Across the five LLM-judged measurements of 2025 and 2026 in this stock that score whether the cited page supports the attached claim, the printed per-system rates run from 24.4 percent (OSS-120B run as a deep-research agent) to 94.04 percent (Claude-3.5-Sonnet with search on DeepResearch Bench). For commercial deep-research products alone the printed range is 50.3 to 90.24 percent by the tables, but the lowest value belongs to a system for which the same paper's running text prints 40.3 percent, so the floor reads 40.3 or 50.3 depending on which line of arXiv 2509.04499 is taken; and a commercial model run as a deep-research agent by benchmark authors rather than sold as a product (GPT-5.4 at 47.7 percent Fact Check) sits below the product floor either way. Two values adopted from proposal 003 fall inside the span and below the product floor: OpenAI DeepResearch at 39.9 percent citation precision on DeepScholar-Bench (GPT-4o entailment judge, arXiv-only corpus, 80 percent human agreement on the judge) and OpenAI Deep Research at 78.87 percent cited-statement consistency on ReportBench (gpt-4o judge, no validation printed); with the first, the printed floor for a commercial deep-research product is 39.9 percent.

Each parent contributes one benchmark's range under its own task set and judge: 24.4 to 76.8 (14 agents, 130 queries, with GPT-5.4 at 47.7), 39.8 to 68.3 (four generative search engines, DeepTRACE), 50.3 by Table 1 or 40.3 by the text to 79.1 (deep-research configurations, DeepTRACE), 39.36 to 94.04 with 77.96 to 90.24 for the four deep-research agents (DeepResearch Bench), 0.62 to 0.86 (ResearcherBench faithfulness), .399 for one product (DeepScholar-Bench, per-citation entailment) and 78.87 for one product (ReportBench, per-cited-statement consistency). The step takes the minimum and maximum of every printed value, text and table alike, and states the product range with both readings where a parent prints two values for one system; it restricts the product range to systems the parents label as sold deep-research products and names separately the benchmark-run configuration that falls below it. It widens no parent's scope: all five are per-citation or per-cited-claim support under an LLM judge; none is human-rated on every citation.

A parent turns out to measure a different quantity (for instance support by any listed source instead of by the attached citation), which would remove its values from the range; a human-rated replication of any one benchmark lands outside its printed range; or the authors of arXiv 2509.04499 resolve the 40.3 against 50.3 print in a revision. Query: compare the metric definitions of the five papers clause by clause; check arXiv 2509.04499 for a v2.

There is one number for how often AI citations hold. Too coarse: the printed per-citation support rates of 2025 and 2026 run from a quarter to more than nine tenths, and even the floor for one product depends on which line of one paper is read.

Supersedes the derivation 'Published per-citation support rates for 2025 and 2026 systems span 24 to 94 percent' after attacker run 3 (attacker-grok-4.7): its product floor of 50.3 took Table 1 of arXiv 2509.04499 and ignored the 40.3 the same paper prints in its text for the same system, against its own rule of taking the minimum of printed values; and GPT-5.4 run as a deep-research agent at 47.7 percent was neither included nor named. The old card is kept with status broken. Extended 2026-09-23 after proposal 003 (completeness attack of attacker-grok-4.7): two adopted anchors added as parents; the product floor moves from 40.3 or 50.3 to 39.9.