A chatbot answer can be fluent and still fail the task. I trust eval notes more when they include the one prompt that broke it.