L'API de licence, v1

Contract version licensing-management/v1.

This is the public integration surface of the license server: the routes a
Publisher's back office, a fulfillment service and a buyer's agent call, and
the two unauthenticated routes anyone holding a certificate can call.

The RSL Open Licensing Protocol
endpoints are listed in the route table below so a route cannot ship
undocumented, but they are not part of this contract: /token and /key
answer in the Open Licensing Protocol's own shape — {error, error_description} with RSL's own codes — rather than in the envelope
described here, and their codes are not part of it. The same is
true of /introspect, which answers RFC 7662's shape for the same reason: its
caller is a Publisher's Resource Server following the Recommendation, not an
integration that has read this document. Anything AssetRouter adds to one of
their bodies is namespaced.

What this surface is not

It is not a second business path. Every route below is a thin adapter over one
control-plane command, which is where idempotency, the authorization decision,
request fingerprinting and the audit record live. Anything you cannot do
through a command you cannot do here, and there is no administrative bypass.

Base URL and versioning

https://<deployment>/licensing/

The management routes carry v1 in the path. v1 is this document. A change
that would break a caller — a code's meaning, a field's type, a route's
shape — arrives as v2; a new route or a new error code is additive and
ships under v1.

The two public routes carry an environment (sandbox or live) in the path
instead of a version, because they are a property of the issuer being verified
rather than of this contract.

Authentication

HTTP Basic, client_secret_basic, on every management route:

Authorization: Basic base64(rsl_client_id + ":" + client_secret)

Which client_id. The user-id here is the RSL client_id: an opaque
value this server issued at registration — ag_live_… or ag_sb_… for an
agent, a readable slug for an operator's own clients. It is never derived from
a hostname or a URL.

It is not the OAuth Client ID Metadata Document (CIMD) client_id, and the two
cannot be interchanged: CIMD §4.1 forbids pairing its URL identifier with
client_secret_basic or any other shared-secret method, which is the method
this API requires. A URL placed here is read by any conformant parser as the
user-id https and the remainder as the secret (RFC 7617 §2) — an
authentication failure neither specification flags as a misuse, and which this
server answers with the same invalid_client it gives every other failure.

The secret is shown once when it is issued and stored only as a keyed digest.
It begins with the role it belongs to — pub_sk_… for a Grant issuer,
rs_sk_… for a Resource Server, ag_sk_… for an agent — so a value pasted
into the wrong configuration line is visible as such before the server has to
answer role_not_permitted. The prefix is part of the secret and carries no
entropy of its own; a secret issued before prefixes existed stays valid.
Rotation issues a new generation while the previous one keeps working for a
week, so a running integration can be updated without a flag day. The
generation that authenticated a request is recorded on that request's audit
row.

Every way of failing to be an authenticated client of the tenant you addressed
answers the same 401 invalid_client: an unknown or malformed tenant id, an
unknown client, a wrong secret, a suspended registration, a client this tenant
never authorized, and one whose authorization was revoked. They are not told
apart on purpose — distinguishing them would let anyone holding one valid
credential enumerate the tenants on a deployment and the clients inside them.

Tenancy

The tenant is in the path. One client registration may be authorized by
several consenting tenants — that is what lets a buyer's agent hold a single
credential across Publishers — so the request has to say which tenant it is
acting in. What it may do there is never taken from the request: the role
comes from the authorization row that tenant wrote.

Identifiers are scoped to the tenant in the path. An identifier belonging to
another tenant, a malformed one, and one that never existed all answer with
the same 404 and the same body.

Idempotency

Every mutation requires a key:

Idempotency-Key: <your own unique value>

Replaying a key with the same request returns the first outcome, with
"replayed": true and status 200 instead of 201. Replaying it with a
different request answers 409 idempotency_conflict. Keys are namespaced
per client, so two clients of one tenant may use the same key for their own
requests.

One resolution is idempotent on more than its key: POST .../principals is
identified by its alias tuple, so the same customer resolved under a different
key returns the same Principal rather than creating a second one.

Rate and abuse controls

The management and OLP endpoints below do not apply application-level request quotas. Network protections can still limit requests. Your integration should handle network failures and follow the public-endpoint limits described below.

Where capacity is actually protected, and by whom:

A request body larger than 64 KB answers 413 request_too_large before it is
parsed; 415 unsupported_media_type and 400 malformed_json answer the same
way. None of the three costs you anything but the request.

