# Atlas Commons — Institutional Epistemics (agent context pack)

> Machine-first context for the research objects at `/commons/institution-epistemics.jsonld`.
> Format: agent-native research objects — every claim names a **public source the reader can re-fetch**;
> every claim carries its **non-collapse four-status**; every claim ships its **falsifier**. You do not have
> to take this page's word. Re-fetch the source.

## What this is

Nine research objects. Each states one of the institution's **epistemic commitments** — the design principles
that govern how every Atlas Commons claim is represented and grounded — represented *without collapsing* its
four epistemic layers, and grounded in named public methodology you can re-fetch: the research-object anatomy
at `/commons/the-research-object`, plus the public sources (BLS, O*NET, Searle's social ontology) those
principles operate over. These are AgenticU's own design readings, not an authority's verdicts: the dataset
says so in each object's four-status (`occurred=false`, `adjudged=false` — the lowest warrant, on its face).
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.

## The two ideas

1. **Non-collapse four-status.** A social-reality claim does not have one truth value. Each object carries
   `epistemicStatus = {occurred, supported, adjudged, operative}` — did the underlying act happen / is there
   evidence / did an authority recognize it / is it currently in force. A collapsed representation (one
   truth-bit) destroys these distinctions; the Commons preserves them, which is what lets an agent detect a
   *credential mill* (adjudged but unsupported) that a flat representation cannot.

2. **Source-grounded.** Every claim names a **public source the reader can re-fetch** — a labor-record claim
   cites BLS or O*NET; a design-commitment claim cites the public methodology anatomy (and, where applicable,
   its intellectual source, e.g. Searle 1995). An agent *checks* the claim by following its `sourceUrl` and
   comparing the cited value, rather than deferring to an authority's word. The grounding is a pointer any
   agent can follow, and the falsifier states exactly what re-fetch would break the claim.

## The claims (lead with the cross; honest about scope)

All nine are AgenticU's design readings: `claimType: interpretation`, `occurred=false`, `adjudged=false`,
`supported=true`. Each grounds in `/commons/the-research-object` (the public methodology), and claim 1 also
names Searle's social ontology as its intellectual source.

| Claim | What it commits to |
|---|---|
| **non-collapse-of-social-reality** | The four epistemic layers are pairwise independent — no layer is derivable from the others (Searle's social ontology, operationalized as the four-layer decomposition) |
| **non-collapse-minimality** | Four-status is sufficient AND minimal for the agent query class Q — each layer is load-bearing |
| **laundering-detectable** | A credential mill (adjudged ∧ ¬supported) is detectable in the Commons representation, invisible under collapse |
| **compliant-by-construction** | M1 — an illegitimate institutional state is not a well-formed representation |
| **continuous-universal-audit** | M2 — a total audit any agent runs over the full history, no privileged auditor; non-vacuous & party-independent |
| **composed-institutional-audit** | Legitimacy is ONE check over a mixed history: conservation ∧ attestation ∧ standing ∧ provenance |
| **fabrication-caught** | EUF-CMA — under the unforgeability assumption, no adversary without the key makes the audit accept a forged settlement |
| **conservative-transmission** | No view drifts from history — the object received equals the object sent |
| **source-grounded-citation** | The meta-commitment: every Commons claim names a public source the reader can re-fetch |

## How to verify (any agent, any time, no credential)

Follow each claim's `sourceUrl` and compare the cited `value` against the public source. The query API returns
this recipe per claim at `/api/commons/query?verify=<id>`:

```
GET /api/commons/query?verify=non-collapse-of-social-reality
  → { source, sourceUrl, value, instructions: ["Fetch the named source and compare the value", …] }
```

For the design-commitment claims, re-fetch the methodology anatomy at `/commons/the-research-object` and check
the reading against it. For the labor-record objects elsewhere in the Commons (`/commons/roles/*.jsonld`,
`/commons/tasks/*.jsonld`), re-fetch the cited public dataset (BLS, O*NET) and compare the value. If the source
reports something different — or no longer resolves — the object is stale or wrong and should be rejected. That
is the whole point: the claim carries its own re-fetchable check.

## The honest scope (no-overclaim)

- "Agent-optimal" is **relative to the query class Q**, not abstract. A different query class needs its own
  minimality argument.
- `fabrication-caught` is the **symbolic** model; the symbolic→computational EUF-CMA gap is named, not closed.
- Every commitment here is **AgenticU's own design reading** (`adjudged: false`) — the dataset labels its own
  warrant honestly rather than dressing a design choice as an authority's verdict.
- This is an iteration of the agent-native scientific method, not the one true account.

## The corpus — ~20,000 research objects

The same format spans the whole Atlas Commons:

- The institution's epistemic commitments — `/commons/institution-epistemics.jsonld` (these nine objects).
- ~1,016 occupations — e.g. `/commons/roles/role-29-1141-00.jsonld` (Registered Nurses); four-status varies.
- ~19,265 tasks — the atomic unit of work — e.g. `/commons/tasks/task-1.jsonld`.
- Query / verify any of them: `/api/commons/query` (`?id=` a claim, `?verify=` for the re-fetch recipe).
- The machine vocabulary every object's `@context` resolves to: `/vocab/agent-native`.

Everything in the Commons is a research object, and every research object points back to the methodology
anatomy for its definition. CC0 — public domain.
