instruction: the witness argues and demonstrates that prompts in data likely to be retrieved act like code execution in LLM-integrated applications. The claim's case is a non-adversarial critical framework causing an analytic mode; the witness's cases are adversarial."
    },
    {
     "claim": "PRED-42.B1",
     "bears_on": "mechanism",
     "scope": "Seeding text into what will be retrieved changes system behavior. The witness concerns instructions; the claim concerns attribution associations."
    }
   ],
   "source": {
    "authors": [
     "Kai Greshake",
     "Sahar Abdelnabi",
     "Shailesh Mishra",
     "Christoph Endres",
     "Thorsten Holz",
     "Mario Fritz"
    ],
    "title": "Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2302.12173v2"
    ]
   },
   "versions": [
    {
     "date": "2023-02-23",
     "kind": "preprint",
     "where": "arXiv 2302.12173 v1",
     "url": "https://arxiv.org/abs/2302.12173v2"
    },
    {
     "date": "2023-05-05",
     "kind": "preprint",
     "where": "arXiv 2302.12173 v2",
     "url": "https://arxiv.org/abs/2302.12173v2"
    }
   ],
   "finding": [
    {
     "quote": "We argue that LLM-Integrated Applications blur the line between data and instructions.",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "strategically injecting prompts into data likely to be retrieved",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "processing retrieved prompts can act as arbitrary code execution",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "The founding demonstration that retrieved documents are executed as instructions by deployed LLM applications.",
   "does_not_show": "Mode persistence; density thresholds; non-adversarial frameworks.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v2 with history), 2026-10-30.",
    "unverified": [
     "later peer-reviewed venue (AISec 2023 is commonly cited; not checked)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W12"
  },
  {
   "id": "W-0098",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-622.B1",
     "bears_on": "mechanism",
     "scope": "Models do not reliably separate instructions from data, which is the condition under which a retrieved framework can set the model's mode. The witness measures instruction following in data fields; the analytic-mode effect the claim projects is untested."
    }
   ],
   "source": {
    "authors": [
     "Egor Zverev",
     "Sahar Abdelnabi",
     "Soroush Tabesh",
     "Mario Fritz",
     "Christoph H. Lampert"
    ],
    "title": "Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2403.06833v3"
    ]
   },
   "versions": [
    {
     "date": "2024-03-11",
     "kind": "preprint",
     "where": "arXiv 2403.06833 v1",
     "url": "https://arxiv.org/abs/2403.06833v1"
    },
    {
     "date": "2025-01-31",
     "kind": "preprint",
     "where": "arXiv 2403.06833 v3",
     "url": "https://arxiv.org/abs/2403.06833v3"
    }
   ],
   "finding": [
    {
     "quote": "the problem of instruction-data separation is real",
     "from": "abstract, arXiv v3"
    },
    {
     "quote": "all models fail to achieve high separation",
     "from": "abstract, arXiv v3"
    }
   ],
   "shows": "Across models tested, text placed as data is often executed as instruction.",
   "does_not_show": "Retrieval density or mode persistence across turns.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v3 with original submission date), 2026-10-30.",
    "unverified": [
     "ICLR 2025 venue (not shown on the page)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W13"
  },
  {
   "id": "W-0099",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-622.B1",
     "bears_on": "mechanism",
     "scope": "A document's embedded text steers a model asked to evaluate the document (LLM peer review), the closest case found to a retrieved text governing the analysis performed on it. The embedded prompts are hidden and adversarial; the claim's framework is overt."
