A reward derived indirectly from model behavior (e.g., policy shift magnitude) rather than explicitly computed by a reward model.