Label 50 to 100 outputs by hand, compute Cohen's kappa against your LLM judge, and test position and verbosity bias before ...