Method

What is published here

285 datasets, 110,579 records, assembled from 851 publishers. Each record states one fact — an error code and its meaning, a version and the date its support ends, a threshold and what it triggers — and each has its own page and its own stable URL.

The datasets are chosen for a particular shape of question: one with a definite public answer that is scattered across many publishers, changes over time, or varies by jurisdiction. Those are the questions where a confident answer from memory is most likely to be a wrong one.

How a record is made

• Sources are chosen first, and recorded. Manufacturer documentation, regulators, standards bodies and official support pages before anything else.
• Values are copied, never normalised. If a source prints a code one way, that is how it appears here, even where another convention would be tidier.
• Each value is stored with the span of text it came from — not a paraphrase, not a summary, the passage itself.
• Every claim is then checked back against that passage before the page is built.
• Nothing is filled in. Where a source does not state a value, the record omits it. A plausible value is worse than a missing one, because it is indistinguishable from a checked one.

What the markings mean

• verifiedThe value appears word for word in the passage quoted directly below it, from the source cited. A program re-checks that match against the fetched page every time the site is built.
• each item verifiedThe value is a list of several things the source prints separately — product codes on one import alert, say. Each item in the list appears word for word in the passage quoted below; the list itself is ours, in the order the source prints them.
• our readingOur classification of what the source says, not its words. Some comparisons only exist once differently-worded rules are put on one scale — that scale is ours, and it is labelled wherever it appears. The quote below it is the evidence; the judgement is ours and can be wrong.
• UnmarkedTaken from the cited source, and shown with the passage it came from, where that passage does not contain the value word for word. Two ordinary reasons: the wording differs — a source printing “carry-on baggage only” for a value recorded as “carry-on only” — or the source states the fact somewhere other than the passage quoted here, in a column heading, a footnote or the document’s own title. Unmarked is not a doubt about the value. It means only that this quote does not settle it, so we do not say it does. The source and the quote are on the page; read them and judge.

What “verified” covers

It means the value appears in the passage quoted from the page cited, on the date shown. It is a statement about what the source says — not a warranty that the source itself is right, or still right today. Checking that a page says a thing is not the same as checking that the thing is true, and only the first is mechanical.

This is why the quote is on every page. You do not have to take the value on trust: follow the link and read the sentence.

Freshness

Reference data decays. Every dataset carries the date it was last checked against its sources and a date by which it is due to be checked again. When a page passes that date without a re-check, it says so at the top rather than looking current.

That first date moves on its own, without any record being edited. Datasets come round in rotation: every page a dataset cites is downloaded again and every stored quote looked for on it. Where all of them are still there, the date shown becomes the day of that pass. Where even one cited page would not load, nothing is claimed for it and the date stays where it was — a partial reading cannot say every number was checked. So the date answers when was every value here last read off its own source, which is not the same as when the dataset was last changed.

A page may show two dates in its opening sentence — recorded from its source on one date, source re-checked on a later one. The first is the day the value was taken off the page. The second is the day that page was downloaded again and the quoted passage found still on it. They are separate claims, and where a value is marked our reading rather than verified the re-check speaks only to the quote.

Pages are not deleted for being stale — a stale page that says it is stale is more useful than a broken link. Anything found to be wrong is corrected or withdrawn.

Disagreement between sources

Where two authorities state different values for the same thing, both appear on the same page, each with its own source, its own quote and its own date. Choosing between them would hide the most useful part of the answer, which is that the disagreement exists.

Corrections

An error in a value, a quote that does not support what it is attached to, or a source that has changed — all are worth reporting, and all are acted on rather than argued with. Write to [email protected].

Reuse

Facts are not owned, but a publisher's expression of them can be. Each dataset carries its own licence, stated on its page, and each source's terms were checked before its rows were included; rows from a source whose terms forbid reuse are not published. If you use this data, cite the original sources — they are listed on every record for exactly that reason.

For machines

Every page except the interactive ones — calculators, list pages and the /play/ puzzles — is also served as Markdown at the same path (append index.md), this page included; every dataset as JSON at /<dataset>/data.json, and the whole catalogue at /catalog.json. Pages carry schema.org Dataset and Article metadata, and each record opens with a single plain sentence stating the fact, so there is something to quote that is not a table cell. No JavaScript is used anywhere on this site and no request leaves it — the HTML you fetch is the whole page. AI crawlers are explicitly allowed in robots.txt.