Where judge decisions on citation support were compared with human labels, two measurements find the LLM judge stricter than humans (false negative rates of 0.183 to 0.470 on a human-reviewed benchmark; GPT-4o judge precision of 79.7, 78.1 and 53.9 against human 88.9, 84.2 and 67.5 on the same responses) and one finds it more generous (both disagreements in 100 validated AI Overview verdicts were the verifier accepting a claim the humans did not).

The judge benchmark contributes the size of over-rejection on an adversarial report in which 81.6 percent of pairs are unsupported by construction. The LongCite human evaluation contributes the same direction on natural model output, for a document-grounded task. The AI Overviews validation contributes the opposite direction on natural web output, on two cases. The step does not average them: tasks, judges and base rates differ, and two cases are not a rate. It follows only that the LLM-judged support rates in this stock cannot be corrected in a known direction.

A validation of one of the stock's judges on natural web-research output with a few hundred human labels that shows a consistent direction of error. Query: human re-labelling of released judge decisions from DeepResearch Bench or arXiv 2605.06635.

LLM judges are lenient, so real support is lower than reported. Not licensed: two of three comparisons in the stock show the judge stricter than the humans.