Using verifier-generated explanations and rationales to improve verification, beyond just binary correct/incorrect signals.