Tracking Cognitive Trajectories in Multi-Agent Dialogue : A Longitudinal K-CDI Demonstration (companion report to Kunisawa, 2026)

What this report is

The companion paper introduces **K-CDI**, a multi-metric instrument that profiles the cognitive-divergence signature of a text-generating agent from externally observable outputs alone. There, K-CDI is applied *cross-sectionally*: one profile per agent per condition, a static snapshot. This short report demonstrates the **longitudinal** use of the same instrument. Instead of measuring a single Q1→Q2 displacement, we treat a multi-turn conversation as a *sequence of displacements* and compute K-CDI **per turn**, tracing how each speaker's cognitive state moves through the exchange, and how the distance between speakers evolves. Only metrics that are well defined for a single response are used (LE, cgc_nodes, cgc_dispersion, and — relative to each speaker's own previous turn — TEI, LN, and the between-speaker topic distance). Metrics requiring replication or variance (e.g. SD) are, by construction, excluded from an n = 1 series. This is a demonstration, not a controlled experiment. Its purpose is to show that the movement itself is measurable.