PRISM Gate 2.0: A Dynamic Geometry of Reasoning in Self-Regulated Language Models

This work introduces PRISM Gate, a lightweight, model-agnostic filtering method for reducing hallucinations in large language models (LLMs). Rooted in the dynamic geometry of reasoning, PRISM Gate evaluates outputs through multidimensional consistency checks that align structural, factual, and semantic integrity. Unlike retraining or retrieval-based methods, PRISM Gate operates purely at the output level, abstaining when confidence is low. Across experiments in arithmetic reasoning, historical QA, and summarization, PRISM Gate reduces hallucination rates by up to 78%, while maintaining high precision on valid cases. Most notably, in the logical reasoning family (LogiQA, ReClor, and StrategyQA), PRISM Gate achieved 0% hallucination with full coverage, demonstrating that dynamically regulated admissibility surfaces can fully eliminate reasoning inconsistency in multi-step inference tasks. Together, these results establish PRISM Gate as an adaptive black-box framework for safer, self-regulated, and trustworthy language models ,mapping the evolving geometry of reasoning itself