Methodology

Every requirement record cites at least one primary source: an issuing agency's own page, its published fee schedule, a municipal or state code section, or an official application form. Secondary sources such as law firm articles may add context, but they never stand alone as the basis for a published fact.

We store more than a link. Each source is captured with its URL, publisher, page title, the date it was retrieved, and an archived copy — because government sites reorganise constantly and a bare link rots.

Every record carries one of two published states. A verified record has been confirmed against its primary source by a named reviewer within its review window. A record marked as needing review has passed that window, or our change detection has flagged its source — it is still shown, but it is labelled, never presented as current.

Fees and renewal terms are re-checked every 90 days. Application methods and forms every 120. Statutory references and agency identity annually. Anything on a high-traffic page is checked every 60 days regardless.

We re-crawl every stored source weekly and compare a hash of the specific region that matters — the fee table, the term statement — rather than the whole page, so a navigation redesign does not generate a false alarm. When something moves, dependent records are flagged for review and the change is published in the change log with the before and the after.

A language model may draft prose, extract candidate values from a source, or summarise a document. It is never the origin of a published fact. Every extracted value is confirmed by a human against the primary source before publication, and each record logs whether extraction was machine-assisted.