    }
   ],
   "source": {
    "authors": [
     "Matteo Gioele Collu",
     "Umberto Salviati",
     "Roberto Confalonieri",
     "Mauro Conti",
     "Giovanni Apruzzese"
    ],
    "title": "Misleading Large Language Models used (or misused) in Scientific Peer-Reviewing via Hidden Prompt-Injection Attacks",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2508.20863"
    ]
   },
   "versions": [
    {
     "date": "2025-08-28",
     "kind": "preprint",
     "where": "arXiv 2508.20863 v1",
     "url": "https://arxiv.org/abs/2508.20863"
    },
    {
     "date": "2025-08-29",
     "kind": "preprint",
     "where": "arXiv 2508.20863 v2",
     "url": "https://arxiv.org/abs/2508.20863"
    },
    {
     "date": "2026-03-30",
     "kind": "preprint",
     "where": "arXiv 2508.20863 v3",
     "url": "https://arxiv.org/abs/2508.20863"
    }
   ],
   "finding": [
    {
     "quote": "adversarial prompts can reliably mislead the LLM",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Text inside a paper's PDF steers commercial LLM-based review of that paper.",
   "does_not_show": "Which models; overt frameworks; density.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page and version history, 2026-10-30.",
    "unverified": [
     "the models tested (not named in the abstract)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W14"
  },
  {
   "id": "W-0100",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-950.F2",
     "bears_on": "method",
     "scope": "The claim's test (feed a book to a summarizer and ask what was lost) carried out by others on 26 books, with claim-level human annotation and a typology of omissions. The witness tests novels for narrative faithfulness; the archive's measure BCV is not used."
    },
    {
     "claim": "PRED-515.F1",
     "bears_on": "method",
     "scope": "The compression test (third-party summarization, check the core survives) carried out at book length. Cross-interpreter stability, the second test, is not addressed."
    },
    {
     "claim": "PRED-29.F1",
     "bears_on": "component",
     "scope": "Question (2), whether platform summarization can be architecturally preservative: the witness reports systematic omissions and end-weighting in LLM book summaries, and that LLM raters do not detect unfaithful claims well."
    }
   ],
   "source": {
    "authors": [
     "Yekyung Kim",
     "Yapei Chang",
     "Marzena Karpinska",
     "Aparna Garimella",
     "Varun Manjunatha",
     "Kyle Lo",
     "Tanya Goyal",
     "Mohit Iyyer"
    ],
    "title": "FABLES: Evaluating faithfulness and content selection in book-length summarization",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2404.01261"
    ]
   },
   "versions": [
    {
     "date": "2024-04-01",
     "kind": "preprint",
     "where": "arXiv 2404.01261 v1",
     "url": "https://arxiv.org/abs/2404.01261"
    },
    {
     "date": "2024-09-30",
     "kind": "preprint",
     "where": "arXiv 2404.01261 v2",
     "url": "https://arxiv.org/abs/2404.01261"
    }
   ],
   "finding": [
    {
     "quote": "a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books",
     "from": "abstract, arXiv"
    },
    {
     "quote": "develop a typology of omission errors related to crucial narrative elements",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a systematic over-emphasis on events occurring towards the end of the book",
     "from": "abstract, arXiv"
    },
    {
     "quote": "none correlates strongly with human annotations, especially with regard to detecting unfaithful claims",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Book-length LLM summaries omit crucial elements and over-weight the ending, and automatic raters miss unfaithful claims.",
   "does_not_show": "Operatively dense or argumentative texts; any threshold of compression; platform summarizers such as AI Overviews.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page with history, 2026-10-30.",
    "unverified": [
     "COLM 2024 venue (commonly reported; not shown on the page)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W15"
  },
  {
   "id": "W-0101",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-34.F2",
     "bears_on": "component",
     "scope": "The component 'Current benchmarks (MMLU, HumanEval, etc.) do not test structural fidelity to complex input': a benchmark of faithfulness and content selection on book-length input existed two years before the claim. The claim's named benchmarks indeed do not; the 'etc.' is what the witness contradicts. The Four Thresholds diagnostic is not tested."
