Zovda
Every dataset has undocumented decisions: why this category was merged, why that month is missing, why the filter excludes a value that looks harmless.
They live in one person's memory. When that person changes team, the dataset keeps working and stops being maintainable.
What has to be written down
- The unit: what one row is
- The filters, in order, with reasons
- The known gaps and what caused them
- The decisions that were close calls
- The date each of those was made
The fourth list is the one that saves the most time later. A close call that is not recorded gets re-argued every year.
Write it next to the rows
A wiki page that is not linked from the data is a page nobody reads. The notes belong with the collection, versioned with it, and shipped with it.
Make the handoff a test
Give the document to someone who has never seen the set and ask them to answer five questions about it without help. The questions they cannot answer are the documentation's gaps, not theirs.
Include the queries
The three queries people run most often, written out. They are the fastest way into a set and the hardest to reconstruct.
Zovda ships each collection through Datuva with the decisions attached, so a change of team is not a change of meaning. The corpus catalogue shows the format.
Tell us what nobody has written down yet. We reply within one business day.