Housing data stewards debate when labeled datasets enter the lifecycle versus when the model is monitored in production. How should they separate those phases?

Housing data stewards should separate labeling and training from production monitoring. Labeling and training prepare the model before release; production monitoring watches live behavior after deployment—latency, accuracy proxies, drift, and incidents on the inference path. Labeling and training do prepare the model, but evaluation suites do not detect live feature drift after deployment; live feature drift detection is a production-monitoring responsibility after deployment, not an evaluation-suite function, and this option misplaces post-release observation inside evaluation. Feature engineering does build training inputs offline, but training pipelines do not validate live behavior after deployment; validating live behavior after deployment is production monitoring, not training, and this option wrongly assigns monitoring duties to training pipelines. Data labeling does create training inputs pre-release, but model registry approval does not track live drift after deployment; tracking live drift after deployment is production monitoring, not registry approval, and model registry approval is a separate release gate that does not define the labeling-versus-monitoring split. Exam caveat: weak labels or human feedback collected from production can later feed retraining, but that feedback loop still distinguishes the monitoring/observation phase from the preparatory labeling and training phases. Operational check: keep labeled housing datasets and training runs under the development track, deploy a monitored inference service separately, and only promote new labels into scheduled retraining after review—not by using model registry approval to track live drift after deployment.