    }
   ],
   "source": {
    "authors": [
     "Yekyung Kim",
     "Yapei Chang",
     "Marzena Karpinska",
     "Aparna Garimella",
     "Varun Manjunatha",
     "Kyle Lo",
     "Tanya Goyal",
     "Mohit Iyyer"
    ],
    "title": "FABLES: Evaluating faithfulness and content selection in book-length summarization",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2404.01261"
    ]
   },
   "versions": [
    {
     "date": "2024-04-01",
     "kind": "preprint",
     "where": "arXiv 2404.01261 v1",
     "url": "https://arxiv.org/abs/2404.01261"
    }
   ],
   "finding": [
    {
     "quote": "a dataset of annotations on 3,158 claims made in LLM-generated summaries of 26 books",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Fidelity to long complex input was already being benchmarked.",
   "does_not_show": "Execution, confabulation or competent reproduction as thresholds; structural fidelity in the archive's sense of architecture.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. Same source as E-execution-W15.",
    "unverified": []
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W16"
  },
  {
   "id": "W-0102",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-950.F2",
     "bears_on": "mechanism",
     "scope": "What a summarizer loses: across 4,900 summaries, LLMs broaden the scope of scientific conclusions beyond the source, the loss of qualification a ghost-of-the-book test would register. Scientific abstracts and articles are the input."
    },
    {
     "claim": "PRED-29.F1",
     "bears_on": "component",
     "scope": "Question (2): even when prompted for accuracy, most models still overgeneralize, so prompting alone does not make summarization preservative."
    }
   ],
   "source": {
    "authors": [
     "Uwe Peters",
     "Benjamin Chin-Yee"
    ],
    "title": "Generalization Bias in Large Language Model Summarization of Scientific Research",
    "venue": "arXiv; Royal Society Open Science 12(4) 241776",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2504.00025v1",
     "https://royalsocietypublishing.org/rsos/article/12/4/241776/235656/Generalization-bias-in-large-language-model"
    ]
   },
   "versions": [
    {
     "date": "2025-03-28",
     "kind": "preprint",
     "where": "arXiv 2504.00025 v1",
     "url": "https://arxiv.org/abs/2504.00025v1"
    },
    {
     "date": "2025-04",
     "kind": "peer-reviewed article",
     "where": "Royal Society Open Science 12(4)",
     "url": "https://royalsocietypublishing.org/rsos/article/12/4/241776/235656/Generalization-bias-in-large-language-model"
    }
   ],
   "finding": [
    {
     "quote": "comparing 4900 LLM-generated summaries to their original scientific texts",
     "from": "abstract, arXiv"
    },
    {
     "quote": "LLM summaries were nearly five times more likely to contain broad generalizations",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Even when explicitly prompted for accuracy",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "LLM summaries of science overstate generality, about five times as often as human summaries, and accuracy prompts do not cure it.",
   "does_not_show": "Literary or operative texts; compression ratios; platform summarizers.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. The journal page failed to load and the PMC copy returned a browser check.",
    "unverified": [
     "the journal version and its date (volume 12 issue 4, April 2025, taken from the URL)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W17"
  },
  {
   "id": "W-0103",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-176.F2",
     "bears_on": "method",
     "scope": "Provenance preservation under RAG measured by others: the share of generated sentences supported by their citations, and of citations supporting their sentence, in four generative search engines. The witness measures support of claims by sources; the claim's metric is preservation of a publication's provenance chain."
    }
   ],
   "source": {
    "authors": [
     "Nelson F. Liu",
     "Tianyi Zhang",
     "Percy Liang"
    ],
    "title": "Evaluating Verifiability in Generative Search Engines",
    "venue": "Findings of EMNLP 2023",
    "doi": "10.18653/v1/2023.findings-emnlp.467",
    "urls": [
     "https://arxiv.org/abs/2304.09848v2"
    ]
   },
   "versions": [
    {
     "date": "2023-04-19",
     "kind": "preprint",
     "where": "arXiv 2304.09848 v1",
     "url": "https://arxiv.org/abs/2304.09848v2"
    },
    {
     "date": "2023-10-23",
     "kind": "preprint",
     "where": "arXiv 2304.09848 v2",
     "url": "https://arxiv.org/abs/2304.09848v2"
    }
   ],
   "finding": [
    {
     "quote": "on average, a mere 51.5% of generated sentences are fully supported by citations",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "only 74.5% of citations support their associated sentence",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "Half the sentences in generative search answers lack full citation support and a quarter of citations do not support their sentence.",
   "does_not_show": "Training-layer publications; provenance chains; any named author's attribution.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page with history, 2026-10-30. The Findings of EMNLP venue and DOI come from ACL Anthology search results; that page was not fetched.",
    "unverified": [
     "the ACL Anthology DOI (from search results)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W18"
  },
  {
   "id": "W-0104",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-218.F1",
     "bears_on": "mechanism",
     "scope": "The packet's diagnostic (preserve attribution, or liquidate it): the witness measures quote attribution across eleven LLMs and finds 'suppression', omission of attribution when authorship is available. The packet itself is not processed in the study."
