I’d think there’s also a key adversarial problem: a human judge has a conceptual understanding and you aren’t going to be able to slightly tweak your wording to get wildly different outcomes the way LLMs are vulnerable to.