Replication materials: controlled-injection audit of evaluation-protocol leakage in football player evaluation

Replication materials for a registered, controlled-injection audit of evaluation-protocol leakage in machine-learning player evaluation in professional football. From an honest evaluation protocol (team-blocked cross-validation, contemporaneous features only), single protocol violations are injected one at a time (random splitting, identity features, future-information features at three contamination strengths, label-period position features, a label-collinear feature, and a composite sloppy protocol) across ten professional competition units, two tasks, two model families, and 50 paired resamples per condition. The archive contains the analysis code, the derived player-level feature matrices, and all per-seed battery outputs and registered-analysis result files. The analysis plan is registered on OSF with dated amendments; the registration link is disclosed on the article title page. Raw event data are not redistributed; they are openly available from the Wyscout public dataset (CC BY 4.0) and StatsBomb open data.