Epistemic Legitimacy as a Governance Layer for Large Language Models: Architecture and Implementation

Large language models are increasingly used in settings where an answer can become advice, a record entry, a recommendation, or a decision trigger. In such settings, the central problem is not fluency alone, but whether a model-generated conclusion is permitted by the available evidence, persistent state, and role-specific authority. This paper introduces epistemic legitimacy as a governance requirement for high-stakes LLM deployment. It presents an architecture in which interpretation is conditioned by persistent existence rather than reconstructed from local sequence alone, admissibility is determined before output becomes consequence, refusal and suspended judgment are valid governed outcomes, and post-refusal continuation is itself lawfully constrained. The paper distinguishes Aurora as the broader reasoning substrate, the Persistent Existence Framework (PEF) as the persistence substrate on which meaning depends, Lens as the runtime admissibility gate, and Governor as the continuation layer that constrains what may happen after refusal, revision, or ambiguity. It also describes a provider-agnostic transport-layer implementation and a tamper-evident forensic audit record generated at decision time.The author used AI writing assistance for editorial refinement of this manuscript.