Abstract

Hallucination in large language models is conventionally treated as a single, model-internal pathology — a defect of the generator to be reduced through better training, retrieval, or decoding. This paper argues that the category is not homogeneous. Two structurally distinct phenomena are conflated under the one word. Class 1, floor hallucination, is the statistically irreducible error rate intrinsic to next-token prediction over a finite parametric store; recent work argues that, under standard next-token prediction and evaluation regimes, a positive hallucination floor is structural rather than incidental. Class 2, field-maintenance hallucination, is the loss of established accuracy that occurs in extended interaction when the relational field between human and model — the active configuration of constraints, definitions, and frame on which the output is conditioned — is no longer maintained. Class 2 is not a property of the model in isolation; it is a property of a relation that has ceased to be sustained. Building on Primordial Separation Ontology (PSO) and Triadic Relational Ontology (TRO), the paper develops a breakdown taxonomy for Class 2 (context truncation, constraint evaporation, frame drift, scope dissolution), anchors the taxonomy in independently documented machine-learning phenomena (positional salience decay, error snowballing, sycophancy), and grounds the account in the attentional substrate of the transformer. The framework yields a falsifiable operational discriminator — the re-injection test — that separates the two classes empirically, though the test measures recoverability, not relationality as such. The claim is bounded: TRO does not dispute the statistical floor, does not assert a categorical discontinuity where there may be only a gradient, and does not deny that what it tracks could be redescribed in purely mechanistic terms. It isolates a practically significant and frequently misattributed subset of confabulation, relocates the relevant intervention from the generator to the maintained relation, and proposes a protocol whose execution would establish whether the distinction earns its keep beyond nomenclature. This is a hypothesis paper; the protocol is proposed, not yet run.