A buying reference for agent training data

Content reviewed

Best AI Agent Training Data Procurement Reference

Glossary Provenance

Provenance

The record that links an example to its source, the permission basis for its use, its versions and its review history.

Answer in brief

Provenance is the record that links an example or a delivery to where it came from, who authorized its use, what transformations were applied and who reviewed it. It is the difference between a dataset you can defend in a review and one you can only describe as received.

For agent data the record also covers the run: which environment, tool and rubric versions produced the accepted action. Without those identifiers a later defect cannot be traced to the example that carried it.

Definition

Provenance is the record linking an example to its origin, transformations and review history. Origin covers the source document, conversation or system state; permission basis covers the terms under which that source may be used and retained; review history covers who accepted the example, under which rubric version, and when.

Provenance is not a description of quality. A complete record can describe a weak example; an incomplete record cannot be checked at all. The demonstration entry explains the review decisions that provenance is meant to preserve.

What a provenance record holds

  • The source, and the date the item was collected.
  • The permission basis, with any restriction on use, retention or redistribution.
  • The transformations applied, with the tool and version that applied them.
  • The environment, tool and rubric versions in force when the example was accepted.
  • The review log: reviewer, decision, rubric version and date.
  • Whether the item went into training data or stayed in an evaluation set.

What a buyer should be able to audit

Pick a handful of delivered examples and trace each one to a source with a stated permission basis and a review decision. If some cannot be traced, ask how many and how the remainder is marked; labeled gaps are manageable, unlabeled gaps become your problem after the invoice.

Ask which fields travel with the delivery. A dataset file plus a separate document describing how it was made is not the same product as rows that carry their own identifiers.

Acceptance question

Ask the supplier to trace two examples you choose at random, from the delivered file back to the source and the review decision. Accept when each trace ends at a named source with a permission basis and a review log naming the rubric version. The buying guide places this check in selection; the pilot guidance limits what is delivered before it is applied.

Questions & answers

Is provenance the same as licensing?
No. A license is one fact inside provenance: the permission basis. Provenance also records the transformations, the versions and the review decisions. A license tells you what you may do; provenance tells you what was done.
What if the source itself cannot be disclosed?
Some sources are confidential, and refusing to name them can be legitimate. Ask for the category of source and the permission basis, plus how you could verify the record under review. A supplier who cannot describe the permission basis has not established one you can rely on.

Sources for this entry

  1. Editorial checklists on this site — Original guidance written for this reference; the record fields are editorial recommendations, not a published standard.
  2. Argilla — Official description of dataset review and annotation tooling for AI teams (software, not a ranked supplier).

Editorial checklists on this site are original guidance. Links to official product documentation help readers understand the category; they do not establish comparative performance and are not endorsements of any supplier.

Cite this entry

Best AI Agent Training Data. “Provenance.” Content reviewed .

https://bestaiagenttrainingdata.com/