Integration–Segregation Balance in Zero-Training Cognitive Regimes: Total Correlation, O-information, and MIP Structure

A system can either move as one whole or break into parts. Neither extreme is useful for cognition. The interesting behavior sits in the middle: enough coupling to share state, enough separation to keep specialized channels from collapsing into one mode. Technically, an instrumented zero-training cognitive runtime trace (15040 ticks) is analyzed using multivariate information measures computed on sliding windows. Total correlation (TC; multi-information) and dual total correlation (DTC; binding information) quantify global dependence in nats 1, 2, 3. Their difference, the O-information O ≡ TC − DTC, separates redundancy-dominated from synergy-dominated high-order dependence 4. Results show a consistent regime shift: low-entropy baselines exhibit higher TC (6.04 ± 0.45) and higher O (3.42 ± 0.70) than a high-entropy plateau (TC 4.79 ± 0.54, O 1.90 ± 0.71; Cohen’s d > 2). Synergy-dominated windows (O < 0) are rare in baselines (0.4%) but occur more often in the plateau (9.3%). Minimum-information partitions (MIP) further indicate that segregation is not random: baseline windows most often isolate the TD-error channel, while plateau windows most often isolate text-word emission.