ResearchPosition
AI governance
AI alignment at the point of deployment.
Alignment is usually discussed as a property of frontier models. For the systems that run infrastructure, it is a property of the deployment, and it is engineered or it is absent.
Whose intent, exactly
Alignment is the property that a system pursues the intent of the people it serves, rather than a proxy that drifts from it. In the research literature this is a question about training objectives. At the point of deployment it becomes concrete and unglamorous: whose intent, written down where, checked by whom. A model that is aligned in the abstract can still be misaligned in an operations room, because the operator's actual mandate, serve these customers, protect this equipment, never invent a number, was never made explicit anywhere the system could be held to it.
Our position is that deployed alignment begins with writing the mandate down as policy the system enforces, not prose the vendor gestures at. What the system may decide alone, what it may only recommend, what it must never do: these are specifications, and a deployment that cannot produce them on request is not aligned, whatever its model card says.
The envelope is the alignment mechanism
We do not rely on a model's judgement to keep it inside its mandate. Every model output in our systems runs inside a deterministic envelope: a bounded set of actions it can influence, hard limits it cannot cross, and a defined failure mode when its inputs degrade. The envelope is ordinary software, testable and auditable, which is precisely its virtue. Alignment enforced by architecture survives model updates, prompt drift and vendor changes. Alignment enforced by behaviour has to be re-established every time anything upstream moves.
The same applies to failure. An aligned system fails in the direction its operators chose in advance. Ours degrade toward inaction and escalation: a forecast that loses its inputs falls back to the honest baseline and says so; an automation that meets an unexpected state stops and routes to a human. Choosing the failure direction in advance is alignment work, done where it is cheap.
Alignment enforced by architecture survives model updates, prompt drift and vendor changes.
Figure 01
Alignment as architecture. The mandate enters as enforceable policy, the model runs inside a deterministic envelope of ordinary software, and only actions inside the envelope reach the world, with operator override on the way out. The hard limit holds regardless of how confident the model is, and degraded inputs fail toward inaction and escalation.
An aligned system fails in the direction its operators chose in advance.
Alignment is demonstrated, not asserted
A claim of alignment that cannot be checked is marketing. Because every consequential action in our systems leaves an append-only record, the question "did the system stay inside its mandate" is answerable from evidence, month after month, by someone other than us. That is the standard we think deployment alignment owes its users: not a promise about values, but a mandate written as policy, an envelope that enforces it, and a record that shows it held.