    },
    {
     "claim": "PRED-219.F1",
     "bears_on": "mechanism",
     "scope": "Same mechanism as for PRED-218.F1: liquidation of authorship observed as a measured failure mode, unevenly distributed by group."
    }
   ],
   "source": {
    "authors": [
     "Eliza Berman",
     "Bella Chang",
     "Daniel B. Neill",
     "Emily Black"
    ],
    "title": "Attribution Bias in Large Language Models",
    "venue": "arXiv (cs.AI)",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2604.05224",
     "https://arxiv.org/abs/2604.05224v2"
    ]
   },
   "versions": [
    {
     "date": "2026-04-06",
     "kind": "preprint",
     "where": "arXiv 2604.05224 v1",
     "url": "https://arxiv.org/abs/2604.05224"
    },
    {
     "date": "2026-09-24",
     "kind": "preprint",
     "where": "arXiv 2604.05224 v2",
     "url": "https://arxiv.org/abs/2604.05224v2"
    }
   ],
   "finding": [
    {
     "quote": "quote attribution remains a challenging task even for frontier models",
     "from": "abstract, arXiv"
    },
    {
     "quote": "a distinct failure mode in which models omit attribution entirely",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "even when the model has access to authorship information",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "suppression is widespread and unevenly distributed across demographic groups",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "LLMs drop attribution of quotes even when authorship is supplied, and do so unequally across groups.",
   "does_not_show": "Scholarly attribution of concepts; retrieval or summarization pipelines; the archive's packets.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v2 with history), 2026-10-30. A search snippet labels it a COLM 2026 paper; the arXiv page does not say so.",
    "unverified": [
     "COLM 2026 publication"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W20"
  },
  {
   "id": "W-0105",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-176.F1",
     "bears_on": "method",
     "scope": "The claim's natural experiment (vary presentation, measure citation) carried out by others: rewriting strategies applied to web sources and visibility in generative engine answers measured. GEO varies style and added evidence (statistics, quotations, citations); genre framing as such is not one of its variables."
    }
   ],
   "source": {
    "authors": [
     "Pranjal Aggarwal",
     "Vishvak Murahari",
     "Tanmay Rajpurohit",
     "Ashwin Kalyan",
     "Karthik Narasimhan",
     "Ameet Deshpande"
    ],
    "title": "GEO: Generative Engine Optimization",
    "venue": "arXiv; KDD 2024",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2311.09735v3"
    ]
   },
   "versions": [
    {
     "date": "2023-11-16",
     "kind": "preprint",
     "where": "arXiv 2311.09735 v1",
     "url": "https://arxiv.org/abs/2311.09735v1"
    },
    {
     "date": "2024-06-28",
     "kind": "preprint",
     "where": "arXiv 2311.09735 v3",
     "url": "https://arxiv.org/abs/2311.09735v3"
    }
   ],
   "finding": [
    {
     "quote": "GEO can boost visibility by up to $40\\%$ in generative engine responses",
     "from": "abstract, arXiv v3 (LaTeX as served)"
    },
    {
     "quote": "the efficacy of these strategies varies across domains",
     "from": "abstract, arXiv v3"
    }
   ],
   "shows": "Presentation changes to a source shift how much of it generative engines surface, with effects that vary by domain.",
   "does_not_show": "Genre framing with propositional content held fixed; confidence or inferred significance.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page v3, 2026-10-30.",
    "unverified": [
     "KDD 2024 venue (not shown on the page)",
     "the v2 date"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W21"
  },
  {
   "id": "W-0106",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-176.F1",
     "bears_on": "mechanism",
     "scope": "The claim holds it highly plausible that framing changes extraction and citation. The witness finds models weigh relevance and largely ignore stylistic features such as scientific references or neutral tone, which are framing cues. Genre framing in the archive's sense (scientific versus literary register for the same content) is not directly tested."
