Cross-Border AI Architecture Patterns: Three Patterns, and Why the Fourth Is Not a Choice

An institution told that data may not leave a jurisdiction usually treats the statement as an obstacle to a deployment. It is not an obstacle. It is the first fact about the deployment's architecture, and treating it as an obstacle is how an institution arrives at a topology chosen for cost and then argues that the clauses permit it. This specification inverts that order. It is the worked instance of the Boundary Invariant applied to the data residency boundary class: the clause is the boundary, the topology is the optimization, and the patterns are optimizations under a clause that may not be crossed. Three patterns are enumerated at the level of a single workload class. Sovereign Silo, where no data may leave. Federated, where named derived artifacts may move and the underlying data may not. Regional Hub, where transfer is permitted to a destination every participating jurisdiction accepts. Hybrid is not a fourth pattern, which is what the subtitle means. It is the composition that appears when workload classes in one estate resolve differently, reported across an estate rather than selected for a class. An architect who chooses Hybrid has skipped the step that produces it. The specification takes no residency rule as its own. Its single input is the residency rule set produced by the AI Data Governance Framework, taken at a stated version and consumed through a schema published there rather than restated here. Two copies would drift. A selection references clause identifiers and must not introduce a clause, which prevents an architect supplying, at the moment of selection, the permission the selection needs. Where a clause is unresolved or past its re-verification date the instrument returns Blocked rather than a pattern. It is entry REG-09 of the Defensible AI Framework Registry, and the only subordinate entry occupying no MESA altitude, sitting instead under the Boundary Invariant applied to residency. Section 4 publishes the decision instrument the registry recorded as this entry's missing artifact, the absence of which meant the method could not be applied consistently by two architects working independently. What has not been tested is stated rather than implied. No institution unconnected to the author has been observed operating the selection, and Sections 2 through 6 are architectural. The determinism claim is untested: the specification claims two architects given one rule set produce one assignment, and that has not been tested with two architects — a procedure can be deterministic on paper and still admit judgement where a workload is assigned to a class. The registry's relationship register records no authority grant to this framework, while the registry's own account holds that evidence cannot be produced without authority and this framework produces evidence; the specification names that as a gap in the corpus rather than resolving it. And the worked cases in the source treatment are disclosed there as composites: they illustrate a method and are not evidence that any named institution operates any pattern. This specification names no jurisdiction and states no transfer rule. Jurisdiction-specific rules age, and a specification carrying them would age with them while continuing to read as authoritative. That is why no statute, regulation or supervisory instrument appears in its reference list. It is not a certification scheme.