For a long time, medical AI evaluation was dominated by knowledge benchmarks: medical licensing exams, multiple-choice question answering…