Hamel Husain shows how to use binary scores in LM-based evals even when outcomes feel subjective. Scope your evaluation narrowly to a specific failure like calling the wrong tool or leaking user IDs. Vague, broad evaluations produce noisy signals that don't validate well against human labels.
If you're using an LM judge to eval your AI application, the output of that judge should be a binary score. A lot of people get confused by this. They say, hey, the outcome that we're trying to grade is subjective. How can you possibly assign a binary score that's black and white? And the answer is you need to scope your eval small enough so that it is binary. So instead of saying, is something just good or bad, which is vague and not actionable, you need to narrow it down to the specific error, such as incorrect tool being called, or user IDs being emitted in the output, or some very specific failure. And the specific failure is something you find out from error analysis. It's really important that you scope your evals properly, otherwise they'll just cause a lot of noise. And if your evals are too vague, you won't be able to verify them easily against human labels. I have a blog post that I'll put in the comments where you can learn more.