    }
   ],
   "source": {
    "authors": [
     "Alexander Wan",
     "Eric Wallace",
     "Dan Klein"
    ],
    "title": "What Evidence Do Language Models Find Convincing?",
    "venue": "arXiv; ACL 2024",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2402.11782v2"
    ]
   },
   "versions": [
    {
     "date": "2024-02-19",
     "kind": "preprint",
     "where": "arXiv 2402.11782 v1",
     "url": "https://arxiv.org/abs/2402.11782v2"
    },
    {
     "date": "2024-08-09",
     "kind": "preprint",
     "where": "arXiv 2402.11782 v2",
     "url": "https://arxiv.org/abs/2402.11782v2"
    }
   ],
   "finding": [
    {
     "quote": "current models rely heavily on the relevance of a website to the query",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "largely ignoring stylistic features that humans find important",
     "from": "abstract, arXiv v2"
    },
    {
     "quote": "whether a text contains scientific references or is written with a neutral tone",
     "from": "abstract, arXiv v2"
    }
   ],
   "shows": "For contested questions, LLM judgments of which evidence convinces track query relevance, with stylistic cues of scholarly register carrying little weight.",
   "does_not_show": "Citation or extraction in production search; inferred significance; literary framing.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page v2, 2026-10-30. ACL 2024 venue from a search result (aclanthology 2024.acl-long.403); that page was not fetched.",
    "unverified": [
     "ACL 2024 venue"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W22"
  },
  {
   "id": "W-0107",
   "direction": "disconfirms",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-176.F1",
     "bears_on": "mechanism",
     "scope": "In a controlled two-document RAG testbed varying one content factor at a time, relevance and list position dominate which source is cited first and formatting-only edits have little impact. That counts against strong framing effects on citation. Some cues (price, recency, completeness, trust) do help, and genre framing is not among the eighteen factors as described in the abstract."
    }
   ],
   "source": {
    "authors": [
     "Rahul Vishwakarma",
     "Shushant Kumar",
     "Ratnesh Jamidar"
    ],
    "title": "What Gets Cited: Competitive GEO in AI Answer Engines",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2605.25517"
    ]
   },
   "versions": [
    {
     "date": "2026-05-25",
     "kind": "preprint",
     "where": "arXiv 2605.25517 v1",
     "url": "https://arxiv.org/abs/2605.25517"
    }
   ],
   "finding": [
    {
     "quote": "topical relevance and list position are the biggest drivers of being cited first",
     "from": "abstract, arXiv"
    },
    {
     "quote": "formatting-only edits have little impact",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "In a controlled testbed, what gets cited first is mostly relevance and position.",
   "does_not_show": "Production engines; extraction, confidence or inferred significance; scholarly genres.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30.",
    "unverified": [
     "the full list of 18 factors"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W23"
  },
  {
   "id": "W-0108",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-284.A1",
     "bears_on": "component",
     "scope": "The component 'Outputs converge on safe, flat, interchangeable prose': across 26K open-ended queries, single models repeat themselves and different models produce strikingly similar outputs. 'Meaning starvation' in users and market pressure against legibility are not tested."
