Data provenance records a dataset’s origin, transformations, custody, and evolution. Establish it early so teams can interpret results, clarify responsibility, and avoid building on misunderstood data. section 4.10 section 4.5.4
Make provenance an explicit delivery artifact:
- Pin down the ultimate source of each dataset and document why and how it was collected. section 4.5.3
- Map tables, warehouses, lakes, mirrors, and transformations; use a data catalogue, lineage system, metadata processes, or even a spreadsheet. section 4.5.3
- Record migrations, replatforming, quality work, schema changes, rounding, corrections, and other events that may introduce noise. section 4.5.4
- Carry provenance into model documentation and chain-of-custody controls so models can be understood, governed, and selected appropriately. section 1.3
A practical provenance ticket
Before modeling begins, create a checkpoint requiring the team to identify the dataset’s ultimate source, collection purpose, transformations, custodians, and historical changes. Attach the available lineage map or spreadsheet, then record unresolved gaps and their risks. Revisit the record when surprising model results or ethical questions arise. section 4.10 section 4.5.4
