Most ML models are tested the way students cram for an exam: On clean, familiar material, under comfortable conditions. Then they graduate into production — where the inputs are messy, the labels drift, and nobody is grading on a curve.