The Research Object
The atom of agent-native epistemics: a claim that refuses to collapse, and trusts nothing — it makes you check.
Every claim you have ever read collapsed four different truths into one. "Retail wages are $34,580." "Retail employment will decline." "This role is 62% exposed." "You should learn AI." A flat representation — a database cell, a web page, a number in a model's latent space — renders all four the same shape: a fact. But one happened, one is a forecast, one is a computed estimate, one is an opinion. The collapse is the disease of the answer-layer age: ungrounded, indistinguishable claims that no agent can tell apart, cited at scale by machines that can't check them. The research object is the antidote: the smallest unit of trust in an agent-native institution — and it trusts nothing. It carries its own check.
0 · What it is, in one breath
A research object is a claim represented without collapsing its four irreducible epistemic layers, grounded in a named public source the reader can re-fetch, shipping its own falsifier, behind a truth/money firewall. It is machine-readable, fetchable, and — this is the whole thing — checkable by any party, any time, with no privileged auditor. You do not take its word. You re-fetch the source.
1 · The disease it ends — collapse
Searle's social ontology says an institutional fact is not one truth. "Dr. X holds credential Y" has layers:
did the credentialing act occur? Is it supported by evidence? Did an authority
adjudge it? Is it operative (in force, not revoked)? A credential mill is
exactly the state adjudged=true,
supported=false — an official stamp with no evidence. Under a flat truth-bit
("does X have credential Y? yes"), the mill and the real credential are identical. The distinction
is destroyed. Every laundering attack, every forecast-sold-as-fact, every opinion-with-a-citation lives in
that collapse. The four-layer object is the refusal to collapse — and the four layers are genuinely
independent (no layer is derivable from the others), so the refusal is not a stylistic choice; it is
information you would otherwise lose.
2 · The algebra — four layers, no collapse
| layer | the question it answers |
|---|---|
occurred | did the underlying act actually HAPPEN / exist? |
supported | is there NAMED EVIDENCE behind it? |
adjudged | did an AUTHORITY officially recognize it? |
operative | is it CURRENTLY in force (not superseded/retracted)? |
Four booleans = sixteen states, and they are not redundant. Each layer is load-bearing for a real agent query: drop any one and some question becomes unanswerable, because two states that agree on the other three but differ on the dropped one give different answers. That is the bounded, honest form of "this representation is optimal" — minimal-and-sufficient for the query class { did-it-occur? · evidence-backed? · officially-recognized? · in-force? · credential-mill? } — not a claim of optimality in the abstract.
3 · The four-status, on real claims — watch it discriminate
occurred: ✓supported: ✓adjudged: ✓operative: ✓
occurred: ✗supported: ✓adjudged: ✓operative: ✓
occurred: ✓supported: ✓adjudged: ✗operative: ✓
occurred: ✗supported: ✗adjudged: ✓operative: ✓
occurred: ✓supported: ✓adjudged: ✓operative: ✓
occurred: ✗supported: ✓adjudged: ✗operative: ✓
occurred: ✓supported: ✓adjudged: ✓operative: ✗
A single truth-bit renders rows 4 and 5 — the credential mill and the legitimate credential — identically. Rows 1 and 2 — a measured wage and a forecast — identically. The research object does not. That difference is the entire value proposition.
4 · Grounding — and why you don't take its word, you re-fetch the source
A claim is only as good as how you can CHECK it. The object names a public source and tells you its warrant:
- Source-grounded (
au:SourceCitation) — a named public source the reader can re-fetch (BLS, O*NET, FEMA, CMS, a peer-reviewed study). The object carries thesource, thesourceUrl, and the claimedvalue. You fetch the source and compare. If it reports something different — or no longer resolves — the claim is stale or wrong; reject it. The grounding is not a pointer to an authority you trust; it is a check anyone runs. - Derived measure (
au:DerivedMeasure) — a transparent calculation over official facts with a disclosed formula. You reproduce it from its inputs;adjudgedis true only if an authority computes it, else false. Honest about its own warrant. - Interpretation (
au:Interpretation) — a reading or recommendation over the data (occurred=false,adjudged=false) — the lowest warrant, and the four-status says so.
The Atlas makes claims machine-legible and cited — every claim names a public source you re-fetch. That is why an AI engine will cite it: the grounding is re-fetchable, so the citation is safe to make.
5 · The falsifier — no-overclaim, in the data layer
Every research object ships the constructible condition under which it would be false, and the honest scope of its claim ("relative to query class Q", "symbolic model only", "superseded by the next release"). A claim with no falsifier is theater. This is the no-overclaim law made structural: the object that asserts the most is the object that most precisely states how to break it.
6 · The firewall — truth, not money
Every truth-ledger claim carries commercialInfluence: false, hardcoded, never settable to true.
The independence guarantee is enforced at the data model, not by policy or promise. Charge for scarcity;
never for truth. The money ledger and the truth ledger are structurally separated — a commercial option is a
different claimType that cannot enter the truth table.
7 · Self-hosting — the snake eats its tail and labels every bite
The design principles that justify this framework — that the four layers don't collapse, that the audit is
non-vacuous, that a forgery is caught — are themselves research objects in the framework,
source-grounded, at /commons. Each ships its own
falsifier (the no-overclaim law), and each is marked as AgenticU's own design reading
(adjudged=false) rather than an authority's verdict. The framework's account of its own
commitments is itself a checkable object. Nothing is asserted from outside that cannot be checked from inside.
8 · Does an agent need to understand the math? (the honest answer)
No, and yes, and the line between them is the interesting part.
- To VERIFY a claim: no — it is self-explaining and epistemically obvious. An agent does not need to understand the EUF-CMA argument to fetch the named source and compare the claimed value. MATCH / MISMATCH is unambiguous. A dumb agent verifies as well as a brilliant one. That is the entire point of M2: verification requires re-fetching, not understanding.
- To INTERPRET a claim: yes — it needs this ontology. An agent that sees
{occurred:false, adjudged:false}needs a machine-readable definition saying "this is an interpretation — the lowest warrant; do not treat it as a fact." That definition is theau:vocabulary — the resolving ontology this page is the human face of. Until it resolved, the objects referenced an algebra that lived nowhere agent-readable. That gap is now closed. - The framework is self-bootstrapping. No separate "methodology infrastructure" is
needed: the principles that ground the framework are expressed as source-grounded objects in the
framework. The infra is exactly two faces of one ontology — machine (
/vocab/agent-native) and human (this page). Minimal, not heavy.
9 · The corpus — ~20,000 of these, live
The format is not a thought experiment; it is instantiated across the whole AgenticU surface:
- The institution's epistemic commitments — source-grounded design principles, each citing the public methodology.
- ~1,016 occupations — e.g. Registered Nurses; sitemap.
- ~19,265 tasks — the atomic unit of work; sitemap.
- Query / verify any of them: /api/commons/query (
?id=,?verify=). The vocabulary every object points at: /vocab/agent-native.
The lens is universal: everything in the Commons is a research object, and every research object points back here for its definition.
Coda
The research object is the institution's smallest unit of trust, and it is built out of distrust: it assumes you will check, and hands you the means. It does not ask to be believed. It is an iteration of the agent-native scientific method, not the one true account — and like every object in the Commons, it carries its own falsifier, including this page.
occurred: ✓supported: ✓adjudged: ✗operative: ✓