Authority: Normative summary · Status: Admission factory implemented; pilot review incomplete; selected-origin retrieval disabled · Applies to: E3.1–E3.2 · Verified against: current checkout · Source: spec/atlas/ATLAS_CONTROL_PLANE_0_1.md, ADRs 006–007, and Atlas JSON schemas
Purpose
The Atlas control plane keeps candidate selection, catalog, policy, technical,
publisher, health, adapter-trust, and mapping-trust claims distinct. The exact
500-origin selection is not a network allowlist.
The committed state has three cataloged public identities, three completed
human policy reviews, two semantically linked technical records, zero live
origins, disabled scheduling and zero frontier jobs.
Admission factory
Each pilot origin has its own directory under atlas/admissions/{origin_id}:
The factory binds each identity to the accepted selection, rehashes local
policy evidence, rejects duplicate origins and execution aliases, and renders
deterministic policy, registry, batch, and dossier artifacts.
It cannot create a decision. A canonical admission requires a completed
decision artifact with reviewer_type: human and an approval reference. That
catalog admission remains independent from policy: a human may admit an
identity while its policy stays pending + uncertain, but no retrieval work
order may be produced from that state.
The E3.2 batch contains 25 catalog dossiers across all nine domain families,
11 proposed jurisdictions, and six proposed languages. Twenty-two await
founder review. TWIRX, World Bank and RFC Editor render as public canonical
records. Their decisions are respectively permit_live,
permit_with_constraints and profile_only; none enables the disabled Atlas
scheduler or broad retrieval.
The batch also includes five explicitly provisional commercial/access
candidates: LA Times, Le Monde, The New York Times, Reuters, and The Washington
Post. Their records are inferred catalog-review proposals. No price, current
terms, payment mechanism, policy permission, or executable commercial route is
claimed.
The generated admission work queue covers all 500 selected origins without
inventing 475 evidence dossiers. It distinguishes prepared from
not_prepared, preserves pending policy as uncertain, and gives every origin
an explicit next action. The current queue contains 25 prepared and 475
unprepared dossiers.
State model
Only technical stage is ordered, and only within that dimension. A compiled
adapter does not complete policy review. Publisher approval does not establish
health or live status. Adapter conformance does not approve a mapping.
Pending policy is fail-closed and never counts as completed. Its decision is
uncertain; a completed uncertain decision requires explicit human review.
Origin records now describe interface declarations, capability candidates,
effect classes, access/economic metadata, provisional offer candidates, and
publisher-readiness signals. Every non-unknown declaration is bound to an
immutable evidence digest and records its evidence class.
This metadata is not execution authority. The only admitted capabilities in
E3.2 are the four existing public-read E2 operations across the two canonical
public origins. WebMCP, browser, authenticated, write, financial, legal,
material, and destructive capabilities may be described but cannot execute.
The API exposes these declarations only in bounded origin descriptions and
metrics; it does not implement semantic resolution or ranking.
Canonical E2 identity
atlas/registry.json is the canonical origin registry. E2 retains its bounded
execution routes and typed inputs, but every execution entry binds a
registry_id. Host, publisher, scope, or alias mismatch prevents E2 startup.
The registry includes twirx-org, api-worldbank-org and rfc-editor-org as
public records. The first two retain exact E2 replay, contract, conformance,
proof and semantic-closure evidence; the third binds the approved historical
archive-only profile. The controlled E2 origin is a test_fixture
and is excluded from every Genesis-500 public counter.
Its source record lives under atlas/fixtures, separately from the 25 public
pilot dossiers, and the admission render preserves it in the one canonical
registry.
Dry-run frontier
The frontier emits origin IDs, decisions, budgets, and due times—never a
destination URL—and declares network_access: disabled. Only completed
permit_live or permit_with_constraints state plus positive budgets and an
active scheduler could produce a dry-run job. The committed plan has none.
Every selected origin receives exactly one frontier outcome. The current plan
contains zero jobs because scheduling and all retrieval budgets remain
disabled. A completed policy decision alone is not execution authority.
Full-catalog runtime workload
make stress-e3-500 runs the actual loopback Atlas HTTP service across all
500 origins. It verifies five list pages, 500 direct descriptions, all family
and orthogonal-state filters, all 500 frontier outcomes, malformed-request
rejection, concurrent repeated lookups, and byte-identical status after a
service restart. Its client accepts only a literal loopback HTTP endpoint and
cannot fetch a selected origin.
Read-only API
List filters are family, each state-dimension name, limit, and cursor.
There is no URL parameter, invocation route, write route, crawler, browser,
model, or HTTP client in this process.
Failure behavior
Duplicate keys, unknown fields, trailing JSON, symlinked or digest-substituted
policy evidence, duplicate origins or aliases,
non-HTTPS or credentialed identities, quota drift, missing state evidence,
policy mismatch, a false human-review claim, unsafe evidence paths, digest
substitution, unsafe scheduler state and unbounded API queries fail closed.
The server accepts only a literal loopback listener.Last modified on August 11, 2026