A public-health team must choose among MLflow-tracked runs for a disease-outbreak early-warning model by comparing AUC and recall. What should they do?

Comparing AUC and recall across MLflow-tracked runs and selecting the run that best meets early-warning criteria is how a public-health team should choose among Azure Machine Learning experiment candidates. Outbreak early-warning models often trade precision-like behavior against recall of true events; reading both metrics against policy thresholds beats picking by convenience. MLflow-tracked history makes those comparisons repeatable for after-action review. Choosing the run that best meets the training-split criteria fails by concept: it optimizes for reproducing training-split conditions rather than the early-warning acceptance thresholds for AUC and recall, so it can promote a run that looks good on the split but misses true outbreak cases. Choosing the run that best meets precision-focused criteria fails selection discipline: it prioritizes avoiding false alarms over catching true outbreaks, and it ignores the early-warning acceptance criteria the stem requires where recall of true events is critical. Selecting the run that best meets latency-based criteria fails governance: latency is an operational speed metric, not predictive quality, and optimizing for it skips applying the early-warning acceptance thresholds to AUC and recall that the stem required. Exam caveat: Domain 2 job/run comparison items expect metric-based selection (AUC and recall together against early-warning criteria); optimizing for training-split criteria, precision-focused criteria, or latency-based criteria are classic wrong answers. Operational check: list MLflow runs for the outbreak experiment, tabulate AUC and recall, apply the team's early-warning acceptance thresholds, register only the selected run's model, and document why others were rejected.