Why this exists in ML Training uses finite samples; evaluation uncertainty comes from sampling variability.