7,317 recorded changes across 30 editions. Derived from the repository history rather than written by hand, so it cannot describe a change the data did not make — and cannot omit one it did.
A change is classified by the most specific thing it did, so one commit that re-dated a record while revising a coverage cell is reported as the revision, not as both.
Where the dataset has grown. A record counted here arrived; a record corrected later appears in the editions below rather than twice here.
These are totals for the whole history of the dataset, not for a period of operation. The history starts where the dataset does, on 2026-09-06, so most of what is counted here is the record being assembled for the first time rather than a month’s worth of change up to 2026-09-07. Read the 4,274 additions as the size of the record, and the editions below as its movement.
Newest first. Each edition is one change to the committed data, with the records it touched and the fields that moved. There is no subscription on this page: it is a page, and it stays one until we have something worth sending.
Every commit that touched a data file is read, the file parsed on both sides of it, and each record compared. A hand-written changelog drifts from the data within a month, and the raw diff is unreadable — nine thousand changed lines of JSON says nothing about whether a system was added or a source re-read. So the classification is derived, and the page is regenerated on every build: a stale digest shows up as a diff rather than as a wrong page.
The same data is downloadable in full from the open dataset, where every record carries its own verification date.