AI Autonomous Evolution Series — Part I Suppressing Hallucination for Trustworthy LLMs: A Layer–Knot Framework and HR/GR/CR Metrics Approach

This paper reframes hallucination in large language models not as an isolated factual error, but as a structural collapse of semantic reliability across the generative process. Rather than suppressing creativity through heuristic constraints, we propose the Layer-Knot Framework (LKF), a structural mechanism that stabilizes meaning by embedding semantic knots across representational layers. To enable empirical evaluation, we introduce three complementary metrics: Hallucination Rate (HR), Grounding Rate (GR), and Creativity Rate (CR), capturing factual distortion, evidence alignment, and semantic originality, respectively. Experiments conducted on TruthfulQA and over 2,000 domain-specific prompts demonstrate a 50% reduction in hallucination alongside a 12% increase in grounding, while maintaining stable creativity. These results suggest that trustworthiness and generative autonomy are not mutually exclusive, but can coexist when semantic coherence is structurally preserved. The proposed framework establishes a technical foundation for reliable and autonomous language systems grounded in meaning stability rather than output-level correction.