Comparing four common sovereign-AI deployment patterns across regulated operators, we find that only patterns where residency is enforced at the architecture layer — rather than the configuration layer — survive a full audit without exceptions. Configuration-level sovereignty produced a residency exception in every deployment studied.
- Only architecture-level residency enforcement passed full audit without exceptions.
- Every configuration-level deployment produced at least one residency exception.
- The durable pattern shaped the architecture inward from the boundary — sovereign by default.
Architecture vs configuration
The decisive variable was where residency is enforced. When sovereignty is a configuration on a globally-shaped architecture, every integration is a place it can leak — and in practice, one always did. When the residency boundary is the first architectural constraint, leakage has nowhere to occur.
Method
We classified deployments into four patterns by where the residency control sits, then assessed each against a full audit simulation looking for any path by which data, compute, or keys could cross the boundary. Exceptions were counted and traced to their architectural cause.
- Configuration-level residency on global architecture.
- Gateway-mediated residency with shared control plane.
- Regional isolation with cross-region dependencies.
- Architecture-level, sovereign-by-default — the only one that held.