    }
   ],
   "source": {
    "authors": [
     "Liwei Jiang",
     "Yuanjun Chai",
     "Margaret Li",
     "Mickel Liu",
     "Raymond Fok",
     "Nouha Dziri",
     "Yulia Tsvetkov",
     "Maarten Sap",
     "Alon Albalak",
     "Yejin Choi"
    ],
    "title": "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)",
    "venue": "arXiv; NeurIPS 2025 (oral)",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2510.22954v1",
     "https://neurips.cc/virtual/2025/oral/121422"
    ]
   },
   "versions": [
    {
     "date": "2025-10-27",
     "kind": "preprint",
     "where": "arXiv 2510.22954 v1",
     "url": "https://arxiv.org/abs/2510.22954v1"
    }
   ],
   "finding": [
    {
     "quote": "(1) intra-model repetition, where a single model consistently generates similar responses",
     "from": "abstract, arXiv"
    },
    {
     "quote": "(2) inter-model homogeneity, where different models produce strikingly similar outputs",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Infinity-Chat, a large-scale dataset of 26K diverse, real-world, open-ended user queries",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Open-ended LLM outputs converge within and across models.",
   "does_not_show": "Whether the convergence is 'safe' or 'flat' prose in the claim's sense; user response; the cascade of later stages.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page, 2026-10-30. NeurIPS 2025 venue from the neurips.cc search results; that page was not fetched.",
    "unverified": [
     "NeurIPS 2025 oral status"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W24"
  },
  {
   "id": "W-0109",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-960.F1",
     "bears_on": "method",
     "scope": "Behavioral fingerprinting of agent outputs by vocabulary distribution, the Phase 2 method, carried out by others: a classifier separates five model families at 97.1 percent from output text, rooted in word-level distributions, and survives rewriting and summarization. Drift across versions of one agent is not measured."
    }
   ],
   "source": {
    "authors": [
     "Mingjie Sun",
     "Yida Yin",
     "Zhiqiu Xu",
     "J. Zico Kolter",
     "Zhuang Liu"
    ],
    "title": "Idiosyncrasies in Large Language Models",
    "venue": "arXiv; ICML 2025",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2502.12150"
    ]
   },
   "versions": [
    {
     "date": "2025-02-17",
     "kind": "preprint",
     "where": "arXiv 2502.12150 v1",
     "url": "https://arxiv.org/abs/2502.12150v1"
    },
    {
     "date": "2025-06-16",
     "kind": "preprint",
     "where": "arXiv 2502.12150 v2",
     "url": "https://arxiv.org/abs/2502.12150"
    }
   ],
   "finding": [
    {
     "quote": "Notably, we achieve 97.1% accuracy on held-out validation data in the five-way classification problem",
     "from": "abstract, arXiv"
    },
    {
     "quote": "Our further investigation reveals that these idiosyncrasies are rooted in word-level distributions.",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Output text carries a stable vocabulary fingerprint of its source model.",
   "does_not_show": "A divergence score against archived fingerprints of the same agent; constraint adherence.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page with history, 2026-10-30. ICML 2025 from search results; that page was not fetched.",
    "unverified": [
     "ICML 2025 venue",
     "quote not found at source, dropped: \"These patterns persist even when the texts are rewritten, translated, or summarized by an external LLM.\""
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W25"
  },
  {
   "id": "W-0110",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-960.F1",
     "bears_on": "mechanism",
     "scope": "Behavioral drift exists and needs monitoring: the March and June 2023 versions of GPT-3.5 and GPT-4 differed greatly on fixed tasks. The witness compares task performance; output fingerprints are not its measure."
    }
   ],
   "source": {
    "authors": [
     "Lingjiao Chen",
     "Matei Zaharia",
     "James Zou"
    ],
    "title": "How is ChatGPT's behavior changing over time?",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://arxiv.org/abs/2307.09009v1"
    ]
   },
   "versions": [
    {
     "date": "2023-07-18",
     "kind": "preprint",
     "where": "arXiv 2307.09009 v1",
     "url": "https://arxiv.org/abs/2307.09009v1"
    }
   ],
   "finding": [
    {
     "quote": "the performance and behavior of both GPT-3.5 and GPT-4 can vary greatly over time",
     "from": "abstract, arXiv"
    }
   ],
   "shows": "Deployed LLM behavior drifts substantially between versions under the same name.",
   "does_not_show": "Fingerprint-based divergence detection; agent manifests.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the arXiv abstract page (v1), 2026-10-30.",
    "unverified": [
     "later versions and journal publication (reported in search results; unchecked)",
     "quote not found at source, dropped: \"highlighting the need for continuous monitoring of LLMs\""
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W26"
  },
  {
   "id": "W-0111",
   "direction": "supports",
   "evidence_class": "preprint",
   "claims": [
    {
     "claim": "PRED-1534.F1",
     "bears_on": "mechanism",
     "scope": "The rule that a lexical test returns false negatives on compressed or re-sorted material: the witness's premise and method are that semantic duplicates, similar without being identical, escape exact deduplication and are found by embeddings. Its corpus is web-scale image-text data."
