A data scientist runs dozens of ad-hoc training scripts from a SageMaker notebook while iterating on a churn model. Each script trains an XGBoost model with slightly different feature sets and logs metrics. Later, the team cannot tell which parameters and input datasets produced the best validation AUC, because all runs were logged in a flat, unstructured way. Going forward, they want each experimental attempt grouped and comparable so any run's parameters, metrics, and artifacts can be traced. Which approach using SageMaker Experiments best organizes this work?