New Framework Audits AI Evaluator Reasoning, Not Just Labels

A new framework has been developed to enhance the accountability of AI evaluators by focusing on their reasoning processes, not just the correctness of their final labels. Traditional evaluation often overlooks instances where evaluators arrive at correct answers through flawed logic, which is critical for agentic systems that gate actions or provide training feedback. This framework formalizes reasoning accountability through three core sources: grounds, norms, and authority. By varying these sources, an "eight-cell counterfactual judgment cube" is created to characterize how judgments change. "Judgment receipts" are defined as minimal source replacements that reproduce revised verdicts, effectively explaining judgment transitions. A new benchmark, ReasonBench, was introduced with 19,520 cases to test this. Evaluations on ReasonBench revealed that while models like Qwen3-1.7B achieved high receipt accuracy, strong standard accuracy often masked severe robustness flaws. Meaning-preserving source permutations drastically reduced valid receipt recovery, indicating that models struggle with consistent reasoning even when the final verdict remains correct. This suggests that robust reasoning requires decoupling prediction from certification and evaluating transformation consistency alongside standard accuracy.