Autonomous systems increasingly operate under conditions where their internal representations outlive the assumptions under which those representations were formed. While contemporary approaches emphasize optimization, learning, and self-modification, they largely assume a fixed representational frame. When this assumption fails, systems may exhibit brittleness, uncontrolled self-revision, or unsafe behaviour. This paper introduces epistarchia as an architectural principle governing epistemic authority in autonomous systems. Epistarchia enforces a disciplined alternation between committed operation, in which representations are treated as authoritative and guide action, and bounded suspension, in which epistemic commitment is temporarily withheld to enable safe revision of representational frames. We formalize the core invariants required to realize epistarchia, propose a reference architecture that enforces them, and analyze the safety properties such architectures provide. The contribution is not an algorithm or implementation, but a foundational structural layer applicable across substrates. We argue that without epistarchic constraints, autonomous systems face fundamental limits in robustness, accountability, and long-horizon operation.
Riaan de Beer (Thu,) studied this question.