Measuring and Mitigating Silent Narrative Switching in Large Language Models through Epistemic Mediation

Abstract

The paper defines silent narrative switching, the covert change of a large language model’s dominant position across user contexts without disclosing that alternatives exist, and measures it with a Silent Narrative Switching Index (SNSI). Across four models from different providers, SNSI ranged from 0.73 to 0.80. An execution-level epistemic mediation layer over unmodified models reduced the pooled SNSI from 0.82 to 0.09 without retraining and without a detectable loss of contextual relevance; an ablation study locates most of the effect in the interpretation-generation stage.