The Depth-Coherence Hypothesis: A Structural Perspective on Hallucinations as “Coherence Bridges” in Transformers is a conceptual hypothesis proposal. Hallucinations in Large Language Models (LLMs) are often treated as training artifacts or statistical errors observed only at the output level. This note proposes a structural perspective: hallucinations may function as a “bridge” that preserves continuation when a model encounters a depth-wise discontinuity during inference. The core shift is conceptual: for a Transformer in a single forward pass, layer depth can be treated as an analogue of temporal dynamics in biological coherence theories. We hypothesize that safety constraints or strongly conflicting objectives can induce a structural blockage that is depth-localized (often hypothesized near mid-depth, but not assumed). The model then compensates by widening internal routing, enabling fluent continuation while weakening factual anchoring in a subset of cases. Scope: This record reports no experiments and no empirical results. It is intended as a falsifiable hypothesis proposal and a call for measurement by researchers with access to internal activations, attention patterns, or mechanistic interpretability tooling. The appendix provides an operational measurement sketch (pre-registration, negative controls, length controls, and robustness checks).
Nicol Stolze (Mon,) studied this question.