A small neural network component that predicts whether a task will fail by analyzing a model's internal representations.