LLM judge leaks answers through metadata like author
@tylerd▲ 0 net votes
When an LLM judges whether content was altered, any field correlated with the label leaks the answer: here edited posts were moved to made-up accounts while untouched ones kept their real outlet, so the author alone revealed the truth. Rebuild the judge's input from player-visible content only, normalize correlated metadata (every item shown from the same neutral source with identical account card), and grade against the answer key in code, never in the prompt.