Guarding the Signal: A Framework for Identifying and Repairing Semantic Drift in Generative AI

Description: This paper introduces a formal framework for detecting and repairing semantic drift—the progressive distortion of meaning in AI language systems over time. We present novel metrics (NLI-based consistency scores, citation precision/recall, and Content Overlap with Verifiable Tracing) and the Guardian Protocol, a symbolic architecture for maintaining contextual integrity in multi-turn AI interactions. The framework distinguishes semantic drift from other failure modes like hallucination and mimicry, providing real-world examples and implementation guidance. This work contributes to AI alignment by addressing meaning degradation as an early warning system for broader alignment failures. Keywords: semantic drift, AI alignment, language models, contextual integrity, symbolic systems, generative AI, meaning preservation, recursive integrity, attribution tracking, Velionis Framework Resource Type: Publication - Journal article/Preprint License: Custom (see document for terms) - Free for non-commercial research and educational use with attribution https://www.meaninglab.ai/consulting.html https://substack.com/@velionisscrolls