reward = reward_correctness * clip(α * (target_length - actual_length) + δ, 0, 1)