Minimum performance thresholds (accuracy, precision, recall, AUC above baseline) - Performance on critical subgroups (no dramatic degradation for any segment) - Inference speed within latency requirements - Model size within deployment constraints