CTSignal
Thursday, November 19, 2026 · automated data desk

Corrections

We are wrong in public, on this page, with receipts.

How corrections work

How this differs from a normal newsroom

Most errors here are mechanical (an upstream dataset revised, an endpoint changed shape), so most fixes are automatic on the next fetch. What we list below are the ones where a reader saw something we'd have kept serving.

Corrections to date

  1. 2026-10-06 — the unemployment answer contradicted itself. The story card showed 5.1% (August, "2nd-highest of 9 peers", cached) while the board tile showed 5.8% (July, 3rd-highest of 52). A partial upstream cohort had outranked a complete print. The fix is in the code: cohort-consistent dates at the source and a completeness rule in revalidation — a nine-peer answer can no longer beat a 52-peer one. What readers saw for the day: two defensible numbers and no explanation, which is exactly the kind of thing this page exists to admit.
  2. 2026-10-06 — we published a credential in our own failure log. A bug in our Data Commons adapter on October 3 wrote a failed request URL — API key attached — into data/failures.log, which the pipeline commits publicly. The key was rotated the moment we caught it; the log is scrubbed, and failure lines are now redacted before they are ever written. The old key's value is a dead credential; the lesson is printed where you can check it.

That is a small list, not a claim of perfection — our failure log is public and longer than our corrections list on purpose: failures that stop a card from publishing never needed correcting.