AI models are increasingly used to grade other AI models' work - and a new benchmark suggests those graders often can't explain themselves.
Researchers formalize what they call "judgment receipts": the minimal set of evidence, rules, or authority a verdict actually rests on. They built ReasonBench, a policy and logical reasoning benchmark spanning 19,520 cases and 7,200 controls, to test whether evaluator models could reproduce those receipts when the underlying facts changed. In controlled, frozen tests, a small model called Qwen3-1.7B scored 98.41% on receipt accuracy and 96.99% on predicting how judgments should shift - numbers that looked close to solved. Reshuffle the same evidence, rules, and authority into meaning-preserving orders, though, and receipt recovery collapsed to 54.8% and 49.2%.
That gap matters because these evaluators are gatekeepers now: they approve agent actions, route items for human review, and generate feedback used to train other models. A model that gets the label right without a stable reason for it is a liability wearing a passing grade. Retraining on simple reordered cases patched surface consistency to 96.6% but made deeper reasoning prediction worse, not better.
In other words, teaching a model to survive one flavor of scrutiny doesn't teach it to reason - it teaches it to survive that flavor of scrutiny. Anyone grading AI systems on accuracy alone is measuring the wrong thing.