Beyond Content: Proper Nouns and Semantic Governance Failures in LLMs

This paper extends prior work demonstrating that contemporary language models lack architectural primitives for semantic governance. While previous research established these failures at the content level (word meanings and referents), we document identical governance patterns across operational protocols, entity persistence, and deployment contexts. Through controlled experiments and naturalistic observation, we show that models cannot autonomously prioritize among competing semantic states or revoke prioritization when context changes, whether disambiguating polysemous terms, maintaining protocol boundaries, or preserving proper noun identity. Critically, we identify proper noun avoidance as a learned uncertainty management strategy arising from weak semantic axes and subjective interpretability. The findings demonstrate that semantic prioritization and revocation function as missing architectural primitives operating universally across all layers requiring semantic boundary maintenance, not as domain-specific capabilities. We evaluate S-vectors (Evans, 2025) as a proposed architectural solution, finding they address significance weighting but leave unresolved questions about entity knowledge representation and personalized semantic content.

This paper extends a sequence of investigations into coherence collapse, semantic drift, hallucination as fracture-repair, and significance weighting in transformer systems (Evans, 2024–2025). Earlier work established that hallucinations arise from fracture followed by probabilistic repair, and proposed S-vectors as a mechanism for significance weighting. The present work examines a harder case: whether semantic governance failures persist at the operational and entity layers, with particular focus on proper noun persistence. The goal is not to re-establish prior results, but to test whether the same missing primitives appear wherever semantic boundaries must be maintained.