Modules C03, C06 · concept player

Healthy and quietly wrong

A model can score perfectly on every system metric while the world underneath it moves. Shift a live feature's distribution against its training baseline and watch a divergence score climb — then flip whether the label relationship moved too.

Illustrative Gaussian bins, not a live production metric. The divergence score is a standard Population Stability Index computed over ten bins; the 0.1 / 0.25 thresholds shown are a common industry rule of thumb for "watch this" and "act on this", not a number this curriculum measured. The distinction is the citable part — the glossary's drift entry: covariate shift moves P(x) while the model's mapping stays valid; label/concept shift moves the relationship itself, which retraining on old labels cannot fix.