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

Modern large language models are increasingly deployed in high-stakes domains including medicine, law, finance, and government. In these contexts, the critical failure is not a lack of intelligence but a lack of epistemic legitimacy: the ability to determine whether a conclusion is permitted, blocked, or requires clarification given the available evidence structure. Current LLMs default to producing a best guess even when the structurally correct outcome is refusal or suspended judgment. This paper introduces the concept of epistemic legitimacy as a formal governance requirement for AI systems, presents Aurora-Lens, a provider-agnostic governance proxy that enforces epistemic admissibility at the transport layer between applications and LLM providers, and describes a tamper-evident forensic audit architecture suitable for regulatory inspection. Aurora-Lens does not replace large language models; it constrains them. The system maintains parallel admissible interpretations per session, collapses only under eliminating constraints, and treats refusal as a first-class correct outcome. We argue that epistemic governance at the tool boundary represents the highest-leverage intervention point for making LLM deployment defensible in regulated environments.