Yes, agreed. I was speaking in general, of course. This particular topic is of interest to me, so thinking of the edge cases and confounds vs Jev.

In your example, I would expect an LLM to do fine and if you have access to the raw logits you can measure whether or not it was confused and assign a confidence to the answer it gave.

I do think that Jev handles more than this though and, in my early testing, does things that are not easily accomplished with guided decoding techniques.