Learning under Distributional Drift: Reproducibility as an Intrinsic Statistical Resource