How we verify

How we research

We use AI research tools. We are telling you that up front, because you would be right to want to know.

What we use: Claude, ChatGPT, and conventional search including Google. Different tools are better at different things. AI tools are good at reading long government documents quickly, finding where a claim might be documented, summarizing legislative text, and catching holes in our own reasoning. Search is better for finding the document in the first place.

What we do not do: treat any AI tool as a source of fact.

That distinction is the whole thing, so here it is plainly.

The two-stage rule

Until July 2026, nothing published on this site until it cleared both stages, and that sentence used to end there. It no longer can. The two-stage rule still governs everything a person writes: stories, positions, county narratives, and every characterization. A narrow class of mechanical claims now publishes on machine checks alone, and the section on machine-checked claims below says exactly what that means and what it does not.

Stage two is the verification. Stage one is research assistance. We do not treat these as equivalent and neither should you.

Why we work this way

AI language tools have a specific and well-documented failure mode: they produce citations that look correct and are not. A fabricated source reads exactly like a real one. It has a plausible title, a plausible author, and a plausible URL. The confidence of the output carries no information about whether the underlying document exists.

This failure is not rare and it is not going away. Any site using these tools without a human verification gate will eventually publish something that traces back to nothing.

We would rather publish less and be checkable.

Some of our pages publish their pending list openly. If you see a page with a visible verification checklist showing unchecked items, that is not an oversight. That is the process working where you can see it.

What publishes without a human read

Some claims on this site publish automatically. You should know which ones, and what that actually buys you.

A Tier A claim in our claims system is one sentence bound to a quoted passage of a named primary source, and it is always a mechanical restatement: a date on a filing, a figure as certified, the words of a holding. Tier B and Tier C claims can rest on broader sourcing, and every one of them waits for a person. When a Tier A claim passes every machine check, it publishes on the next build. No person reads it between extraction and publication. That is a change from the promise this page used to make, and we are stating it rather than letting you find it.

We changed this because checking every mechanical restatement by hand does not scale to 67 counties, and because the checks below catch a class of error that hand-checking reliably misses. A person clicking a citation confirms the link works. It does not confirm the number in the sentence matches the number in the source, and that is the error we actually kept making.

What the machine checks:

• The claim names its source: URL, title, publisher, and a verbatim quoted span. A claim missing any of these cannot exist in the system at all.
• Every number in the claim appears in the quoted span. A sentence saying 58 percent cannot publish against a source that says 55.9 percent, which is an error we once caught by hand before it published and now catch by machine.
• The quoted span appears in our stored copy of the source's full text.
• An archive link exists, so the citation survives the original moving or dying.
• For court opinions, statutes, and official reports, a recorded answer to whether the document is still in effect. A vacated or superseded document cannot be the source for any published claim, under any framing.

What the machine cannot check. Read this list as carefully as the one above, because it is the honest half:

• Whether the source is any good. A faithfully quoted span from a bad document passes every check we have.
• Whether a true, sourced sentence misleads by leaving something out. The checks confirm what is present, never what is missing.
• Whether the sentence means what the source means. Number binding catches a wrong figure; it cannot catch a wrong verb.
• Whether these were the right claims to publish in the first place.

Because of that second list, everything interpretive still waits for a person. Every characterization, every comparison, every causal claim, and every Returns Ledger stamp is held until a named person approves it, and the approval is recorded in the claim's own data. The machine gate raises the floor under our mechanical claims. It does not replace judgment, and this page is not claiming it does.

The IN REVIEW chip

IN REVIEW marks a claim that a machine extracted from a cited source and that no person has yet reviewed. The claim's text does not appear on the page. What you see is a marker naming what is under review, and nothing more, until a person clears it or rejects it.

This is not the same state as Pending review. Pending means a person knows about the item and sourcing is in progress. IN REVIEW means the extraction has happened and the human read has not. We keep the chips separate because merging them would let a machine's output wear a label that implies a person was involved.

How we use AI in writing

Research is not the only place these tools show up, and it would be a strange kind of honesty to disclose one and not the other.

AI tools assist in drafting and editing content on this site. Structure, phrasing, and first drafts are frequently AI-assisted. Every page's structure, prose, and arguments are reviewed, edited, and approved by a human before publishing, with one exception this page states plainly: machine-checked Tier A claims, covered above, publish without a per-claim human read. The arguments, positions, and editorial judgments are ours.

We mention this because a site built on "check our sources" should not be selective about what it discloses.

When this process fails

It will. When it does:

• The correction goes on our public corrections log with the date and what changed
• If the error came from an unverified citation, we say so specifically
• We do not quietly edit

We have already rejected claims during research that turned out to be wrong, including several we initially believed. Those are listed in the rejected-claims table on our sources page. We publish them because a clean record would tell you less than an honest one.

What you can do

Check us. Every factual claim on this site should have a link. Open it. If the source does not say what we say it says, tell us and we will fix it.

That is not a rhetorical invitation. It is the only reason the process means anything.