Use cases: - When data quality is mostly good, users want to just fix wrong label that they noticed during analytics, rather than going through external labeling. - model-in-the-loop labeling.