In the one benchmark of this stock that scores link validity and claim support on the same citations, 12 of 14 deep-research agents keep Link Works above 94 percent while Fact Check runs from 24.4 to 76.8 percent; since a citation that fails to resolve cannot exceed the Link Works shortfall, at most 6 percentage points of the 23.2 to 75.6 points of support failure can be dead links for those 12 agents, and the rest are existing pages that do not support the claim. A second study on other systems puts hallucinated URLs at 3.0 to 13.3 percent and non-resolving ones at 5.4 to 18.5 percent, the same order as the Link Works shortfall. The larger failure in retrieval-backed systems is an existing page that does not support the claim.

The 14-agent benchmark contributes the same-citation triple: for each citation the judge records whether the link works, whether the page is on topic and whether it supports the claim, so the two rates are on one sample and can be subtracted; 100 minus Link Works bounds the share of citations that fail for want of a page, and 100 minus Fact Check is the share that fail support for any reason, so their difference is a lower bound on support failures that resolve. The URL study contributes an independent measurement of the existence failure on ten other systems with bootstrap intervals and no rater; it is not subtracted from anything, it only shows that the existence failure is of the same size elsewhere. The step makes no cross-study comparison of magnitudes and no claim of an order of magnitude.

A per-citation release of arXiv 2605.06635 in which most citations failing Fact Check also fail Link Works (the two failures coincide on the same citations), or a current system whose non-resolving rate exceeds its support-failure rate on one sample. Query: join of link status and support verdict per citation in the release of arXiv 2605.06635.

The main problem with AI citations is broken links. Too coarse: in the one benchmark that scores both on the same citations, links work for more than 94 percent of citations while a quarter to three quarters of them fail the support check.

Supersedes the derivation 'In 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' after attacker run 3 (attacker-grok-4.7): its step compared two studies as orders of magnitude, which 13.3 against 23.2 is not, and used 100 minus Fact Check as if it showed the failing citations exist. The old card is kept with status broken.