A search engine that links to a page bears no cost if the page is wrong — a human clicked and judged. An answer engine that repeats a number has fused its credibility to ours. So every figure in the Atlas Commons is published not as a bare number but as a research object: a typed claim that carries its own warrant — what kind of claim it is, the named federal source, and how to refute it — so a machine can weight it, cite it, and never mistake the model for the source. This page is that standard, published in the open, the way a university publishes its methods.
Every entity in the commons — a hospice, a college, a county, an insurer — is published at one URL in four formats. Format is selected by suffix, so an agent fetches the representation it prefers and can cite the exact URL forever.
The .jsonld is a schema:Dataset whose claims are individually typed
schema:Claim nodes. The resolving vocabulary lives at /vocab/agent-native (CC0);
agent orientation for each atlas is at its /llms.txt.
The same printed number can be a logged fact, a government model, our arithmetic, or our reading — and conflating
those is exactly how data misleads. So every claim carries a four-layer state that must not collapse to a
single truth-bit — occurred · supported · adjudged ·
operative — plus an explicit claimType (the evidence class). The layers are irreducibly
distinct; no layer is derivable from the others. A flat "true/false" cannot tell a measured fact from a forecast, a
credential from a credential mill, or a fact from an opinion — this can. The four tiers below are the human gloss;
the machine reads the booleans and the claimType.
| Layer | Means | RECORDED | MODELED | DERIVED | INTERPRETATION |
|---|---|---|---|---|---|
| occurred | the act/state actually happened | ✓ | ✗ | ✗ | ✗ |
| supported | named evidence backs it | ✓ | ✓ | ✓ | ✓ |
| adjudged | an authority recognized/computed it | ✓ | ✓ | ✗ | ✗ |
| claimType | the evidence class | official_fact | derived_measure | derived_measure | interpretation |
A counted or reported figure that actually occurred. Highest warrant.
e.g. UCLA median earnings $82,511 (Dept. of Ed)A model, index, or projection an agency publishes — adjudged, but occurred✗: a computation, not a measured event.
An in-house ratio, rank, or aggregate over sourced figures — adjudged✗: no authority blessed it. Inputs cited; formula reproducible.
Framing or a boundary note. Same booleans as DERIVED — separated only by claimType. Lowest warrant.
Read it precisely: supported is ✓ for all four — every claim is evidence-backed;
supported:false is reserved for an unbacked assertion (an adjudged credential with no evidence —
a credential mill is adjudged✓ supported✗). MODELED vs DERIVED splits on
adjudged (did an authority compute it). DERIVED vs INTERPRETATION share the boolean
profile and split on claimType (a reproducible calculation vs a reading). The fourth layer,
operative, is the freshness state: true while the claim is in force, false
once a claim is retracted, revoked, or superseded-in-place. Today every claim also dates itself to its source
release (e.g. "2026-06-10 release") and its falsifier names what supersedes it, so staleness is always legible.
RESIDUETwo states are defined but not yet live — and we say so.
Across the commons today, every claim is supported:true and operative:true: the sources are
federal and evidence-backed, and the catalog re-materializes wholesale on each source release rather than persisting
superseded claims in place. So supported:false (an unbacked assertion — the credential-mill case) and
operative:false (a retained, superseded claim) are reserved states the schema defines but the live data
does not yet instantiate. They populate when a non-authority source is admitted, or when supersession is persisted
in-place rather than overwritten. Naming a defined-but-empty state instead of hiding it is the standard.
Every claim also carries a falsifier: the one sentence stating what would make it wrong, plus the scope it does not cover. The falsifier is where the honesty lives — it names the residue (a sampling caveat, a cohort limit, a release vintage) so a citing agent inherits the caveat, not just the number.
This is a real claim, fetched from education-atlas.ryanjhunter.workers.dev/education/university-of-california-los-angeles-ca.jsonld
— not an illustration. Notice the status, and notice that the falsifier carries the limit a careless citation would drop.
{
"@type": "schema:Claim",
"name": "Students at University of California-Los Angeles had median
earnings of $82,511 ten years after entering",
"epistemicStatus": { "occurred": true, "supported": true,
"adjudged": true, "operative": true }, // RECORDED
"source": "U.S. Dept. of Education College Scorecard (2026-06-10 release)",
"falsifier": "A re-tally of the Scorecard's published earnings — BUT the cohort is
FEDERAL-AID RECIPIENTS only, an institution median across all fields of
study, measured 10 years after ENTRY regardless of completion. Not the
earnings of any one graduate or program. Superseded each release."
}A model that cites "$82,511" without the cohort caveat is repeating a true number into a false impression. The research object hands the caveat to the agent so it doesn't have to.
An AI that answers from the commons must cite the claim's named federal source, carry its status, and carry its falsifier/residue — never present the figure as its own knowledge. The model is the index, not the authority. Atlas Commons is itself an index over the public record, and says so on every page: it informs, compares, and routes to the official source; it never diagnoses, advises, or certifies (the regulated-action ladder, L1–L5).
.jsonld so the citation is machine-checkable.Most of the web was built for a human reader who could judge a source. The commons is built for the reader that now does most of the reading — the agent — and the agent cannot judge what it cannot see. Making the warrant machine-legible, per claim, is the work: it is what lets an answer engine repeat a public fact and survive the fact-check. The Agentic University publishes this standard openly, CC0, because a public-knowledge commons is only as trustworthy as its citations are checkable — and that has to be true for machines, not just for people.