Safety
Autonomy that earns each key and hands it back on request.
A Compute World Model acts, not only predicts. It places work, and when nothing fits, it says what would. Both outputs pass through the same boundary. Safety is therefore a property of the architecture, not a policy applied to it afterward.
Authority is granted in stages
Observe. Recommend. Shadow. Canary. Bounded action. Broader authority.
Each rung is gated on measured reliability at the rung below. You set the rung. You can move it in either direction, at any moment, for any reason, and the system does not argue.
Specification sits at Recommend and never climbs. The model states what the next machine needs. A person decides whether to build it.
The boundary is independent by construction
Declared limits are enforced continuously from outside the learned model. The model does not hold its own leash. A violation escalates immediately, regardless of what the model predicted or how confident it was.
A boundary inside the model fails exactly when the model does. A boundary outside it does not.
Every kernel is verified before it executes
Generated code is checked for semantic equivalence against a reference before it touches hardware. The model proposes. An independent check decides. Confidence is not an input to that decision.
Reversibility is what lets this move fast
Every action is bounded and undoable. A slice of traffic is routed, measured against a preserved baseline, and kept or reversed. Nothing is committed that cannot be walked back.
One output is not reversible once acted on: a specification that becomes an order. That is why it is held at Recommend, why it is stated in units a buyer can verify after delivery, and why machines built to it enter service through the same ladder as any other hardware, from Observe up.
What the agent never touches
The model reads the shared power and thermal envelope and places work inside it. It sets limits on compute it has been delegated. It issues no facility command: no fan curve, no pump speed, no valve position, no supply temperature, no breaker limit. The room belongs to the people who run the room.
Neutrality is a safety property
A compute layer that depends on one supply chain concentrates the decision of who gets to build. We own no silicon. We hold no vendor design seat. We run on whatever hardware exists.
A specifier that held a design seat would be a party to the design it specifies. We hold the intelligence seat and nothing else, because the specification is only worth something if the people reading it know we had no stake in the answer.
What we are building toward
A compute layer understood well enough to predict, predicted well enough to act, and operated under a boundary that does not move when the model is confident.
Every machine ever manufactured, reachable by the work that needs it. Every machine not yet manufactured, specified by it. Both under limits the owner sets and the model cannot lower.