Task Temporal Structure Strength Determines the Value of Learnable Coupling Matrices

This paper investigates how the temporal structure strength of a task determines the engineering value of learnable coupling matrices. We compare four reasoning architectures across three tasks (MNIST no temporal structure / Cartpole weak structure / Collision strong structure): G4 static-fusion baseline, G4+ parameter-matched control, G5 original information-bottleneck (audit vectors from action/precursor streams only), and G5+ repaired variant (full-stream audit). Main results: on Collision, the G4 to G5+ coupling gain is 69.3% (train) / 66.0% (ood), winning on all 11 seeds (Wilcoxon p=0.0005, permutation p=0.0006, Cohen dz=3.56); cross-lag temporal structure std is 0.0099 (Collision) vs 0.0012 (Cartpole), an 8.25x ratio, supporting that temporal structure strength determines coupling value. We also give three testable follow-up predictions (ACF stratification / W_lag null-model control / lead_steps sweep).