A technique to measure what information is encoded in a model's internal representations without retraining.