Code and data for "Non-Stationarity Breaks Permutation Surrogates in Multi-Agent Reinforcement Learning: Diagnosis and Remedies"

Abstract

Analysis code, generated data and the 100-seed sweep for the transfer-entropy null-control experiment reported in the manuscript. Includes the four-way surrogate decomposition, the equal-length window control, the stationarity diagnostics and gate enumeration, the synthetic drift manipulation, the effect-size titration, and the regression guard that checks every reported number against these files.