Why Corrupted Training Data Doesn't Show Up as High Loss

A run on pure shuffled labels — a dataset with nothing left to learn — reduced its loss by 62% on a textbook-healthy curve. Noise is learnable, so it hides.