What we fixed or changed that could affect how you read or reuse a figure, newest first. Spotted something wrong? Email [email protected] with what looks off and where you saw it.
Dates are the day a change was merged; it reaches the live site at the next release. The same list is available for programs as JSON. The source repository is private, so each entry points to a page you can open here rather than to the code.
Plain-English reasons for rejected comparisons
· Correction · Comparison checks (API and chat)
What was wrong. When the comparison check declined to call two measures comparable, the reason it gave could be one of our internal working notes (it mentioned a one-off data fix, another entry, or an internal name) instead of a sentence a reader or an AI assistant could quote.
Why it matters. The reason is what an assistant repeats to a person. It has to make sense on its own. The five affected stored reasons were rewritten on 9 October to say the same finding in plain words, and a check of the live responses found no internal wording left.
A vetted comparison no longer says it was not vetted
· Correction · Comparison checks (API and chat)
What was wrong. A pair of measures that a person had already reviewed and approved could still be told "not vetted by a human" in the comparison check's reasons. That sentence was false for those pairs.
Why it matters. A false reason undermines the one signal that says a person looked at the pair. Reviewed pairs now count as reviewed, and the check gives the real reason (how much data it has) instead.
Clean comparisons could never reach the top verdict; a human rejection could be ignored
· Correction · Comparison checks (API and chat)
What was wrong. Two faults in how a comparison verdict was reached. First, a pair with no outliers was treated as if its robustness re-run had failed, so a clean pair could never be rated higher than "qualified". Second, for a registered pair, a person's recorded rejection was never looked up, so the check could ignore it.
Why it matters. Both made the verdict less trustworthy: one understated good evidence, the other could overstate a pair a person had already rejected. Both are fixed and covered by tests.
A malformed search no longer causes a server error
· Change · Tender awards pages and API
What was wrong. A search containing a NUL (0x00) character, which automated scanners send, made the tender awards pages and several API endpoints fail with a server error (HTTP 500) instead of saying the request was invalid.
Why it matters. A server error looks like an outage and wastes the site's capacity. The request is now refused with a clear HTTP 400 message.
UK pages no longer offer a chat that is not available there
· Correction · United Kingdom pages
What was wrong. United Kingdom pages showed the chat box and "Ask StatsMapped" buttons, but the chat does not yet cover UK data, so a visitor could finish the human check and then be told it was unavailable, in wording that suggested a temporary fault.
Why it matters. Offering something that cannot work is a broken promise. The entry points are hidden where chat does not exist, and the refusal now says plainly that it is not available for that country yet.