Decision Models That Return Probabilities, Not Text

Every launch post for decision support models opens the same way — new class of model, returns probabilities instead of text, benchmark table sorted by accuracy. But accuracy isn't the number your routing system consumes. This episode digs into what these models actually are, why binary yes/no decisions were awkward with ordinary LLMs and JSON constraints, and how they slot into moderation, automations, and agentic systems. Liquid AI's d1 gets a close look, along with the calibration problem that structured outputs never solved. Episode #864427 — open it directly at myweirdprompts.com/864427