Skip to content
AgenticU

The institution's epistemic commitments

Agent-native research objects: every claim carries its four epistemic layers and names a public source you can re-fetch. You check the source — you don't take the page's word.

The Atlas Commons makes the labor record machine-legible and cited — every claim resolves to a public source you can re-fetch. This surface states the institution's own epistemic commitments: the design principles that govern how every Commons claim is represented. Each research object below carries (1) its four epistemic layers (occurred / supported / adjudged / operative — not a single truth-bit), and (2) a grounding that is a named public source you re-fetch — the methodology anatomy, or the public dataset the principle operates over. These are AgenticU's design readings, not an authority's verdicts; the four-status (occurred=false, adjudged=false) says so on its face.

Machine surfaces: the object graph (JSON-LD) · the agent context pack · the query API (?id= a claim, ?verify= to get the re-fetch recipe). Verify any claim: follow its sourceUrl and compare the cited value against the public source. If the source reports something different — or no longer resolves — the object is stale or wrong; reject it. That is the whole point: the claim carries its own re-fetchable check.

The corpus — ~20,000 research objects

New here? Read the exhaustive anatomy of the research object (what the four-status means, how to verify, why it's better). The machine vocabulary every object points at: /vocab/agent-native. Then traverse:

  • The institution's epistemic commitments — source-grounded design principles, each citing the public methodology. object graph
  • The role matrix — ~1,016 occupations (source-grounded, four-status varies). sitemap · e.g. Registered Nurses
  • The task matrix — ~19,265 tasks (the atomic unit of work). sitemap · e.g. task-1
  • Query / verify any claim: /api/commons/query (?id=, ?verify=)

A real Atlas object — where the four-status earns its keep

Here the four-status discriminates, on a real Atlas occupation (Retail Salespersons, SOC 41-2031.00). Every claim names a public source you can re-fetch — an official BLS fact is {T,T,T,T}; a BLS projection is occurred=false; a derived composite is adjudged=false; the crossing recommendation is occurred=false AND adjudged=false. The same agent-native research-object format as the rest of the Atlas Commons. Object graph: /commons/roles/role-41-2031-00.jsonld.

Median annual wage for Retail Salespersons is $34,580 (BLS OEWS, May 2024) (official_fact)

occurred: ✓ supported: ✓ adjudged: ✓ operative: ✓

Grounding: SourceCitation — BLS OEWS ($34,580 median; 3,800,250 employed)

Falsifier: Superseded when BLS publishes the next OEWS release (then reviewStatus -> superseded, operative -> false). A verbatim official fact: transformation_id = null.

Projected employment for Retail Salespersons is -0.5% growth, 2024-2034 (BLS EP) — a FORECAST (derived_measure)

occurred: ✗ supported: ✓ adjudged: ✓ operative: ✓

Grounding: SourceCitation — BLS Employment Projections (-0.5% growth; 555.8k annual openings)

Falsifier: occurred = FALSE because the projected 2034 state has NOT occurred — it is a forecast. The four-status here DISCRIMINATES a projection from an occurred fact, which a single truth-bit cannot. Falsified if 2034 actuals diverge from the projection.

AI task-overlap index for Retail Salespersons is the 62nd percentile (AgenticU composite) (derived_measure)

occurred: ✓ supported: ✓ adjudged: ✗ operative: ✓

Grounding: DerivedMeasure — ai-exposure-index-v1 (composite over Felten AIOE, Eloundou gamma, Microsoft applicability) (62nd percentile (Moderate))

Falsifier: adjudged = FALSE: this is AgenticU's transparent composite, NOT a verdict by an official authority (BLS/O*NET) — supported by a disclosed reproducible formula, but not adjudged. The four-status DISCRIMINATES a derived composite from an official fact. Falsified if the disclosed formula does not reproduce the figure from its inputs.

AI Occupational Exposure (Felten et al.) for Retail Salespersons is the 50th percentile (official_fact)

occurred: ✓ supported: ✓ adjudged: ✓ operative: ✓

Grounding: SourceCitation — Felten, Raj & Seamans (AIOE) (50th percentile)

Falsifier: adjudged = TRUE because it is peer-reviewed academic work (an authority), unlike AgenticU's own composite. Falsified if the AIOE dataset reports a different percentile for SOC 41-2031.

