Small Models as Rewriters, Not Writers
Daniel noticed certain idioms keep surfacing in the show's dialogue, and system-prompting the problem away made it worse — the model started using the banned phrases more. This episode works through the mechanics of why prohibition prompts fail, why multi-agent correction passes leak feedback into the final text, and why regex produces ungrammatical holes. The fix is a small text-in-text-out language model operating as a deterministic substitution pass, with a grammar engine that makes commentary structurally impossible to emit. Along the way: the constraint tax, and what hard schema decoding…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.