A learned component trained to decide whether a language model's generated answer should be kept or rejected.