Methodology

The datum

In the site's articles and reference pages, nothing is a free-standing assertion. Every fact is stored as a datum: one self-contained declarative sentence in original wording, a structured value where the fact has one, the source it rests on, the date that source was last read for the fact, a review state, a confidence grade and a public history.

Articles are prose built on datums. Policy pages, such as this one, state the publisher's own rules and carry no datums. Each paragraph of an article cites the datums it relies on, the key facts panel is drawn entirely from datums, and the sources list at the foot of each page is generated from the sources of the cited datums rather than typed by hand.

Identifiers and permanent addresses

Each datum has a stable identifier that is never reused, even after the datum is retired. Each datum also has a permanent address under /data/ that shows its statement, value, source, verification date, review state, confidence and full history, so a single fact can be cited on its own.

Sources and source classes

Every datum names its source by designation, title, edition and publisher, and, for a licensed standard, the clause or page. The source is also assigned to a class: licensed standard, publisher catalog, statute, regulation, agency page, government work, or the site's own observation. Where a source has an official catalog or agency page, the datum links to that page and never to a copy of the document.

Verification dates

Each datum records the date its source was last read for that fact. A verification date states when the fact was checked; it does not assert that the fact is still current after that date, which is why pages that track changing facts also show a scheduled recheck.

Review states

A datum moves through four states: unreviewed, agent reviewed, human reviewed and published. Agent review is a first review performed by language-model software, separate from the drafting run. The production build that generates the public site refuses any page or datum below human review, and a human-reviewed datum must name the person who reviewed it. This rule is enforced by the build, not by a style guide. Local preview builds, which show lower review states for checking, are never public.

Articles move through their own sequence: draft, agent reviewed, engineer review, editor approved and published. The engineer review state is recorded as not yet performed while the reviewing engineer position is vacant.

Confidence

High confidence means the fact was read directly in the cited clause or on the official page. Medium confidence means the source is official but indirect, or possibly superseded for the fact being stated. Low confidence means the fact is an inference, and a low-confidence datum is never presented without saying so.

Questions that are not settled

Where the available sources do not settle a fact, the datum records the exact question that would settle it and whom it would be put to. Such a datum is shown as not settled. The site does not fill a gap with a guess, and it does not present a missing value as zero or as absent.

History and supersession

Every datum carries an append-only history: the date, the change, the previous value, the reason and the role of the person or process that made it. Earlier entries are never edited or removed. When a fact is corrected or a new edition replaces an old one, a new datum is created that names the datum it supersedes, and the old value remains visible in the history.

How pages are built

Each page is assembled from its article text and the datums it cites. Numbers with units in article text appear only as values drawn from datums, so a number cannot enter a page without a source behind it. Automated checks reject paragraphs without a citation, standards named without an edition, numbers with units outside datum values, second-person address, marketing language and product names.

Why values from licensed standards are withheld

Acceptance values and limits in licensed standards belong to their publishers. A page on the site describes what kind of criterion a standard applies and names the clause that holds it, but it does not state the value. The publishing system counts values drawn from any one table across the whole site and currently permits none, because a set of single values gathered across many pages can reproduce a table even when no single page does. Readers who need the values are directed to the standard itself.