The model's reward function is its own reward. No need to reinvent the wheel.