Use when a retrained ML model is a candidate for promotion and CI must block it if it got worse. Compares held-out accuracy, F1, AUC or RMSE with the production model within a tolerance, and checks key segments with Deepchecks so a global gain cannot hide a slice regression.