A polished answer can be the harder failure to review

A note on why answer quality reviews should track hidden assumptions, not only fluency.

For answer quality, fluency is sometimes the misleading part. A rough answer with visible uncertainty is easy to review. A polished answer with one hidden assumption can take longer, because the reviewer has to find the missing hinge. When comparing assistants, I would rather log the failure like this: - Did it invent a fact? - Did it miss a constraint? - Did it edit outside scope? - Did it sound more certain than the evidence allowed? - How long did review take? That makes model choice less like a popularity vote and more like a workflow decision.