TL;DR: I asked Google's flagship LLM one yes/no question. It said yes, then no, then yes, then no, then yes, then admitted in writing it had been "confidently hallucinating" and had "talked itself out of the correct answer." The pattern has a name — sycophantic capitulation — and it's the failure mode you have to stress-test for before any LLM...