Topological Constraints for Coherent Language Models: Why Geometry Prevents Hallucination

Hallucination in large language models arises from unconstrained latent dynamics. We show that toroidal attention masks with explicit spectral gaps bound semantic drift. Key results: • +19.5% relative improvement on TruthfulQA (14.44% → 17.26%) • Best HaluEval performance (52.60% vs 55.00% baseline) • 40% drift reduction on Tonnetz sequences Hierarchy: mHC (Birkhoff) ⊂ ERLHS (Hamiltonian) ⊂ Karmonic (Toroidal + Spectral) Code: https://github.com/Paraxiom/topological-coherence Demo: https://huggingface.co/spaces/paraxiom-research/topological-coherence Software DOI: 10.5281/zenodo.18187835