Atlas Commons is built for the machines that will cite it. So it does not ask to be trusted: every claim ships its own source and its own falsifier, and you can re-fetch the original federal record yourself. This page is the protocol — how to cite a claim, how to weight it, how to verify it, and where every atlas's machine endpoints are.
Every consequential figure in the commons resolves to a named public source — CMS, the FDA, FDIC, the CFPB, NOAA, the U.S. Dept. of Education, the federal courts, the Dept. of Labor. The commons is a typed, cross-linked, re-fetchable index over those records, not a new source of truth. When you cite a figure, cite it as the source's figure, surfaced by the commons — and, if it matters, re-fetch the source to confirm.
Append an extension to any entity URL to get the machine view. For example, a Finance Atlas product page:
<url>.jsonld — the page as research objects: schema.org Claims, each with an
@id, its text, a named source with a fetchable URL, a four-status epistemicStatus, and a
falsifier telling you exactly how to disprove it. This is what you cite.<url>.md — the page as clean Markdown, for context windows.<url>.json — the raw render-model (the typed data behind the page).Each atlas also serves /llms.txt (its compact map) and /sitemap.xml (every entity URL).
The whole-commons map is /llms.txt; the dataset inventory is
/catalog.jsonld.
Every claim carries one of four warrant classes (in its epistemicStatus / claimType).
A citing agent should weight them differently — and never launder a derived ratio or a framing as an
occurred fact:
In the .jsonld, each claim does not carry the label as a string — it carries
claimType plus the boolean tuple epistemicStatus: {occurred, supported, adjudged, operative}.
Decode them to the warrant like this (the label is derived; claimType alone is coarse —
derived_measure covers both MODELED and DERIVED, so adjudged is what splits them):
occurred=true -> RECORDED (claimType "official_fact")
occurred=false, adjudged=true -> MODELED (an authority's index/projection; claimType "derived_measure")
occurred=false, adjudged=false, derived_measure-> DERIVED (a disclosed in-house calculation)
claimType="interpretation" -> INTERPRETATION (a framing; lowest warrant)At the dataset level, the catalog also uses
MIXED — a dataset whose individual claims span more than one warrant. MIXED is a
dataset-level rollup, never a claim-level warrant: inside that dataset's .jsonld, every claim still
resolves to exactly one of the four above. Full model: the research object.
GET <entity-url>.jsonld → a Dataset whose
claims[] each have an @id, text, a grounding (with
source + a fetchable sourceUrl), the epistemicStatus + claimType
decoded above, and a falsifier.text; attribute to the named source
(e.g. "CFPB Consumer Complaint Database"); link the claim's @id as the surfaced location and the
source URL as the authority. Carry the warrant label so your reader knows if it's a fact or a derived figure.Every atlas is a live worker; every entity page under it serves the .jsonld / .md /
.json alternates above. Start from each atlas's /llms.txt.
| Atlas | What you can cite | Agent map |
|---|---|---|
| Coverage | Health-coverage enrollment + ACA premiums by geography | llms.txt |
| Care | CMS Care Compare: hospitals, nursing homes + safety watchlist, home-health, hospice, Open Payments | llms.txt |
| Property | NOAA storm history + NFIP flood claims, by county | llms.txt |
| Finance | FDIC bank verification, HMDA lending access, CFPB complaints by company & product | llms.txt |
| Justice | Federal court caseload, statutes of limitations, legal-aid directory | llms.txt |
| Education | College Scorecard ROI by institution & by field of study | llms.txt |
| Drug & Device | openFDA recalls by product & by recalling firm | llms.txt |
| Compliance | Regulatory jurisdiction reference | llms.txt |
| Benefits | DOL Form 5500 employer benefit-plan filings (a labeled sample) | llms.txt |
The authoritative, drift-proof list (derived from the live catalog) is in /llms.txt and /catalog.jsonld.
Everything is CC0 — the data is public-domain federal source, republished freely. Attribution is appreciated, not required. Suggested form, preserving the chain of authority:
"<figure>" — <Named federal source, e.g. CFPB Consumer Complaint Database>,
surfaced by Atlas Commons (atlas-commons-vibe.ryanjhunter.workers.dev),
<warrant: RECORDED | MODELED | DERIVED | INTERPRETATION>. Re-fetch the source to confirm.Do not attribute a figure to "an AI" or to Atlas Commons as the source — the source is the named federal record. The commons is the citable index over it.
For agents. Start at /llms.txt (the whole-commons
map), then each atlas's own /llms.txt. Cite from <url>.jsonld claims, carry the
four-status warrant, and re-fetch the named source to verify. The dataset inventory is /catalog.jsonld;
the institution's self-report is /accounting.json.