Modern large language models (LLMs) exhibit strong generative capabilities,but remain prone to producing fluent yet factually incorrect outputs.A key limitation of existing approaches is the absence of an explicitrepresentation of internal reasoning dynamics, with generation andevaluation typically occurring within a single probabilistic process. Weintroduce the Sakshi-Protocol, an epistemic control architecture thatseparates generation, observation, and decision-making through an explicitcognitive state-space representation that captures stability, reactivity,transformation, valuation, and integration. A distortion metric over thisrepresentation estimates epistemic instability and guides interventiondecisions during inference. The architecture is conceptually motivatedby classical frameworks for valid knowledge, formalising six epistemicinstruments as distinct system components. We extend the framework witha three-class hallucination taxonomy and a grounding layer that selectivelyinvokes external retrieval when epistemic risk is elevated. Evaluated acrossthree model families on a structured hallucination evaluation datasetspanning fabricated entities, real-entity specific claims, and niche traditionprompts, the protocol achieves strong and consistent intervention onhigh-risk prompts while preserving baseline accuracy on non-hallucinationprompts. These results establish the Sakshi-Protocol as a structured epistemiccontrol layer for reliable LLM deployment.
N.K. Vidyesh (Wed,) studied this question.