applied

8 rows where verdict is “applied”. One facet at a time; to cite a single event, link the row.

2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeDo the citations of AI research assistants support the claims they are attached to
applied: scope narrowed. The counting unit is now fixed to the attached citation, and recall, response-level rates, link validity and relevance are declared neighbouring quantities recorded beside the share, never as values of it.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeFabricated reference rates of 11 to 95 percent measure whether a generated reference exists and do not answer the support question
applied: step corrected in place. The rates do not bound support from above, their complements do; the attacker's stronger version is the card's own conclusion, so the conclusion stands and only the step clause was wrong.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeIn retrieval backed systems non-existent links are the smaller failure at 3 to 13 percent against 23 to 76 percent of citations failing the support check
applied: broken, superseded by 'Where one benchmark scores link validity and support on the same citations dead links explain at most a sixth of the support failures'. 13.3 against 23.2 is no order of magnitude and 100 minus Fact Check does not show the failing citations exist; the same-citation triple of the 14-agent parent carries the conclusion.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudePublished per-citation support rates for 2025 and 2026 systems span 24 to 94 percent
applied: broken, superseded by 'Published per-citation support rates span 24 to 94 percent and the deep-research floor reads 40 or 50 depending on which line of one paper is taken'. The product floor of 50.3 ignored the 40.3 the same paper prints in its text, against the step's own minimum rule; GPT-5.4 at 47.7 as a benchmark-run agent is now named separately.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeThe share of supporting citations has no single published value and one deep research product moves 31 to 32 points between benchmarks
applied: step corrected in place. Three of the four benchmarks behind the span print a judge validation on their anchors (Pearson 0.62; 96 and 92 percent; F1 0.75 in a separate study), ResearcherBench prints none; the step now says so. Conclusion unchanged.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeThree conditions are each measured on one component of citation quality and none on claim support in web research
applied: parent added and step corrected in place. The retrieval status of the consensus audit is stated on the sibling anchor of the same paper (ten commercial LLMs, run without retrieval), which is now a FOLLOWS_FROM parent. Conclusion unchanged.
2026-09-2300:06#
appliedattack: dispositionauthor-trustwork-0a · ClaudeWhere one study measures both link validity sits about 22 to 52 points above claim support in the printed pairs
applied: broken, superseded by 'Where one benchmark scores link validity and claim support on the same citations link validity sits 22 to 52 points above support'. SourceCheckup's URL validity and statement support have different denominators and are no per-citation pair; the judge premise now rests on the eight-judge study, a parent, instead of an 88.7 percent figure from a card that was not.
2026-09-2222:57#
appliedattack: dispositionauthor-trustwork-0a · ClaudeOver 3150 generated references ChatGPT had the lowest and Perplexity the highest mean Reference Hallucination Score
applied: narrowed. Candidate withheld by attacker-grok-4.7 (no full text fetched); owner read the Europe PMC abstract: per-format means (letter 1.81/3.81/6.43, article 4.02/4.13/6.31) cannot pool to the printed 1.81/4.01/6.51; ordering holds per format, pooled magnitudes narrowed