From Curves to Cascades: Composed Coupling Prediction Among Autonomous Agents

Autonomous agents that share a substrate couple through it, and prior work established that the resulting edges can be detected and their response curves measured — pairwise, with error bars, while the system is live. This paper asks the next question: do those local measurements compose? Can curves measured on neighboring links, chained through nonlinear switching elements and summed where paths converge, predict far-end behavior without walking the whole path? The referee design keeps the test honest: the pieces that compose into a prediction and the direct walk that judges it are disjoint measurements, and the comparison is made against propagated error bars built from the local noise of each piece. Across the campaign — four nonlinear chain seeds, a converging junction, and a live discontinuity relay across three seeds — 95 of 97 referee points fall within propagated two-sigma bars, and the propagated bars are themselves validated: the hit rate matches the nominal rate without over-rejection. The measurement horizon is measured, not assumed: roughly two nonlinear hops at noise σ = 0.02, beyond which the far end falls below the detection floor and composition stops being an economy and becomes the only remaining door. Boundary rows (a dead planted cell, a starved-scheduler pair) are reported as receipts, not discarded.