Editorial Standards and AI Disclosure

Last updated: August 2026

LegislationPatch uses AI to read and draft. This page says exactly where, exactly how it is constrained, and exactly what a human does before anything is published. It is written to be checkable rather than reassuring — the specific gates described below either exist in the code or they do not, and the repository is public.

The one rule everything else serves

Nothing published on this site may come from an AI model's background knowledge. Every figure, date, dollar amount, statutory citation, and section reference must appear in a source document that was actually fetched and stored — the bill's official text, a referenced statute, a roll-call record, or the Congressional Record. A model's recollection of what a bill says is not a source, even when it happens to be right. A claim that cannot be tied to a stored source line is deleted, not softened.

What the automated pipeline does

• Fetches the bill's official text from the Congress.gov API — the latest applicable version at fetch time, because a bill that has passed a chamber is often substantially rewritten from the version introduced. GovInfo supplies the Congressional Record and any statute or chamber manual an analysis rests on. A bill that advances after it has been analysed is not automatically re-fetched, so the stored text can be older than the newest version posted.
• Extracts a facts sheet: every figure, percentage, deadline, date, and heading in the text, with line numbers. Figures in the published analysis are copied from that sheet, not re-typed from memory.
• Fetches referenced statutes when a bill amends or extends another law, so the numbers come from the actual referenced text.
• Drafts the analysis or article against that source material under a fixed set of sourcing rules.
• Runs 27 deterministic check sections before a batch can ship. Fourteen of them block outright — among them that no analysis exists without the bill text it was written from, that every cited source resolves to a stored document, that no attributed quote is unverifiable, and that a stated legislative stage does not contradict a failed roll call. The other thirteen warn rather than block, including the figure-sourcing guard that flags a dollar amount, percentage, or section citation absent from the fetched text. A warning is not a pass: each one must be fixed, re-sourced, or recorded in a tracked adjudication file with the evidence, and a --strict run turns every remaining warning into a blocking error.
• Re-reads the draft against the source in fresh sessions that carry no memory of previous passes, working adversarially — assuming the draft is wrong and treating absence from the source as a flag.
• Decomposes the draft into individual factual claims and judges each one supported, contradicted, or unsupported against a quoted span of source text, including a check for direction-of-change errors such as reading a limit as an increase.
• Records on each audit ledger which model performed the verification pass, so a reader can see who checked what. For standalone explainer articles this is enforced rather than asserted: the publishing script refuses to publish or refresh an article whose ledger names the same model as drafter and auditor. Bill analyses record the auditing model but do not yet record the drafting one, so the same check cannot be applied to them.
• Records the result as a per-page audit ledger: every claim checked, the verdict, and the source line behind it. The ledger covers 212 of the 219 bill analyses in the database; the seven added most recently do not carry one yet.

What a human does

Every analysis and article is reviewed, verified, and edited by James Shearn, editor, before it is published. The editor adjudicates every flag the automated checks raise — confirming it, fixing it, or recording in a tracked file why it is not an error and what evidence supports that. Nothing publishes on a machine's say-so alone, and the decision to publish is always a human one.

The editor is accountable for everything on the site, including anything the pipeline got wrong.

What is never automated

• Quotations are never generated, paraphrased, or reconstructed. Every quoted statement is reproduced verbatim from the Congressional Record and attributed to the member who said it, or it does not appear. Provenance beyond the speaker varies by store: the standalone floor-quote collection records the GovInfo granule the statement came from, while a bill-attached quote records the chamber and the bill it was pulled for.
• Positions are never invented. Where the site describes what a supporter or an opponent argues, that position must trace to a named holder and a real, retrieved document — the organization's own statement or letter, a recorded vote, a Statement of Administration Policy, or a member's own words. A position that cannot be verified is omitted rather than hedged.
• No AI-generated or AI-altered imagery is used to depict people, events, or documents.
• No editorial labels. The site does not call legislation landmark, radical, or common-sense, and does not endorse or oppose any bill. Significance is conveyed through structural facts — sponsorship, stage, recorded vote outcomes — and never through adjectives.
• No invented balance. Manufacturing an opposing side where none is on the record is treated as a neutrality violation equal to taking a side. A bill that passed 414–2 is described as having passed 414–2.

Authorship and bylines

Articles and analyses carry the byline of the editor who reviewed and verified them, together with a note that the draft originated from the research pipeline. This is deliberate: the site does not publish under invented staff-writer names, and it does not present machine drafting as unassisted human writing. Where AI materially shaped a published page, that is disclosed on the page.

What the process does not guarantee

These checks are strongest on numbers and citations, which can be matched mechanically against source text. They are weaker on subtler failures — a claim that is technically sourced but attached to the wrong account, an omission that changes the picture, or an unsourced qualifier that slipped through. Those are the errors the adversarial re-read exists to catch, and the reason corrections are logged publicly rather than quietly edited.

The site describes what bills say. It does not predict what Congress will do, and where the likelihood of a bill advancing is discussed it is framed structurally rather than as a forecast.

Corrections

Material factual errors are labelled on the affected page and logged publicly with the date, the page affected, and what is wrong. While our accuracy checks are being rebuilt (from October 2026), the text is corrected in our next verified pass rather than immediately. Routine updates that reflect new legislative activity are normal maintenance, not corrections. To report an error, use the public issue tracker; the full policy is on the corrections page.

Independence

LegislationPatch has no political affiliation, no outside funding, no advertising, and no sponsored content. It is not affiliated with any government body. See About for ownership and funding.

Related: Methodology · Neutrality policy · Corrections · Editor