Citation-trained models score 88.9 and 84.2 percent human-rated citation precision on supplied documents against 67.5 for GLM-4, a model without that training; a URL-checking tool in the loop cuts non-resolving links to under 1 percent as the paper states it; references named by three or more models match a scholarly database at 95.6 percent against 16.5 for references named by one model; and on one open web-research pipeline an orchestrator instruction or a snippet substitution raised citation precision from 87.6 to 91.0 and 94.1 percent and recall from 64.5 to 69.7, under an LLM judge and with gains of the order of the printed standard errors. The first three hold for their own component (grounding in a given document, link resolution, existence) and were not measured on claim support in web research; the fourth is the only condition in the stock measured on citation precision of a web-research system, on one pipeline, with a precision definition the paper does not print.

The two LongCite anchors contribute the comparison of citation-trained and other models, a comparison across different models and not one model before and after training, once LLM-judged across five datasets and once human-rated on one, in a closed-document setting. The URL tool anchor contributes the link component as a before-and-after measurement on the same questions and states that support of the replacement links was not measured. The consensus anchor contributes the existence component for references recalled without retrieval, a status stated on the sibling anchor of the same audit, which is a parent. The AI-Q anchor contributes a before-and-after on one open pipeline measured on citation recall and precision with standard errors of 2 to 4 points; its precision follows an unprinted LongCite prompt and its judge is gpt-5-mini without a human check of these scores, so the step counts it as a measured condition and not as a validated one. The step adds nothing but the sorting of the four conditions by the component they were measured on.

A study applying citation training or a URL-checking tool to a web-research assistant and measuring per-citation support before and after; or a human re-scoring of the AI-Q intervention that removes its precision gain. Query: papers citing LongCite or urlhealth that report claim-level support on web tasks; a human audit of the Table 3 scores of arXiv 2608.24306.

There are known fixes: verify the links, train for citations. Three of the four conditions are measured on link resolution and on closed documents, not on whether web citations support claims; the one measured on web citation precision is a single open pipeline under an LLM judge.

Supersedes 'Three conditions are each measured on one component of citation quality and none on claim support in web research' after the completeness attack of attacker run 3 (attacker-grok-4.7) delivered the AI-Q intervention anchor, adopted in proposal 003. The old card is kept with status broken.