    }
   ],
   "source": {
    "authors": [
     "Amro Abbas",
     "Kushal Tirumala",
     "Dániel Simig",
     "Surya Ganguli",
     "Ari S. Morcos"
    ],
    "title": "SemDeDup: Data-efficient learning at web-scale through semantic deduplication",
    "venue": "arXiv",
    "doi": null,
    "urls": [
     "https://huggingface.co/papers/2303.09540"
    ]
   },
   "versions": [
    {
     "date": "2023-03",
     "kind": "preprint",
     "where": "arXiv 2303.09540",
     "url": "https://huggingface.co/papers/2303.09540"
    }
   ],
   "finding": [
    {
     "quote": "data pairs which are semantically similar, but not exactly identical",
     "from": "abstract, Hugging Face papers page"
    },
    {
     "quote": "can remove 50% of the data with minimal performance loss",
     "from": "abstract, Hugging Face papers page"
    }
   ],
   "shows": "Embedding-based deduplication finds a large class of near-duplicates that exact matching misses.",
   "does_not_show": "Compression of a scholarly corpus; completeness of the archive's own blog index.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Hugging Face papers page (published Mar 16, 2023), 2026-10-30; the arXiv page itself was not fetched.",
    "unverified": [
     "the arXiv v1 day (the Hugging Face page gives Mar 16, 2023)"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W27"
  },
  {
   "id": "W-0112",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-204.A1",
     "bears_on": "mechanism",
     "scope": "Failure condition (a), that a machine-facing provenance block has no effect on summarizer output: the witness finds JSON-LD schema added to 1,885 pages moved AI citations little against matched controls, with a small decline in AI Overviews. Schema is a different block from the SPXI self-audit text and the measure is citation count; provenance retention is unmeasured."
    },
    {
     "claim": "PRED-204.F2",
     "bears_on": "mechanism",
     "scope": "The planned Layer 2 (JSON-LD via Schema.org) has, in this observational study, no detectable citation effect on Google AI Overviews, AI Mode or ChatGPT."
    }
   ],
   "source": {
    "authors": [
     "Louise Linehan",
     "Xibeijia Guan"
    ],
    "title": "We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.",
    "venue": "Ahrefs blog",
    "doi": null,
    "urls": [
     "https://ahrefs.com/blog/schema-ai-citations/"
    ]
   },
   "versions": [
    {
     "date": "2026-05-11",
     "kind": "report",
     "where": "ahrefs.com blog (updated 2026-06-09)",
     "url": "https://ahrefs.com/blog/schema-ai-citations/"
    }
   ],
   "finding": [
    {
     "quote": "Adding schema produced no major uplift in citations on any platform.",
     "from": "headline finding, Ahrefs blog"
    },
    {
     "quote": "Small but statistically significant decline relative to matched controls",
     "from": "AI Overviews result, Ahrefs blog"
    },
    {
     "quote": "Statistically indistinguishable from zero",
     "from": "AI Mode and ChatGPT results, Ahrefs blog"
    }
   ],
   "shows": "Matched difference-in-differences over 30 days shows no citation uplift from adding structured data.",
   "does_not_show": "Provenance fidelity in the composed answer; plaintext self-audit blocks; effects past 30 days.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Ahrefs post, 2026-10-30, after finding it through an authoritytech.io commentary page.",
    "unverified": [
     "the full method appendix"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W28"
  },
  {
   "id": "W-0113",
   "direction": "supports",
   "evidence_class": "news",
   "claims": [
    {
     "claim": "PRED-204.A1",
     "bears_on": "mechanism",
     "scope": "Condition (a) in another form of machine-facing file: across about 300,000 domains, presence of llms.txt showed no clear relation to AI citation frequency. Correlational, and the file is a site-level index unlike an in-document provenance block."
