It sounds like you are describing a strict cut-off point. But I imagine the reward score is a balance of both models and the balance switches from helpful to safe over a smooth curve, not abruptly. Am I right?