Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis

Fuente: Zenodo
Enregistré dans:
Détails bibliographiques
Auteur principal: Teixido, John
Format: Recurso digital
Publié: Zenodo 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866902308331192320
author Teixido, John
author_facet Teixido, John
contents <p>The tendency of Large Language Models (LLMs) to confidently generate factually incorrect information—termed “hallucination”—remains a critical barrier to enterprise deployment. Current mitigation strategies rely on semantic evaluation or external retrieval (RAG), which are computationally expensive and post-hoc. We propose a fundamental, pre-generation solution grounded in Topological Analytical Homeostasis (TAH). We introduce a method for calculating the Topological Strain of a generated token vector relative to the machine-verified Teixido Envelope (τ = −0.5). By measuring the geometric deviation of the activation manifold in real-time, the system provides a deterministic, mathematical confidence score before the token is rendered to the user. We detail the implementation of this ”Hallucination Meter” and the Teixido Inhibitory Gating (TIG) mechanism, which physically intercepts and suppresses out-of-distribution confabulations at the logic level.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19652657
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis
Teixido, John
LLM Hallucination
Topological Strain
AI Safety
Teixido Envelope
Contextual Monitoring
Zero-MAC
<p>The tendency of Large Language Models (LLMs) to confidently generate factually incorrect information—termed “hallucination”—remains a critical barrier to enterprise deployment. Current mitigation strategies rely on semantic evaluation or external retrieval (RAG), which are computationally expensive and post-hoc. We propose a fundamental, pre-generation solution grounded in Topological Analytical Homeostasis (TAH). We introduce a method for calculating the Topological Strain of a generated token vector relative to the machine-verified Teixido Envelope (τ = −0.5). By measuring the geometric deviation of the activation manifold in real-time, the system provides a deterministic, mathematical confidence score before the token is rendered to the user. We detail the implementation of this ”Hallucination Meter” and the Teixido Inhibitory Gating (TIG) mechanism, which physically intercepts and suppresses out-of-distribution confabulations at the logic level.</p>
title Real-Time Hallucination Detection in Large Language Models via Topological Strain and Geometric Homeostasis
topic LLM Hallucination
Topological Strain
AI Safety
Teixido Envelope
Contextual Monitoring
Zero-MAC
url https://doi.org/10.5281/zenodo.19652657