CliniCARE-Bench

What CliniCARE-Bench measures

Foundation models now match or exceed clinicians on medical-knowledge exams, yet exam performance is a poor proxy for what clinical deployment demands. A nurse at change of shift, a quality officer auditing sepsis-bundle compliance, or a nephrologist assessing a returning patient does not answer a self-contained multiple-choice question, each conducts a small research project against a heterogeneous, longitudinal, uncurated electronic health record.

CliniCARE-Bench (Clinical Calibrated Audit of Medical Reasoning in EHR) is a deployment-oriented benchmark for selective autonomy in retrospective clinical audit. To our knowledge it is the first clinical agent environment to support end-to-end auditing of an agent's reasoning process, retrieval, evidence grounding, policy use, and the decision trace that links them, scored end to end rather than on the final answer alone.