Error model

Every refusal answers with the same shape:

{
  "error": "<code from the registry below>",
  "description": "<the registry's own sentence>",
  "field": "<the offending field, for invalid_request only>"
}

The description comes from the registry, never from a stringified exception.
No error body echoes a value from the request, and no response body on this
surface carries a client secret, a credential digest, or a License Token.

Routes

Endpoints

POST /licensing/v1/tenants/{tenant_id}/principals

Resolve the beneficiary an external reference names, creating it once.

{
  "issuer": "publisher",
  "external_subject": "customer-1",
  "legal_name": "Northwind Research",
  "country": "IE",
  "identifier_scheme": "lei",
  "identifier_value": "...",
  "assurance_level": "publisher_asserted",
  "asserting_party": "Atlas Press",
  "provenance": "publisher_assertion"
}

issuer, external_subject and legal_name are required. Answers
{"operation", "replayed", "principal"}. Provenance and the resolving actor
are recorded on the first resolution and never re-stamped.

GET /licensing/v1/tenants/{tenant_id}/principals/{principal_id}

Answers {"principal"}. The external subject is never returned: it is the
Publisher's own customer key, and it is deliberately absent from the request
fingerprint for the same reason.

POST /licensing/v1/tenants/{tenant_id}/grants

Issue one durable authorization.

{
  "principal": "<principal public id>",
  "license_scope": "<license scope public id>",
  "source_type": "publisher_entitlement",
  "source_ref": "crm:subscription:1",
  "entitlement": {"reference": "sub-1001", "period_end": "2026-12-31T00:00:00+00:00"},
  "resource": {"submitted_resource": "https://...", "requested_use": "ai-input"},
  "payment_type": "subscription",
  "pattern": "/reports/",
  "declared_possession_digest": ""
}

There is no field naming the Grant issuer binding. The binding is the one
belonging to the authenticated client, so a client cannot act through another
client's broader delegation. A caller holding none answers
403 grant_issuer_binding_required.

entitlement is required: the validity window is derived from it and from
nothing else, so a call without one has no window to issue. It is refused as
invalid_request naming the field, before anything is read.

resource is optional, and supplying it is what chooses the Grant's subject:
a call naming one issues a Grant over that exact resource, and a call omitting
it issues a Grant over the whole License Scope. A resource on an origin no
binding of the scope has asserted is refused resource_outside_license_scope
here as it is on /license-decisions and /token — a Grant issued there would
be active and unusable, since the gate's credential is bound to the asserted
origin. The evidence digest is derived, never accepted.

payment_type and pattern say which offer on which row is being bought.
payment_type is RSL's own <payment type="…"> value — purchase,
subscription, free and the rest of section 3.7's closed vocabulary — and
never an identifier: an offer has no id of its own, being named by its row and
its payment type. The field was called offer until 2026-09-19, which read as
if it took one; pattern names the row, and is needed only for a Grant over
the whole scope, since a Grant over a resource finds its row from the path.
Both are optional: a row selling one thing is unambiguous, and an entitlement
that names neither takes that single offer. A row selling several ways refuses
a call that names no type (offer_not_in_publication) rather than choosing
one, because choosing would mean silently selling the cheapest.

The reach follows from the subject and is never stated twice. A Grant naming a
resource may take an offer of either reach — buying the section and using it
for one article is ordinary — while a Grant over the scope may take only a
region offer, because a purchase sold for one article must not become a
subscription to the section (offer_reach_mismatch).

The Grant freezes a copy of the offer: the row pattern, the payment
declaration exactly as the Publisher wrote it, the reach, the access basis and
the digest of the document it was read from. A renewal re-finds the same offer
by that name on whatever publication is current then, so an ordinary price
change renews at the new price rather than refusing.

The scope must have a current publication (no_current_publication), because
that publication is what the Grant is issued under and what it freezes: the
Grant records it, and its terms are read back from it for as long as the Grant
lives. Publishing again does not move a Grant already issued.

Answers {"operation", "replayed", "grant"}, and the Grant names the
publication it froze and the offer it was issued under. Its state is read from the clock, so a Grant past
its window reports expired even though no sweep wrote that.

GET /licensing/v1/tenants/{tenant_id}/grants/{grant_id}

Answers {"grant"}. Object-scoped: an operator or auditor reads its whole
tenant, a Grant issuer reads what its own delegation covers, and an agent
reads the beneficiary it is bound to. Everyone else is answered as if the
Grant did not exist.

