Data and code for "A Reproducible Benchmark for Leakage-Resistant Machine Learning on Multi-Site Brain Connectivity Graphs"

Stored model predictions, analysis code, and trained artifacts for the manuscript "A Reproducible Benchmark for Leakage-Resistant Machine Learning on Multi-Site Brain Connectivity Graphs" (submitted to Diagnostics, MDPI). Everything needed to reproduce every table, figure, and statistic in the paper. We audited an autism-screening pipeline of the kind this literature builds, traced its reported balanced accuracy of 0.999 to five data-leakage pathways, and closed each one. Under a leakage-resistant leave-one-site-out protocol on 871 participants from ABIDE-I, a graph attention network (area under the curve 0.648) did not outperform a support vector machine on the same connectivity (0.679); the result held on a locked external cohort. Contents: predictions/ (per-participant out-of-fold predictions), results/ (derived metrics), code/ (analysis and figures), models/ (trained artifacts). No participant-level imaging data are redistributed; the cohort is the public ABIDE-I release (http://preprocessed-connectomes-project.org/abide/). Code MIT, data CC BY 4.0.