Two models train on the same labeled data with different hyperparameters. The one with higher hold-out accuracy gets promoted to production. Useful for tuning learning rate, feature subsets, or testing new training strategies without breaking the live model.