AI Governance Already Has a Track Record. We Just Haven't Read It. — The same governance architecture we are now trying to build in software has been running, imperfectly, in constitutional history for centuries.

The Governance Twin (GT) is a runtime architecture for AI oversight in which an operational mind — the agentic AI doing the work — runs alongside a structurally independent governance mind that observes, deliberates, and feeds back guidance without controlling execution.The same architecture exists in two domains: the AI-GT, an engineered runtime system, and the polity-GT, the same five functions read as constitutional structure. The polity-GT has been running in real life for centuries; the historical record supplies empirical evidence for two structural variables that decide whether any GT works — embedding depth (how many independent civic substrates carry each function) and coevolution rate (whether the governance mind can adapt at the tempo of the operational mind). This article reads three completed failures (Qing, Romanov, Weimar), two survivals through significant Ethics-Morals-Values (EMV) transitions (Hungary, Singapore), and three live cases (the United States, Turkey, Russia) through the framework. Success and failure here are properties of the system, not of the EMV the system happens to stabilise on. Because agentic AI operates at machine speed, the failure modes that polity-GTs took generations to exhibit will play outfar faster; the historical record is therefore the primary evidence base for what the AI-GT specification must engineer against. This article is an appetizer; the full academic paper develops the methodology, the formal definitions, and the engagement with objections.