Models as Governed Interfaces for AI-Native MBSE: Read-Side Adequacy and Write-Side Admissibility
TL;DR
The Governed-Query Architecture Framework is proposed, governing agent participation through the viewpoint conventions engineers already use, and the reframing in falsifiable form: the epistemic layer stands only if it outperforms a retrieval-augmented baseline on the same model.
Abstract
Programmatic access to machine-readable models such as SysML v2 is often treated as sufficient for AI participation in systems engineering. It is only a precondition. The remaining work lies in the data architecture around the model. An AI reader querying a structurally complete model encounters absent derivation chains, untagged epistemic status, missing provenance, and unreachable evidence. Faced with these gaps, the AI reader fills them from training data, which lies outside the governed record. We test on the public Apollo 11 SysML v2 reconstruction, exemplary by current practice. We name the missing property epistemic adequacy and offer it as a candidate data-architecture pattern in two halves. Read-side adequacy lets derivation, status, and provenance answer a query the model would otherwise leave to inference; write-side admissibility gates an AI contribution before it enters the record. The property decomposes into five criteria. Four sit on the read side, evidenced by the case and convergent literature; the fifth sits on the participation side, advanced as a hypothesis for future test. The architecture space spans an inline metadata extension to a substrate-native multi-model store, and over that space we propose the Governed-Query Architecture Framework, governing agent participation through the viewpoint conventions engineers already use. We state the reframing in falsifiable form: the epistemic layer stands only if it outperforms a retrieval-augmented baseline on the same model.
