Everything the Library publishes rests on sources: court opinions, statutes, state reports, archives, scholarship. A claim a reader cannot trace is a claim a reader has to take on faith, and this Library was founded on the opposite idea — that public trusts failed for two centuries precisely because nobody could check the books. So we hold our own books to the standard we ask of trustees. This page explains how.
One register, every book
Every authority we rely on — every case, statute, report, and archival
document — lives in a single register that records what the source is,
where the original can be found, and one more thing most bibliographies
omit: whether we have actually verified it yet. Each
entry is marked. Either we hold the primary document and have checked
our quotation and page citation against it, or we are carrying the
citation from a reliable intermediate source and say so, or the item is
still awaiting verification and is flagged. Those flags print. When a
page of the hornbook says [VERIFY], that is the register
showing through — a promise still being run to ground, visible to you
at the same moment it is visible to us.
Harvest, verify, correct
The method has three motions, and they run continuously.
Harvest. We pull primary documents into the Library's own collection — full court opinions, session laws, agency reports — rather than relying on summaries or secondhand quotations. A scheduled task works through the queue of still-unverified authorities week after week.
Verify. With the primary in hand, we compare. Does the opinion actually contain the sentence everyone quotes from it? Is the page number right? Does the case actually hold what the brief said it held? Is the fund figure the audited value or somebody's projection? Every comparison is logged.
Correct. When verification finds a problem, the text is fixed — and the fix is recorded, not buried. Where an error had already reached a published page, the page says what was wrong and what changed. Institutions that conceal their corrections drift; institutions that publish them can be trusted with the next claim.
What the method caught this week
A verification pass over the Library's books and the hornbook ran this week. A few of its findings show why the work is worth doing.
A famous sentence that does not exist. A line about trust duties — "they must be performed; they admit of no substitution; they admit of no addition" — circulates widely in school-trust writing, attributed to a 1919 United States Supreme Court decision, Ervien v. United States. We read the opinion. The sentence is not in it. The opinion's genuine language — duties "special and exact," precluding "any supplementary or aiding sense" — is strong enough on its own, and that is what our books now quote. The phantom sentence has been removed everywhere we had repeated it.
A fund figure that was really a forecast. One state's school trust fund was carried in a draft at $8.4 billion. Verification traced that number to a projection of future mineral royalties — not the fund's actual reported value, which was $7.94 billion as of December 2025. The books now carry the audited figure, with its as-of date.
A statistic we refused to print. A draft asserted that a specific number of states publish no complete accounting of their school trusts. It is a vivid number. It is also unsupported — no national survey of the question exists, and the figure traced back to an AI system's confident guess that had then been quoted as if it were a finding. The claim now states the documented reality without the invented precision: the absence of consolidated reporting is itself the finding.
A rule against quoting ourselves. The audit also enforced a discipline that sounds obvious and is surprisingly easy to violate: our own books, model legislation, and memos are never cited as evidence for anything. A project that quotes itself in a circle can prove anything. Where a draft chapter had leaned on our own commentary, it now rests on the underlying cases and statutes, or says plainly that it is describing our proposal.
Where AI fits, honestly
AI systems do much of this work — the reading at scale, the comparing of quotations against opinions, the bookkeeping of hundreds of register entries. AI is also one of the reasons the work is necessary: as the phantom statistic shows, machine confidence is not evidence. The arrangement is the same one this Library recommends for every perpetual institution: let the machines do the exhaustive checking, let the rules be written down and public, and let humans set the standard and review the result. Errors found are logged and mined for patterns, so the process improves the way endowments are supposed to grow — compounding.
Help us
The flags you see in our drafts are open invitations. If you can supply a primary source, correct a citation, or catch an error we missed, write to the Reference Desk. Corrections are verified against primary documents and carried into the next edition — with the change noted, in the open, the way it should be.
