A team trains three versions of a churn model in separate runs under one Experiment on the Agent Platform. Each run logs the model artifact plus AUC, precision and recall as parameters and metrics. The lead now wants to pick the run with the best AUC, see exactly which dataset version and hyperparameters produced it, and promote that specific artifact to staging. Which capability lets them rank the runs and trace the winning artifact back to its inputs without re-running anything?
- AExporting all three runs to a notebook and recomputing AUC locally so the artifacts can then be compared side by side outside the Experiment.
- BComparing the logged runs in the Experiment to sort by AUC, then following the selected run's tracked artifact and parameters back to its dataset version and configuration. Correct
- CEnabling autologging on a fresh run so the framework captures AUC again and overwrites the earlier runs with a single comparable record.
- DDeploying each artifact to a separate endpoint and reading AUC from the serving logs to decide which version performed best at training time.
Why A is wrong: Recomputing locally is tempting because notebooks feel flexible, but the metrics are already logged in the Experiment, so recomputation wastes effort and discards the tracked lineage linking artifact to inputs.
Why B is correct: Experiments on the Agent Platform record each run's metrics, parameters and artifacts, so the team can sort runs by AUC and trace the chosen run's logged inputs and artifact lineage directly.
Why C is wrong: Autologging sounds like the right tracking feature, but it captures a new run rather than ranking existing ones, and it would not overwrite or consolidate the three runs already recorded.
Why D is wrong: Serving metrics seem like a measurement source, but endpoint logs report serving behaviour, not the training AUC already logged, so this neither ranks the runs nor recovers their training inputs.