LLMs can’t really evaluate things. They’re far too suggestible and can always be broken with the right prompt no matter how many layers you apply.