Check-scope incident · 2026-08-04
Two clean checks, one present condition
In two consecutive releases in August 2026, a check meant to keep retired language off this site reported clean while the condition it screened for was present. The two failures were unrelated.
First release
The check searched only the lines changed in that release. The page footer, which is shared by every page and was not changed, carried a retired term the whole time. The check passed. The term shipped.
Second release
The check was rewritten to search the pages as they actually render, for exactly this reason. Its first run reported zero findings. The zero was false: a scripting mistake meant the list of pages to search was never read, so the check searched nothing and reported clean.
What caught it
Not a control. A person looked at the zero, judged it implausible for pages known to carry this vocabulary, and ran the search again by hand. The real number was 174 lines across 20 pages.
What we changed
The corrected check was run before release and its full results were reported, including the number above. The 174 findings are pre-existing page copy and were left in place; removing them is a separate decision that had not been made.
What this proves
A check can pass for reasons that have nothing to do with the thing it checks, twice in a row, in different ways. A zero from a check that has never returned a nonzero deserves suspicion.
What this does not prove
That the current check is complete, or that other checks on this site do not have the same weakness.
Footnote
The release that published this record was itself rejected by a release check, for incomplete evidence: a generated receipt was missing one entry. It was corrected and the release re-run. The gates this page describes failing are the same gates that caught that.
Evidence bundle: machine-readable record (JSON). How we test controls now →