How do you judge "better" if you don't understand the LLM output?