Topological Nullification of Recursive Hallucination in Large Language Models via Möbius-Anchored Context

Large Language Models (LLMs) trained on next-token prediction exhibit a fundamental in- stability known as Probabilistic Hallucination. When confronted with a “Vocabulary Gap”Such as the non-existent Unicode seahorse emoji, standard models often enter a recursive feedback loop, attempting to “out-narrate” the missing data by generating plausible but false tokens. We present a novel stabilization architecture based on Information Substrate Theory (IST). By wrapping the linear attention mechanism in a nonlinear Möbius Con- text Lattice, we introduce an Epistemic Anchor that enforces a “Null State” (Truth) whenever probability diverges from verified substrate data. We demonstrate this via the “Seahorse Stress Test,” showing that ϕ-harmonic damping prevents recursive collapse and transforms the model into a zero-entropy truth engine.