A regulated institution is about to let AI agents act on its behalf: settle a claim, release a payment, merge a change into production. The models are good enough for the act. What they act on is another matter. The rule the agent applied, the interpretation it relied on, the pattern it took to be current: each is a claim somebody once made, and nobody is keeping.
Institutions already have controls for their data and systems. Our work examines how those controls need to extend to the interpretations and assumptions an AI system relies on when it acts. So their quality decides whether the act can be defended afterwards.
Governed intelligence is the name for the layer that holds them, and for the body of work that specifies it. A reference architecture in which every piece of institutional knowledge is a claim with provenance, standing, and a shelf life. A check before each act that compares the standing of what the agent relied on with the consequence of what it is about to do. A record that lets a supervisor replay the basis of any decision. And above the architecture, a control plane that says who may let an agent act, on whose authority, and within what limit.
