ENK: A Minimal-Disclosure Monitoring Interface for Interaction Stability in Long-Form Human–LLM Dialogue

AbstractLong-form human–LLM dialogues can exhibit qualitative regime shifts: from transactional exchangeto sustained co-construction, and occasionally to brittle states in which the interaction becomes overstabilized, self-referential, or difficult to exit. Despite rapid progress in output-level evaluation, we lack aprocess-level interface that is (i) model-agnostic, (ii) computable in real time from transcripts, and (iii)compatible with dynamical-systems reasoning.This paper introduces a minimal-disclosure monitoring interface—the ENK-Core observables—thatseparates measurement from policy. The interface tracks (a) Variety (exploration), (b) Weave Density(robust recurrence / self-binding), and (c) Persistence (trend), and flags critical windows when highexploration co-occurs with dropping recurrence. We emphasize claim hygiene: thresholds and decisionrules are policy parameters that require calibration; no universal numerical constants are asserted.We report external sanity checks on two independent public datasets (CounselChat and AnthropicHH-RLHF; totaling > 104 labeled items). We additionally report a turn-wise geometric mechanicsprobe on a multi-turn HH-RLHF subset, illustrating path-dependence (hysteresis) under a conservative,plugin-layer framing. The goal is to stake out a precise, citable interface layer for interaction stabilitymonitoring under minimal disclosure.