Provenance answers practical questions. Which system produced these records. When they were exported and by whom. What filters were applied. How personal data was removed and with what tools. Whether any records were added from another source or generated by a model. The answers are recorded in a data card and, in more detail, in a processing log kept by the seller.
Provenance is related to chain of title but is not the same thing. Chain of title is about rights: who owns the data and may license it. Provenance is about facts: where the data came from and what was done to it. A dataset can have clear rights and poor provenance if nobody recorded how it was assembled, or good provenance and defective rights. Buyers want both. Checksum manifests support provenance by fixing the exact files at each stage, so that a delivered file can be matched to the version that was reviewed. Clear provenance also helps a buyer answer questions from its own customers and regulators about training sources.
Provenance is easiest to record as the work is done. Reconstructing it afterward is slow and often incomplete.