Where this sits: Home › About › Methodology. For how to use the site with a person, see the guide for AI assistants.
This page describes how the site's content is produced, so you can judge how much weight to give it.
1. Collecting sources
Official sources (Minnesota Statutes and Rules, federal law and regulations, agency policy manuals, bulletins and program pages) are fetched one page at a time, not by bulk crawling. Each file is kept exactly as served, or as the browser rendered it where that is all a site allows (labelled as such). It carries a fetch record: requested and final address, HTTP status, method, the retrieval time recorded when fetched, a SHA-256 fingerprint, and any dates the page itself shows. When a live page can't be reached, an Internet Archive copy may be used, labelled with its capture date. CAPTCHAs are never solved or bypassed by the system.
Coverage is checked by cross-walking programs across several independent catalogs (state agency directories, 211, House Research, the state budget, federal assistance listings, county program lists) and recording how deeply each program is documented. Primary sources cited by other benefits projects are gathered into one list and fetched.
2. The archive
Each substantive document becomes an archive entry: clean text anchored by section, subsection or program block, carrying its original fetch record. A ledger lists every source considered as absorbed, pending, or discarded with a reason (duplicate, not substantive, blocked, failed). The archive is evidence, not published content.
3. Units of knowledge
Knowledge is organized by subject, not by source document. A unit covers one subject, across programs where that helps (for example, how child support counts as income). Each unit lists its evidence: the archive entry and anchor, the source's layer (federal law or regulation, state statute, rule, manual, bulletin, public page), how the passage relates to the claim (states, implements, restates, conflicts, superseded), and effective dates. Units have stable identifiers and can sit under more than one parent, such as a program and a cross-program concept.
Realistic questions are kept as test questions linked to the units that answer them. A question the units can't answer is recorded as a gap.
4. Review
A reviewer agent that did not write a unit, started fresh, checks it claim by claim against the cited passages and marks each claim verified, corrected, unsupported or conflicting. A second AI model, from a different vendor, answers test questions from the archive alone without seeing the units; a fresh reviewer settles the differences by reading the sources. Reviews are recorded on the unit with a date. Conflicts between sources are logged for reporting to the publisher. Legal readings that the sources can't settle are marked as unsettled on the page.
5. Pages
Each page is defined by a specification: the pieces of knowledge it must convey, by reference to units, and how fully. The text is written to that list and checked against it, so every item is present and nothing is added beyond it. Pages link only to other pages of this site, except to things a person acts on directly (applications, forms, phone lines). The site is built from these definitions into plain HTML by the hosting service.
6. Changes
Every citation records the fingerprint of the copy it was read from. When a source is fetched again and has changed, the units citing it are found and reviewed, and pages built from them are rebuilt. There is no fixed re-fetch schedule yet, and pages don't yet show a last-checked date.
What this means for you as a reader
Treat pages as careful general explanations, not determinations. They may be out of date, and they are not legal advice; the county or agency decides any individual case.
Where a page marks something as unsettled, conflicting or not covered, pass that on to the person. Companion pages listing each page's sources are planned but not built yet.