Semantic Fidelity Lab Failure Modes Series: LLM Drift, RAG Evaluation, and Meaning Preservation in AI Systems

This bundle contains the Semantic Fidelity Lab Failure Modes Series, a collection of short papers documenting recurring failure modes in large language model systems where outputs remain coherent while losing alignment with meaning, intent, and underlying reality. The series frames these failures as expressions of a shared structural problem: meaning degrades across transformations while systems continue to function. Across eight papers, the bundle examines semantic misalignment, benchmark failure, embedding similarity errors, multi-agent drift, agent drift measurement, retrieval versus interpretation failure, multi-step inconsistency, and RAG evaluation beyond accuracy. Together, the papers argue that modern AI systems often fail not by visibly breaking, but by remaining fluent, plausible, and operational while drifting away from the reality they are meant to represent. The series introduces semantic fidelity as a practical evaluation lens for assessing whether meaning and intent are preserved across representation, retrieval, reasoning, and generation. This upload includes PDF versions of the eight Failure Modes papers and a README overview for repository-style navigation.