Reality Image Model: A Unified Risk Modeling and Metrological Governance Framework for LLM Factual Hallucinations (RIM-H)

Reality Image Model: A Unified Risk Modeling and Metrological Governance Framework for LLM Factual Hallucinations (RIM-H) Large Language Models (LLMs) have demonstrated remarkable capabilities in complex knowledge reasoning and multi-task generation. However, "factual hallucination"—where models generate fluent yet factually unsupported or logically contradictory statements—remains the fundamental bottleneck impeding their deployment toward high reliability, safety alignment, and rigorous scientific applications. Existing governance paradigms (such as post-hoc fact-checking, Retrieval-Augmented Generation (RAG), or probabilistic calibration) largely remain fragmented, patch-oriented engineering heuristics, lacking a unified risk modeling framework grounded in information source epistemology and a rigorously falsifiable metrological system. Grounded in the forefront of independent research and building upon the author's foundational work in cognitive science and systems theory—Reality Image Theory (RIT)—this project systematically constructs the Reality Image Model for Hallucination Governance (RIM-H). Moving beyond conventional purely probabilistic perspectives, RIM-H reconceptualizes LLM factual errors as structural system risks within the full-chain evolution of "External Reality -> Accessible Information -> Reality Image -> Information Mediator -> Evidence Constraint." To address the critical metrological blind spot of "Atomic Claim Drift" and "Statistical Unit Drift" inherent in prior evaluation frameworks, this project has fully developed and statically frozen the Atomic Claim Parsing Protocol (ACP-v1.0.1). This establishes a fully auditable evidence chain spanning raw model responses to a four-layer isolated data architecture (Response -> Span -> Claim Instance -> Semantic Claim) paired with deterministic Log Loss recalculation. Achieving methodological saturation in theory and a rigorously isolated double-blind pilot execution framework in engineering, this project delivers a reproducible, scientifically rigorous governance paradigm for LLM factual hallucinations suitable for rigorous international peer review.