Negative Regulative Language as a Control Instrument for AI State-Flow Stability

This working paper examines how negative, non-descriptive language functions as a control instrument for regulating implicit instability in AI-generated state flows. Drawing on interaction traces between a human expert user and a large language model, the study focuses on instances where explicit rules and prompts remain formally correct while generative behavior exhibits structural drift. The paper adopts a state-based analytical framework to trace how human interventions—specifically those articulated through negation, proportional correction, and boundary-setting—affect AI output stability. Rather than proposing a new optimization technique or architectural modification, the paper limits its scope to observational and analytical claims regarding interaction-level control mechanisms. It does not evaluate performance metrics, benchmark accuracy, or generalize findings beyond interactional contexts.