Why can’t we just use USMLE or bar exam scores to evalu...

There is a critical distinction between quizzing and asking. Licensing exams like the USMLE or the bar exam test whether a model has memorized structured knowledge under exam conditions. They do not measure whether the model can execute the actual, complex workflows required in a professional setting. Consequently, a model might pass a standardized test but still fail to produce usable professional artifacts, such as a discharge summary.