TopoDialogue III: Beyond the Transcript — Observable Implicit Interactional Structure in Human–AI Dialogue

What is often described as the “unspoken” part of human–AI dialogue should not be operationalised by guessing a participant’s hidden state. This instrumentation pilot extends TopoDialogue with seven observable interaction-event candidates: revisit, repair, rejection, adoption, handover, reframing, and explicit pause. It applies deterministic extraction to one 114-turn Japanese human–AI dialogue while withholding every original utterance from the public derivative. The released representation retains speaker order, coarse descriptive metadata, pseudonymous recurrence regions, and event candidates. The paper reports a representation-sensitive trajectory comparison and a low projection-stability failure case. It includes an annotation protocol and per-label agreement software, but does not claim that two-person human agreement has already been measured. The work makes no psychological, clinical, causal, or mathematical-topology claim.