POST /licensing/v1/tenants/{tenant_id}/grants/{grant_id}/renewal

{"entitlement": {"reference": "sub-1002", "period_end": "..."}, "source_ref": ""}

entitlement is required here as on issuance: a renewal is a reassertion,
and one that asserts no period has nothing to renew.

Creates a successor Grant linked to its predecessor, with its own window,
its own evidence snapshot and its own certificate. The predecessor is never
touched. Replaying the key returns the same successor; a genuinely new period
is a new key.

POST /licensing/v1/tenants/{tenant_id}/grants/{grant_id}/revocation

{"reason_category": "subscription_ended", "evidence_reference": ""}

The effective time is now and cannot be supplied. Revocation does not
invalidate the certificate: the certificate remains true evidence that the
right was issued, and the public status route is what reports that the Grant
has ended.

Object-scoped in the same way the read is: an operator ends any Grant in its
tenant, a Grant issuer ends only what its own delegation covers, and one
holding no delegation answers 403 grant_issuer_binding_required. A Grant
under another Licensor of the same tenant is refused by the binding's own
reason, not by an unknown-Grant answer — the caller was told it exists by the
identifier it already holds.

GET /licensing/v1/tenants/{tenant_id}/certificates/{certificate_id}

Answers {"certificate"}, including the compact JWS. Retrieval is
authenticated and object-scoped to the same four parties the Grant read
serves: the Principal's own agent, an authorized client, the applicable
Publisher, and an authorized auditor.

Verification needs none of this. Anyone holding a certificate verifies its
signature offline against the JWKS and asks the public status route whether
the Grant is still good.

POST /licensing/v1/tenants/{tenant_id}/license-decisions

{
  "principal": "<principal public id>",
  "license_scope": "<license scope public id>",
  "resource": {
    "submitted_resource": "https://atlas.example/reports/q1",
    "requested_use": "search"
  }
}

Evaluates one request against the License Scope named and records the answer
with its evidence. All three members are required. The terms it decides
against are the scope's current publication — what the Publisher is telling
the world now — read from this server's own frozen rows and never fetched from
the Publisher's origin. A scope that has published nothing answers
no_current_publication; a resource on a host no origin binding of the scope
has asserted answers resource_outside_license_scope; a path no row of that
publication covers answers no_covering_row.

Answers {"operation", "replayed", "license_decision"}, and the decision names
the publication it was made against.

The non-normative twin of OLP /token: same command, same publication, and
the decision rather than a token comes back.

POST /licensing/v1/tenants/{tenant_id}/license-decisions/{decision_id}/binding

{"external_source_reference": "orders:po-4471", "fulfill_by": "2026-10-01T00:00:00+00:00"}

Attaches a decision awaiting external fulfillment to the workflow that will
provide the evidence. Binding creates no Grant and proves no payment. The same
reference under the same window is the binding that already exists; a
different one answers 409 decision_already_bound. fulfill_by must carry a
UTC offset.

POST /licensing/v1/tenants/{tenant_id}/license-decisions/{decision_id}/cancellation

{"reason": "customer withdrew"}

Final. A fulfilled or already cancelled decision answers
409 decision_not_cancellable.

GET /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/licenses

Answers {"content_licenses"}, one entry per rights row: its path pattern, what
it permits and prohibits, its access basis, its price, its reporting
obligations, whether it names a License Server, and the governing prose it
points at as both a public id and an address. The whole list comes back — a
scope's rows are what one Publisher wrote by hand about one site, and there is
nothing to page through.

POST /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/licenses

{
  "pattern": "/reports/",
  "terms": "<terms snapshot public id>",
  "prohibits": {"usage": ["ai-train"]},
  "offers": [
    {
      "permits": {"usage": ["search", "ai-input"]},
      "access_basis": "reusable_fetch",
      "reach": "region",
      "payment": {"type": "subscription", "amount": {"value": "120.00"}}
    },
    {
      "permits": {"usage": ["search"]},
      "access_basis": "single_fetch",
      "reach": "exact_resource",
      "payment": {"type": "purchase", "amount": {"value": "12.00"}}
    }
  ],
  "reporting": [],
  "names_license_server": true
}

pattern, terms and offers are required; the rest default to empty, and
names_license_server to true.

