We allude to 2 when talking about using explanations first, but I totally agree. One minor comment is explanations after can sometimes be useful for understandi

We allude to 2 when talking about using explanations first, but I totally agree. One minor comment is explanations after can sometimes be useful for understanding how the model came to a particular generation during post-hoc evals. Point 1 is also a good callout. I added something on this for llm judge but it’s relevant more broadly.