Suppressing Hallucination for Trustworthy LLMs (Part IV): Semantic Integrity and Autonomous Meaning Flow

This paper reconceptualizes hallucination in large language models (LLMs) as a collapse of semantic reliability rather than a simple technical error. To address this, we propose the Layer-Knot Framework (LKF), which stabilizes meaning by embedding semantic knots across network layers, thereby preventing contextual drift without diminishing generative autonomy. The framework is evaluated using three quantitative indicators: Hallucination Rate (HR), Grounding Rate (GR), and Creativity Rate (CR). Experiments on TruthfulQA and 2,000 cross-domain prompts demonstrate a 50% reduction in HR, a 12% improvement in GR, and stable CR performance. These findings show that accuracy and creativity are not mutually exclusive; when semantic integrity is structurally enforced, a model can maintain factual reliability while preserving originality. The study establishes LKF as a principled approach to trustworthy AI, positioning hallucination suppression as a process of semantic alignment rather than data expansion. This provides a foundational methodology for developing high-reliability generative systems grounded in coherent internal meaning structures.