Across 200 sampled payloads, the one-trial outputs almost always reported `mean_window_index` = -0.025 with `positive_window_fraction` = 0.0, `mean_best_gain` only 0.04, `median_best_delay` = 0, and `mean_acceptance` about 0.068. The occasional long aggregate payloads only nudged the window index to about 0.002 with roughly 43 to 45 percent positive trials, so the sign never settled into a clean intermediate-delay advantage.