A lightweight component added after a model is trained to estimate confidence without retraining the base model.