A good AI reviewer knows when to stop

Every evaluative system eventually confuses volume with quality. A critic who flags more issues is assumed to be sharper; a screening system that pushes more objections is assumed to be more thorough. The newest AI peer-review tooling is now admitting, in numbers, that those assumptions were wrong: the right model is not the one that complains most, but the one calibrated enough to know when the paper in front of it is fine.

EquiReview-R, a preprint on arXiv as 2609.03943, treats two failure modes as separate things rather than one tradeoff. Overcritique is the model inventing a flaw the text does not support. Omission is the model missing a flaw the text clearly contains. The first is loud and easy to spot; the second is quiet and shows up downstream. The fix for one is not the fix for the other: a refinement pass over an existing concern set, scored against localized evidence, reduces the first without inflating the second. On a frozen cohort of unseen papers, fabricated major criticisms drop from 15.5% to 8.1%, the omission upper bound holds at 9.9%, and the system returns a "stop" verdict on 52.4% of papers.

The mechanism is portable. A trustworthy AI critic, in any domain, is one whose objections cite evidence that actually supports them, and one calibrated enough to stay quiet when no objection does. Tools that never say "this is fine" are not more rigorous; they are louder.