A deterministic reward function based on verifiable rules rather than subjective judgment, enabling automated training without human annotation.