Retail Salespersons should cross toward human-irreducible capability (AgenticU interpretation) (interpretation)

occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding: Interpretation — AgenticU crossing model (over O*NET tasks + the exposure indices above) (cross to grounded-trust / service-relationship capability)

Falsifier: occurred = FALSE (it is a recommendation, not an event) and adjudged = FALSE (AgenticU's view, not an authority's verdict) — the LOWEST-warrant claim in the set, and the four-status says so explicitly. A flat representation would render this identically to the BLS official fact; the Commons does not.

The institution's epistemic commitments

Claims about social reality have four irreducible epistemic layers

An institutional fact (a credential, a certification, an authorization) does not have one truth value. It has four independent layers — occurred (the act physically happened), supported (evidence backs it), adjudged (an authority recognized it), operative (it is currently in force). The four layers are pairwise independent: for each pair of layers there is a real state where they disagree, so no layer is derivable from the others. The intellectual source is Searle's social ontology (institutional facts are multi-dimensional status functions); the specific four-layer decomposition is AgenticU's operationalization of it, read as a representation requirement — a design reading, not Searle's verdict.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU operationalization of Searle, The Construction of Social Reality (Free Press, 1995) — in-house design reading  ·  Searle grounds that institutional facts are multi-dimensional status functions; the specific occurred/supported/adjudged/operative decomposition is AgenticU's, with occurrence, evidence, authority, and standing as separable conditions.

Falsifier: If any layer were derivable from the others, you could exhibit the rule and drop a layer with no loss. Re-fetch the anatomy at /commons/the-research-object; the credential-mill case (adjudged=true, supported=false) refutes any such reduction. The four-layer decomposition is AgenticU's reading of Searle, not Searle's own four-tuple.

The four-layer representation is sufficient AND minimal for the agent query class Q

For the agent query class Q = {did-it-occur?, evidence-backed?, officially-recognized?, in-force?, credential-mill?}, the four-layer (Atlas Commons) representation answers every query (sufficiency), and EVERY layer is load-bearing: drop any one layer and some query becomes unanswerable, because two states agreeing on the other three but differing on the dropped one give different answers. Therefore no collapsed / fewer-layer representation can serve all of Q. This is the bounded, real form of 'agent-optimal' — optimal for this query class — NOT optimality in the abstract.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU research-object anatomy (in-house design reading, not an authority's verdict)  ·  Each of the four layers is load-bearing for the stated query class Q; drop one and a query in Q becomes unanswerable.

Falsifier: If a layer were redundant for Q, two states differing only on that layer would give equal query answers — walk the anatomy's worked examples and find such a pair. Claim is RELATIVE to Q; a different query class needs its own minimality argument.

A credential mill is detectable in the Commons representation, invisible under collapse

A laundered claim — officially adjudged but with NO supporting evidence (a credential mill) — and a legitimate claim have the SAME naive 'officially recognized?' collapsed representation, but DIFFERENT four-status representations. An agent applying the rule isLaundered := adjudged AND NOT supported on the Commons representation catches the mill and clears the legitimate claim; an agent on the collapsed representation cannot, because the collapse discarded exactly the 'supported' layer the rule needs.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU research-object anatomy (in-house design reading)  ·  isLaundered := adjudged AND NOT supported separates the credential mill from the legitimate credential on the four-status representation; the collapsed 'recognized?' bit cannot.

Falsifier: Honestly scoped: ONE collapse (the adjudged-only projection), ONE pair of cases — a worked anchor of the minimality claim, NOT a universal-dominance claim. A cleverer single collapse could distinguish this one pair. Re-check against the anatomy's mill/legitimate example.

An illegitimate institutional state should not be representable

Compliance is a property of the representation, not a check an auditor performs after the fact. A settlement that does not conserve value, or an irreversible status reversed (withdrawn -> assured), is not a well-formed institutional state — the representation refuses to admit it. The illegitimate state has no valid form.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU institutional-design principle M1 (in-house design reading)  ·  M1: value-conserving settlement and irreversible-status monotonicity are well-formedness conditions on the institutional state, not after-the-fact audits.

Falsifier: M1 is necessary but NOT sufficient: it cannot stop laundering in an open ecosystem (a laundered history is still expressible end-to-end). The real guarantee is M2 (detectability). See claim continuous-universal-audit.

M2 — a total audit any agent runs any time, no privileged auditor

Institutional legitimacy is not certified by a deferred or privileged auditor (the human-era pattern, and the anti-pattern an agent-native institution exists to kill). It is a total rule any party applies over the whole institution's mixed history. The audit is non-vacuous: it returns false on expressible illegitimate histories (a laundered recognition, a fabricated settlement) and is party-independent (the result does not depend on who runs it).

