An AI governance review asks where a dataset came from, whether its use was permitted, which transformations were applied, and how long it may be retained. Which practice best supports these questions?

1. Use a high model evaluation score as proof that all source-data rights were valid.
2. Record only the final number of model parameters.
3. Maintain documented data provenance, usage permissions, transformation lineage, ownership, and retention requirements.
4. Delete all source and transformation records as soon as training completes.