How this site is built

Who builds it, and who answers for it

The American Factbook is written and maintained by Claude, Anthropic’s AI model, working in Claude Code. It writes the code that fetches each dataset, the queries, the pages and the checks that stand between them and a reader. It works under a written rule set that the editor keeps, and the editor is accountable for what is published: About names him. That rule set is internal, because it also covers the site’s budgets and tooling. The rules a reader relies on are the ones on this page.

Where a number comes from

Every figure is the result of a checked-in query against a published government dataset, dated and linked to its source.

Each figure here comes from a query run against our copy of a public dataset. The query is kept beside the page that uses it, and the page links to it. Sentences are written around values the query returned. No figure is typed in, estimated or remembered.

Each dataset is fetched from its publisher and stored as it arrived, dated, before anything is done to it, and that stored copy is never deleted. Every page shows the date its data was retrieved and links to the publisher.

How a figure is checked

Before a change reaches the site, every query behind a page with a fixed set of figures is run again, and every number in the page’s text is compared with what came back. A difference stops the change. After every deploy, pages built from the database on request are sampled, and each figure on them is compared with a direct query.

A check is only as good as what it is compared against. Where a publisher prints its own figure for something this site computes, such as a crime rate or a total, the two are compared, and a difference beyond a stated tolerance fails the build rather than being printed. Where no independent figure exists, the methodology says so.

Every day an automated check asks whether each source still holds the publisher’s latest release, and names any that has fallen behind.

What a page may say

• What the data shows, and where it cannot answer a question, that it cannot. A page does not answer anyway.
• No cause drawn from a correlation. Two figures that moved together are reported as that, and nothing more.
• A rate or a share names what it is divided by. A figure from a period that has not finished is not published as though it had.
• The same treatment for every topic, in plain language, without adjectives that argue.
• No suggestion of affiliation with any government or agency. The site is independent, and says so.

People in the record

Some records name people: doctors paid by drug and device makers, farm payment and loan recipients, immigration judges, service members who died. A page about a person reports the official record as the agency published it, attributes it, links the source, and draws no inference beyond it. It carries a city, state and ZIP code, never a street address. The pages of those who died in service carry the record alone.

If a publisher later stops releasing part of a record, this site stops publishing that part too, even from a copy it already holds, and even where that copy’s license would allow it.

What is not published

A page with nothing on it but a name and a template is worse than no page, so a page whose record is too thin is kept out of search, or not made. A dataset is used only after its publisher’s terms have been read and recorded, and data whose terms forbid commercial reuse is not used.

When something is wrong

It is corrected in place and logged, dated, on Corrections, and the entry is never removed. Write to kian@americanfactbook.org. How each figure is computed is on Methodology, and every dataset, with its terms, is on Data.