Dialogue as a Coupled Neural System: Interpersonal Synchronization, Shared Reference Frames, and Informative Divergence

Abstract

Two people in dialogue must become similar enough to establish common reference while remaining different enough to contribute independent evidence, correct errors, and explore alternatives. This paper develops a source-faithful Self Aware Networks framework for dialogue as an environment-mediated coupled neural system. It separates shared stimulation, movement, physiological covariation, neural synchrony, representational alignment, directional information transfer, comprehension, and task success rather than treating them as one measure. The article defines a Dialogue Coupling Benchmark, receiver-relative Phase Wave Differentials, task-conditioned partial-information quantities, strongest-form countermodels, matched perturbation and rescue tests, and a prospective dual-EEG and dual-fNIRS protocol. A bounded Lean 4 kernel machine-checks finite route, consent/privacy, multiplicity, decision-history, provenance, and lineage invariants within the declared model. A completed controlled pilot used fixed local Qwen2.5-1.5B-Instruct and Qwen2.5-0.5B-Instruct models under frozen prompts, tasks, parsers, and decision rules. Its preregistered superiority gate failed: current dialogue tied isolation on complementary-object decisions, and the current-over-stale correction difference missed the practical and multiplicity criteria. Exact replay reproduced every live prediction. The adverse result is retained and narrows the application claim. It is evidence about this deterministic language-model preparation, not evidence from human participants or neural recordings.