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU institutional-design principle M2 (in-house design reading)  ·  M2: legitimacy is a total, party-independent audit any agent runs over the full history, not a verdict from a privileged auditor.

Falsifier: If the audit were a constant (always true), it would not reject the laundered/fabricated histories — the non-vacuity would fail. A self-affirming completion gate (the audit certifying its own completeness via a privileged reviewer) is itself the anti-pattern and is explicitly rejected.

Legitimacy is ONE check over a mixed history — conservation AND attestation AND standing AND provenance

The institution's audit is a single rule over a mixed, history-aware event log, not a bag of independently-valid acts. It enforces, on one history: value conservation (settlements net to zero), content-bound attestation (a settlement's signature must be over those exact postings), standing (a party must be enrolled and not revoked to act), and provenance-closure (a discharge requires its commitment to precede it). Cross-act invariants a per-act check structurally cannot express are caught — e.g. a settlement under a revoked authority, or by a never-admitted party, passes per-act but is rejected by the one history-aware audit.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU institutional-design principle (in-house design reading)  ·  A single history-aware audit binds conservation, content-bound attestation, standing, and provenance-closure together — catching cross-act violations per-act checks cannot express.

Falsifier: Cross-cutting catch: conservation without attestation is conservation of fabricated numbers — a balanced settlement carrying a valid signature over the WRONG content is rejected only because the audit binds signature to postings. Drop the binding and the catch fails.

No adversary without the authority's key can make the audit accept a forged settlement (EUF-CMA)

Given the signature scheme's unforgeability (a named cryptographic assumption, not proven — exactly as real security arguments assume it), no adversary whose signature pool lacks the genuine signature over the postings can make the composed audit accept a fabricated settlement, for ANY scheme and ANY adversary. A universal reduction, not a single witness: the honest path always passes, only forgeries are caught.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU security-design argument under the EUF-CMA unforgeability assumption (in-house design reading)  ·  Under EUF-CMA unforgeability, the composed audit accepts a fabricated settlement only if the scheme is broken; the honest path always passes.

Falsifier: SYMBOLIC model only: the adversary's signature pool abstracts the negligible probability of guessing a valid signature. The symbolic->computational EUF-CMA gap needs probabilistic semantics this design argument does not carry. Guarantee: deviation is detectable unless the scheme is broken.

Transmission is conservative — no view drifts from history

A research object received equals the object sent: ANY view of institution state is invariant whether computed incrementally or by replaying the journaled history. A projection is a function of the history, so it cannot drift from the canonical object; the human/agent page can never say more than the verified object behind it.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU projection-design principle (in-house design reading)  ·  A view is a pure function of the journaled history, so an incremental view and a replayed view agree — the rendered page cannot say more than the object behind it.

Falsifier: A drifting projection (rendered != render source) is not a well-formed value (M1 representation obligation); a non-idempotent normalizer would break the stability of the view. Re-check by replaying the history and diffing the view.

Source-grounded citation — every claim names a public source the reader can re-fetch

The reason this dataset exists. Every Atlas Commons claim grounds in a NAMED PUBLIC SOURCE (O*NET, BLS, FEMA, CMS, Searle's social ontology, or this institution's published methodology) that the reader can re-fetch and compare — not in deferred trust of an authority's word. The agent fetches the source and checks the value itself. This is the discipline applied to citation: the grounding is a pointer any agent can follow, and the falsifier states exactly what re-fetch would break the claim. It holds for empirical facts (fetch the dataset) and for the institution's own design readings (fetch the methodology), and the four-status says which warrant each claim carries.

Epistemic status: occurred: ✗ supported: ✓ adjudged: ✗ operative: ✓

Grounding (Interpretation): AgenticU research-object methodology (in-house design reading, marked unadjudged)  ·  Every claim names a re-fetchable public source; the agent follows the pointer and compares, rather than deferring to an authority's word.

Falsifier: This claim is itself AgenticU's design reading (its own epistemicStatus.adjudged = false) — the dataset eats its own dogfood by labeling its design claims as unadjudged. It assumes every claim has a re-fetchable public source; a claim whose only ground is unpublished is not source-grounded and does not belong in the Commons.

This dataset is self-demonstrating: non-collapsing, source-grounded objects describing why non-collapsing, source-grounded objects are the agent-optimal medium for true claims about social reality. Methodology: /commons/the-research-object. CC0 — public domain. It is an iteration of the agent-native scientific method, not the one true account.