Trains AI reviewers to identify substantive, evidence-based concerns about papers. Its reinforcement-learning objective balances the precision of critiques with their coverage of known high-quality concerns.