Why isn’t ranking the score of an llm output against what it is “supposedly” predicting?