Most classification metrics for oncology AI models (AUC, ICC, MAE) measure global agreement. They don't answer the question that actually matters at the point of care: how reliable is this model at the exact cutoff that decides whether a patient gets flagged, biopsied, or treated? I built oncothresh
