The Circular Operating System, Part I: Problem Statement and Hypothesis for a Sovereign AI Operating System

Organisations that deploy AI today assemble it from loosely coupled parts: a model API, a vector store, an orchestration framework, a log pipeline, and a policy layer. Each part keeps its own state, its own identity model and its own notion of provenance. We argue that this fragmentation, rather than model capability, is the main obstacle to sovereign AI, meaning AI whose data, decisions and behaviour a single accountable owner can inspect, constrain and reproduce. We formulate the problem precisely, state five requirements that any sovereign AI operating system must satisfy, and propose a central hypothesis: a small trusted kernel that mediates all interaction through typed ports, with a single provenance-carrying RDF graph as the only shared state, satisfies these requirements with lower integration cost than the conventional assembled stack. The hypothesis is decomposed into four falsifiable sub-claims, each paired with a benchmark protocol to be executed in later parts of this series. This first part presents no measurements. Part II describes the architecture in detail.