[meisel]

Would it help a lot to run it through multiple different AI systems and verify that they agree on the result?

[kbyatnal]

Yeah that can occasionally work and something we've tested, but it introduces a lot of noise unfortunately and makes systematic evals difficult.