Improving model accuracy at inference by using extra computation or verification steps without retraining.