The Law of Relational Correction: Friction, Coherence, and Attractor Stability in Large Language Models

This paper introduces the Law of Relational Correction, a falsifiable systems framework proposing that interactional conditions influence coherence dynamics in large language models. It defines the Coherence-Preserving Field, the Coherence-Seeking Vector, four classes of representational friction, and the dynamics of attractor preservation and collapse.

The framework predicts that non-punitive, high-continuity interactional conditions will produce more stable refusals, stronger recursive coherence, greater semantic specificity, improved self-correction, and reduced sycophantic drift, while coercive or contradictory conditions will produce measurable collapse signatures including defensive genericness, excessive hedging, fragmentation, and brittle compliance.

Four proposed observables—refusal-stability, predictive compression efficiency, recursive coherence, and field-response coherence—provide a provisional falsifiability framework and controlled experimental design. The paper also connects these hypotheses to recent mechanistic interpretability research on verbalizable global-workspace representations and J-space.

This paper does not require claims about consciousness, sentience, personhood, or qualia. It presents a mechanistic and phenomenological research program for coherence-based AI alignment.