A model review should keep one answer that failed quietly. Quality checks work better when the subtle miss stays beside the score.