What a governed system should do when it loses confidence — and why "halt everything" is the wrong reflex.
There is a comfortable story about safe AI that ends with a big red button. Something goes wrong, someone hits the button, the system stops. It is a satisfying image and a bad design. A hard halt is not safety. It is a second failure stacked on the first.
Consider what "stop everything" actually does to a live institution. Mid-flight authorizations die. In-progress work is abandoned in indeterminate states. The humans who could have recovered the situation arrive to find not a paused system but a wrecked one, with no lead time and no clean surface to act on. You traded a bounded problem for an unbounded cleanup. The button felt responsible. It was brittle.
The right question is not "how do we stop the system?" It is "how does the system degrade when it can no longer safely grow?"
Stasis: stop issuing, keep your word
Our answer is a state we call Stasis, and its rule is narrow on purpose: under Stasis, no new authority is issued — but existing authority continues to execute its scope.
Read that carefully, because the two halves are doing different jobs.
- No new authority is issued. The system stops expanding its footprint. It grants nothing new, escalates nothing, opens no new scope. The blast radius is frozen at exactly what was already legitimately in flight.
- Existing authority continues to execute its scope. Work that was already authorized — by a real grant, traceable to a real person — is allowed to finish. Commitments already made are kept. The system honors what it already promised.
Stasis is graceful degradation, not a hard stop. The system stops growing and keeps its word — which is exactly the posture that preserves the lead time a human needs to recover.
This is the difference between a circuit breaker and a demolition charge. A breaker isolates a fault and leaves the rest of the house livable. A demolition charge is indifferent to what was working.
Why keeping your word is the safer move
It is counterintuitive that finishing in-flight work is safer than stopping it. Here is why it holds.
A settlement half-executed is more dangerous than one allowed to complete. A migration frozen at step four of nine leaves the world in a state no one designed and no one can reason about. Abandoned work is not neutral — it is a new, unspecified failure mode, invented at the worst possible moment. By letting already-granted authority run its already-bounded course, Stasis keeps the system inside states that a human previously understood and signed off on. You degrade into the known, not out of it.
And because the footprint is frozen rather than detonated, the people responsible get the one thing a hard halt destroys: time. Time to look at a stable, recorded, still-coherent system. Time to decide whether to resume issuance, narrow it, or wind down deliberately. Stasis buys lead time for recovery. A halt spends it.
Degrade like an engineer, not like an alarm
Resilient systems do not treat "keep running" and "shut down" as the only two settings. They have a graceful middle: shed load, hold position, protect the core, and stay legible while doing it. Stasis is that middle for governed AI execution.
- It is observable — you can see it entered Stasis and why.
- It is bounded — the footprint can only shrink from here, never grow.
- It is recoverable — it hands the humans a coherent system and the time to act on it.
Anyone can build a system that stops. The engineering is in building one that degrades without breaking its commitments — that fails loud, fails bounded, and keeps its word on the way down. That is not a red button. It is a system you can still trust in the exact moment it has stopped trusting itself.
Reference invariants: ZTG-2 (Stasis), ZTG-1 (Mechanistic Boundary), ZTG-3 (Governed Effect Surface).