Your LLM judge gives a different answer on re-runs. How do you test with it?

Run the same LLM-as-judge eval twice and you can get pass, then fail, on identical input. Now try to...