A team needs to identify which input dataset and code produced a registered model version. Which approach best fits?

1. Minimize unnecessary sensitive inputs and enforce approved data access.
2. Define labeling guidelines and measure and resolve annotation disagreements.
3. Maintain model, dataset, and training-code provenance.
4. Use time-aware splits that respect prediction chronology.