A held-out dataset used to estimate model uncertainty and construct statistical guarantees without retraining.