Sakshi-Protocol: State-Space Control and Distortion-Guided Grounding in Large Language Models

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 explicit representation
of internal reasoning dynamics, with generation and evaluation typically
occurring within a single probabilistic process. We introduce the
Sakshi-Protocol, a control-layer architecture that separates generation,
observation, and decision-making through an explicit cognitive state-space
representation. This state-space captures interpretable properties of internal
behavior, including stability, reactivity, transformation, valuation, and
integration.We define a distortion metric over this representation to estimate
epistemic instability and guide intervention decisions during inference. We
demonstrate empirically that internal signals are fundamentally insufficient
to detect high-confidence hallucinations, establishing a boundary condition
for this class of approaches. The framework responds by integrating
a distortion-guided external grounding mechanism, selectively invoked
when epistemic risk is elevated. This enables the system to regulate when
verification is required rather than attempting to directly classify correctness.
Evaluation demonstrates that distortion-guided control produces consistent
separation between control regions, while selective grounding reduces
hallucination rate and preserves baseline accuracy. These results motivate
state-space modeling and distortion-guided intervention as a principled
approach to improving reliability in LLM systems.