How we verify

Every claim on this site is traceable to a numbered section of adopted ordinance text that we archived and can show you. This page describes how that works, and — more usefully — where it stops.

1. We archive the ordinance text itself

We do not work from a summary of the code. We fetch the adopted text, store it, and record where it came from and when. 6 archived artifacts back this site today, covering 36 sections.

Every artifact records its own provenance. For ordinance chapters that means the identifiers needed to re-fetch exactly the same text; for anything else, the URL it was taken from. All of them record the date retrieved.

2. Every claim names its section

A page is not prose with citations added afterwards. Underneath each page is a list of claims, and each one carries the section it rests on and the archived artifact it was read from. That list is published — see the open dataset — so the structure is checkable rather than asserted.

3. Machine checks, which annotate rather than block

Three run over every page:

These annotate. They never withhold a page. That is a deliberate reversal of how this project first worked: an earlier version gave a machine a veto over publication and shipped nothing for a day. Findings now go in front of a person, and a person decides.

4. Independent review by models that did not write the page

Pages are read back against the archived text by external models — currently OpenAI’s and a separate Claude process with no memory of having written the page. They are asked one question: does any passage overstate or misstate the section it cites?

They find real errors. On the first four pages put through this, thirty-three objections were accepted and corrected, several of them on pages that had already passed every machine check cleanly. Some objections we reject, and the page records that we did and why.

5. Nothing publishes without a person approving it

Editing a page changes nothing a reader sees. The live site serves an approved snapshot; an edit sits in a review queue until a human promotes it. The corollary matters as much: a page already published never comes down automatically — not for a failed check, not for a typo. Withholding is safe only until something is live.

What this cannot tell you

Three limits, stated plainly, because a method that only advertises its strengths is marketing.

We describe the text, not the practice. Codes are administered by people, and how a department reads a marginal provision is not something we can source. Where a question turns on that, we say so and tell you to ask them.

We can be out of date. Each page carries the date we last checked it. An ordinance adopted after that date will not be reflected until we re-read it.

Attribution is not comprehension. Our checks confirm a figure appears in the section we cite. They cannot confirm we understood the provision correctly — that is what the independent review and a human reader are for, and neither is infallible.

When we are wrong

Every page names its sections, so a correction can be precise. Tell us the page and the section, and we will check it against the archived text. Corrections are recorded on the page itself, with what was wrong and what changed — you can see that history on pages that have been through it.