RIM-H M6.1 A Two-Layer Model of Hallucination (Project Description)

RIM-H M6.1 is a theoretical framework and pre-registration-style research design for analysing hallucination in large language models, built on Reality-Image Theory (RIT). Hallucination is defined as a violation of one of two norms. The formation norm requires that the functional evidence strength of an Image not exceed what its evidence base can support (representation hallucination). The use norm requires that the strength of an output assertion not exceed the Image’s own evidence strength (commitment hallucination). Deviation of the evidence base itself from Reality is treated as a predisposing factor, not as hallucination. The total deviation between output and Reality is decomposed as an identity into a commitment gap, a conversion gap and an evidence-base gap. In M6.1, evidence strength is a functional property defined relative to a family of probes, not an intrinsic property of an internal representation. The framework includes five axioms, risk and consolidation models, a gap-based governance framework, and nine falsifiable hypotheses (H0–H8), with H0 testing construct validity. Research is organised as gated phases (RRP-0 to RRP-3), starting with measurement validation. Status: This record contains theory and research design only. It contains no experimental data, and all hypotheses are untested. The theory is frozen as M6.1; later changes are limited to typo fixes, added literature and procedural clarification. This record is the English project description. The Chinese theory paper and working documents are archived separately.