How these figures are verified
No figure on this site is typed in by hand and no figure is calculated. Each one is read out of a government document by an automated process whose every numeric claim is checked by code before the page is written. What follows is what that process does, and what it does not do.
The steps
- A candidate document url is proposed and then checked against a list of authoritative government domains held in the repository. A url outside that list is discarded and the run stops. A new domain has to be added to the list in its own commit.
- The document is fetched and its text stored unchanged. That stored text is evidence rather than a record: it is never edited to fit a figure, and replacing it is a deliberate, separately named operation rather than something a routine run does.
- The figures are read out of the stored text. Each one arrives as a key, a number, and a quote copied from the document.
- Every figure is then checked by code with no language model involved. The quote must occur in the stored text exactly once and at a known position. The number must appear inside that quote as a standalone amount with its currency or percent marker, not as part of a larger number, a date, a section number or a form number. The quote must state the rule rather than illustrate it, and must be about the one figure it is offered as evidence for. If it names any years at all, one of them has to be the year being published or the year either side of it, because a document published in one year routinely sets the next year's amounts, and a passage tracing a scheduled increase may name the year the schedule began as well. A quote that names no year is accepted: documents often state a rule in one place and date it in another. Every figure the program declares for that year must be present. If any check fails, nothing is written.
- The short passages of text on each page are written last, and every number in them is scanned. A number is allowed only if it is one of the figures just verified, one of the previous year's verified figures, or a year within one of the year being published. Appearing somewhere in the source document is deliberately not enough - that allowance is what once let a section number and a filing date be published as amounts. A number that is genuinely cited rather than stated as an amount, such as a Code section or a form number, is exempted by the words around it instead.
- Before publication the checks run again against what is about to be written. A page is only published if they pass a second time.
What this does not tell you
The checks compare a number against a sentence. They do not read the sentence. A figure that is correctly quoted but used to answer the wrong question is a mistake this process cannot catch, and neither is a change in the law that no document on file mentions yet.
This site is not reviewed or endorsed by a tax professional, and nothing here is advice. The figures are offered with the document they came from so that you can check them yourself, which is the only assurance this site claims. If a number disagrees with the document linked beside it, the document is right.
Permanence
A published figure never changes. Once a number has been verified and dated it is a matter of historical record, and the build refuses a commit that alters one unless the change is made deliberately and on its own. Past year pages are never deleted and never redirected. How a mistake is corrected is described under corrections.