How we verify

Claim log

A page is not verified as a whole. It is broken into atomic claims — one assertion each — and every claim is routed to the archive or agency that could settle it. The source log on each page is that list, in order, with the figure or wording quoted as the source states it rather than as the page paraphrases it.

The quoted figure is typed from the source by the person doing the checking. It is never copied from the draft and never filled in automatically. That friction is the point: the failure this archive is built to prevent is confirming a claim without opening the record.

Source list

Lower is closer to the record. Every claim type carries a minimum tier, and a claim that cannot meet its tier does not publish.

• Source name
• Source link
• Quote
• Locator

Claims and sources

History does not expire. The present tense does. Claims about access and about what is physically there now carry a re-check window, and once it passes the page says so at the top rather than quietly leaving the date off.

Claim log status

• Sourced
• Needs source

A page publishes only when every claim on it is confirmed, adjusted, disputed. A single failed claim blocks the whole page. The claim is cut or the page waits; it is not published around.

Disagreements logged

• Waiting for review
• Approve and post
• Reject and keep private

Where two sources of comparable standing conflict, the claim is marked disputed and the page prints the disagreement. This archive has no standing to decide between two agencies, and pretending otherwise would be the most expensive kind of error to make.

What the pages refuse to say

Every page carries a “what this page doesn’t say” section listing the things a reader might reasonably expect and the reason each is absent. An absence with no reason attached cannot be told apart from an oversight.

Access and safety

The access block at the top of every page is generated from a recorded access status, not written by hand. Where a site is closed to the public, or where its location is restricted, the page carries no coordinates and no directions, and the map pin falls back to the nearest town rather than the site. That rule is enforced by the code that builds the page, not by remembering to apply it.

Corrections

Corrections are appended. Once resolved, a correction is never edited or removed — the build refuses content whose resolved corrections no longer match the committed record of them. A later change to the same fact is a further correction, not a rewrite of the first. The full log is public.

Where AI is used, and where it is not

Drafts are assembled with AI from facts that a person has already confirmed against a named source. AI does not supply facts, choose sources, or decide what has been verified. Every figure on this site was read off the record it cites by a person.

Concretely: a model helps split a short piece of writing into separate claims, proposes search terms for an archive a person has already chosen, arranges already-confirmed facts into first-draft prose, and flags things for a person to re-check. It cannot fill in a quoted figure, set a claim’s status, supply a source or a link, decide whether a source is good enough, write the access and safety text, generate an omission, or touch a correction. Those are refused by the shape of the system, not by policy. Where a draft needed a fact nobody had confirmed, the gap is marked and the page cannot publish until a person resolves it.

Automatic checks

Three checks run before anything publishes, because these are the errors that have actually happened: a body or agency named alongside a date that precedes it; a superlative with no scope, which no record can settle; and a figure or date in the summary line at the top of a page that no claim on that page supports.