Do you think constitutional approaches would help here? (Verifiable reward for the main score, but then asking the model to self-critique for security and quality.)