Open Data and Attribution: Doing It Right Without the Footnotes

Attribution is the cost of using open data. Done badly, it is a wall of footnotes. Done well, it is a single line at the top of the dataset and a single line in the prompt.

The minimum

• Source name (the entity that maintains the canonical version).
• Source URL (the URL where the canonical version lives).
• Retrieved date (the date you pulled the data; refresh cadence implied).
• License (the license under which the source publishes the data).

Where to put it

• Top of the dataset, as a provenance block.
• At the row level, for derived values that come from a specific source.
• In the prompt, as a single line above the table.

Common mistakes

• Attributing the publisher instead of the maintainer. They are not always the same.
• Forgetting the retrieved date. Open data changes; the date scopes the snapshot.
• Attributing in a way that is invisible to the model. The line must be inside the prompt, not in a comment.

YJTOON's rule

Every published dataset carries a provenance block in the JSON, a header comment in the TOON, and a source_url column where it makes sense. The same URL shows up in the prompt. One URL, one source of truth.