An AI eval note should mark which answer became worse after adding detail. More context helps only when it keeps the task stable.