From Requirements to Testable Boundaries: Measuring Change Containment Under Technology Substitution in Enterprise AI Platforms

Enterprise AI platforms depend on technologies that change far faster than the applications built on them, and capability-based design has been shown to improve tolerance to requirements change. Whether a specific capability boundary contains a heterogeneous technology substitution, measured at its interfaces, has not been tested in the sources examined here. This paper reports preliminary, bounded evidence on that question, treating containment as a prediction that can fail. Following a design science method, twelve requirements derived from practice and the literature are mapped to ten capability domains using three explicit criteria: independent substitutability, ownership and assessment. Decisions and four tested alternatives are recorded; one alternative the criteria favor is rejected on a judgment outside them, so the procedure is structured decision support rather than a deterministic derivation. Each boundary yields a prediction: replacing the technology that realizes a domain should change no interface outside it. The prediction was tested at three boundaries of an author-built instantiation, each under a failure criterion, interface inventory and behavioral assertions fixed before execution, with containment and behavior reported separately. The three outcomes differ. Replacing a hosted model service with a self-hosted one changed no interface outside Model Services, and a negative control confirmed that the inventory detects a planted cross-boundary dependency. Replacing the runtime’s in-process response cache with Redis exposed two latent defects, an unpackaged client and process-dependent cache keys; both were repaired inside the cache component, after which every behavioral assertion passed. Two retrieval engines, PostgreSQL with pgvector and Qdrant, were built against a pre-existing retrieval contract that had no implementation and then switched by configuration alone. No protected interface changed, but frozen behavioral assertions failed for both engines: the existing consumer returned raw engine errors to callers, and the instantiated contract carried tenant identity but not the access decisions the architecture assigns to that boundary. Interface containment held in all three tests; in the third it did not imply adequate behavior across the boundary. These are bounded results, not general validation: the instantiation, realizations and repairs are the author’s, and the results do not isolate the derivation procedure from the adapter patterns it relies on. Commercial platforms and external sources assess the decomposition descriptively.