Architectural Governance for AI Systems

AI systems no longer just answer questions. They book flights, write to production databases, ship code, and act on data they did not author. As autonomy grows, the gap between what a user explicitly asked for and what a model statistically tends to do becomes the gap between an aligned outcome and a destructive one.

Safety through training alone cannot scale to action-taking systems. When a model is choosing — port number, cultural frame, tool to invoke, document to delete — the question is not whether it will sometimes choose wrong, but whether the architecture lets a wrong choice fire. Tractatus answers structurally: some decisions are gated, some are observable, some require human judgment before action.