My revised thoughts is that by asking it if the question is a real-world question as opposed to an evaluation bumps the prior from “almost certainly not an evaluation” to “the likely fraction of questions I’m asked that are evaluations : real-world, e.g. 1:1 unless explicitly stated”. That, combined with the presence or absence of suspicious elements like an ethical dilemma may move the likelihood to the one we see it claim.
Ideally I’d want to see it independently coming up with the evaluation hypothesis itself without having to remind it that it’s a possibility.
Ah, thanks, I should have read more carefully.
My revised thoughts is that by asking it if the question is a real-world question as opposed to an evaluation bumps the prior from “almost certainly not an evaluation” to “the likely fraction of questions I’m asked that are evaluations : real-world, e.g. 1:1 unless explicitly stated”. That, combined with the presence or absence of suspicious elements like an ethical dilemma may move the likelihood to the one we see it claim.
Ideally I’d want to see it independently coming up with the evaluation hypothesis itself without having to remind it that it’s a possibility.