A fraud detection model is trained using a feature that represents whether the transaction was flagged for manual review. In production, this flag is only assigned after a human reviews the transaction. What kind of leakage is this?