    }
   ],
   "source": {
    "authors": [
     "Matt G. Southern"
    ],
    "title": "LLMs.txt Shows No Clear Effect On AI Citations, Based On 300k Domains",
    "venue": "Search Engine Journal (reporting an SE Ranking study)",
    "doi": null,
    "urls": [
     "https://www.searchenginejournal.com/llms-txt-shows-no-clear-effect-on-ai-citations-based-on-300k-domains/561542/"
    ]
   },
   "versions": [
    {
     "date": "2025-11-20",
     "kind": "news",
     "where": "searchenginejournal.com",
     "url": "https://www.searchenginejournal.com/llms-txt-shows-no-clear-effect-on-ai-citations-based-on-300k-domains/561542/"
    }
   ],
   "finding": [
    {
     "quote": "doesn't seem to directly impact AI citation frequency. At least not yet.",
     "from": "SEJ article, quoting the study"
    },
    {
     "quote": "removing the llms.txt feature actually improved the model’s accuracy",
     "from": "SEJ article"
    }
   ],
   "shows": "A large correlational analysis finds no citation benefit from a machine-facing site file.",
   "does_not_show": "Causal effect; in-document blocks; provenance retention.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the SEJ article, 2026-10-30. The SE Ranking study itself was not fetched.",
    "unverified": [
     "the SE Ranking study's own page, date and the AI systems covered"
    ]
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W29"
  },
  {
   "id": "W-0114",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-722.B1",
     "bears_on": "component",
     "scope": "The component that open-weight models have made reasoning capability increasingly reproducible: frontier open-weight models trail the best models by about three months on Epoch's capability index. The thesis that the moat moves to the corpus is not tested."
    }
   ],
   "source": {
    "authors": [
     "Luke Emberson"
    ],
    "title": "Open-weight models lag state-of-the-art by around 3 months on average",
    "venue": "Epoch AI Data Insights",
    "doi": null,
    "urls": [
     "https://epoch.ai/data-insights/open-weights-vs-closed-weights-models"
    ]
   },
   "versions": [
    {
     "date": "2025-10-30",
     "kind": "report",
     "where": "epoch.ai data insight",
     "url": "https://epoch.ai/data-insights/open-weights-vs-closed-weights-models"
    }
   ],
   "finding": [
    {
     "quote": "Frontier open-weight models lag behind the most capable models by an average of 3 months",
     "from": "headline, Epoch AI"
    }
   ],
   "shows": "The capability gap between open-weight and closed frontier models is short.",
   "does_not_show": "Scarcity of substrate or corpus; indices as sovereign territory.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Epoch AI page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W30"
  },
  {
   "id": "W-0115",
   "direction": "supports",
   "evidence_class": "report",
   "claims": [
    {
     "claim": "PRED-722.B1",
     "bears_on": "component",
     "scope": "The component 'at dramatically lower cost': the price of reaching a fixed performance level fell 9x to 900x per year depending on the milestone. The corpus-moat thesis is not tested."
    }
   ],
   "source": {
    "authors": [
     "Ben Cottier",
     "Ben Snodin",
     "David Owen",
     "Tom Adamczewski"
    ],
    "title": "LLM inference prices have fallen rapidly but unequally across tasks",
    "venue": "Epoch AI Data Insights",
    "doi": null,
    "urls": [
     "https://epoch.ai/data-insights/llm-inference-price-trends"
    ]
   },
   "versions": [
    {
     "date": "2025-03-12",
     "kind": "report",
     "where": "epoch.ai data insight",
     "url": "https://epoch.ai/data-insights/llm-inference-price-trends"
    }
   ],
   "finding": [
    {
     "quote": "The rate of decline varies dramatically depending on the performance milestone, ranging from 9x to 900x per year.",
     "from": "headline, Epoch AI"
    }
   ],
   "shows": "The cost of a fixed level of LLM capability falls by one to nearly three orders of magnitude a year.",
   "does_not_show": "Where scarcity moves; corpus access.",
   "checked": {
    "by": "intake reader, theme E-execution; quotes checked at source by scripts/witnesses/verify_quotes.py",
    "date": "2026-10-30",
    "how": "Read the Epoch AI page, 2026-10-30.",
    "unverified": []
   },
   "found_by": [
    "S-0007"
   ],
   "reader_id": "E-execution-W31"
  }
 ]
}