From ConOps to Running Code: An MBSE-in-Practice Experiment using AI-Augmented SysML v2 Modeling

Model-Based Systems Engineering promises traceability from requirements through architecture to implementation, and that well-defined interfaces make integrating components across teams easier. In practice both promises are hard to keep. The link between architecture models and running code is usually created by hand and breaks as soon as either side changes. Interfaces described separately in each system's documentation drift apart, and the integration that MBSE was supposed to simplify becomes a manual reconciliation exercise. We built an experimentation platform to test whether native AI integration into a SysML v2 toolchain, grounded by continuous validation and a curated knowledge base, can close those gaps, and exercised it in a controlled development and integration experiment.

The generated code interoperated correctly, and its automated tests passed: component tests and end-to-end tests, all traced to requirements and verification cases in the model. Equally valuable were the honest lessons. A validator confirms a model is well-formed, not that it is complete; several defects sat inside a model that passed every check, from a behavior the code only half-implemented to values the language model invented where a requirement left a number unstated. Code generation made each gap visible and traceable to the requirement that caused it.

The presentation introduces the platform, walks through the experiment, shares what we learned about where model-driven generation helps and where it does not, and lays out our path from experiment toward adoption guidance for programs that must qualify software.