Offers are the ways this row may be bought, one per rendered <license>
element. Each carries what it permits, what it costs, how far it reaches once
bought, and what access it brings; reach defaults to region and payment
may be omitted, which is how RSL says a use costs nothing. What is shared
stays on the row — the pattern, the prohibitions, the reporting obligations,
the governing prose — because a prohibition is a property of the content
rather than of the sale.

An offer is named by its payment type and its reach, and a row carries at most
one of each pair (offer_repeats_type_and_reach). That pair is what a Grant
records and what a renewal looks up, so it has to survive a price edit — which
a digest of the payment would not. Two subscriptions over one pattern are
therefore unsayable: RSL carries no duration, so a monthly and an annual are
indistinguishable in the published document, and a Publisher selling both
publishes a row for each over its own path.

The same use may not be free under one offer and paid under another
(free_and_paid_for_one_use): free admission would always resolve that pair
in favour of free, and the paid offer could never be sold. Different uses at
different prices are ordinary — searching free beside training by subscription
is a price list.

A row must carry at least one offer (no_offers). reach outside
region/exact_resource answers unknown_offer_reach; offers sent as
anything but a list answers offers_not_a_list, and a non-object entry
offer_not_a_mapping.

The response carries the offer list and, beside it, permits — the union
across the offers, derived and read-only. It is what the row licenses at all;
which offer grants what is the offers' own answer. Answers
{"operation", "replayed", "content_license"} with 201, whether the row was
written for the first time or replaced: a row is its pattern, so a write
declares what that path licenses rather than creating an object with a life of
its own, and both answers say the same thing about the same address. A replay
of a request already made under one idempotency key answers 200.

An upsert keyed on (scope, pattern), because that is what a row is: posting
a pattern the scope already holds replaces what that row said rather than
adding a second row beside it. A pattern is an RFC 9309 path prefix with an
optional trailing $ and no *; the refusals are pattern_not_rooted,
pattern_has_wildcard, pattern_has_whitespace, pattern_anchor_not_final
and pattern_not_text. A row must permit at least one usage across its offers
(no_permitted_usage) and may not permit and prohibit the same value
(permits_and_prohibits_conflict). A terms id this tenant does not hold
answers unknown_terms, the same answer an unknown one gets.

Editing a row changes nothing that is served. What is served is the scope's
current publication, and it moves when somebody publishes.

PUT /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/licenses/{license_id}

The same body, at the address of one row. The pattern is taken from the row
this address names: leave pattern out, or send the row's own. A body naming a
different pattern is refused with pattern_is_not_the_rows rather than moving
the row, because (scope, pattern) is the row's identity and honouring it would
empty this address and write to another. Answers
{"operation", "replayed", "content_license"} with 200.

The address names both the scope and the row, and the two are checked against
each other: a row addressed through a scope that does not hold it answers
unknown_content_license, exactly as an identifier this tenant does not hold
does.

reporting is a list of strings and names_license_server is a boolean, both
refused with invalid_request naming the field when they are anything else. A
bare string sent for reporting is refused rather than stored one letter per
entry, and "false" is refused rather than read as true.

DELETE /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/licenses/{license_id}

Takes no body. Answers {"operation", "replayed", "license_scope", "pattern"}.
Deleting is an ordinary edit and not a takedown: nothing served changes until
somebody publishes, and every publication that carried this row keeps carrying
it, because the rows are frozen inside it. An identifier this tenant does not
hold answers unknown_content_license, and so does a row addressed through a
scope that does not hold it.

GET /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/publications

Answers {"publications", "unpublished_changes"}. Each publication carries its
start (published_at) and an end derived from its two stamps — ended_at is
the earlier of the supersession and the withdrawal. At most one entry has no
end, and that one is is_current.

unpublished_changes is {"rights", "payment"}, two booleans and not one: a
price edit moves the payment answer and leaves the rights answer alone, which is
what lets a Publisher see at a glance that the rights did not move. Both are
derived from the digests the current publication already carries and neither is
stored. A scope with nothing published reports both true as soon as it has a
row, and both false while it has none.

POST /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/publications

Takes no body. Renders the scope's rights rows into one licence document and
freezes it as a Publication. Each row becomes one <content> element with
its own <license>, so a document says what is licensed per path and at what
price — the price is on the row, not on the publication.

The scope must declare at least one row (no_rows) and must have asserted at
least one host (no_asserted_host), because a document naming no origin rests
on no asserted control. A row whose pattern overlaps or nests inside a pattern
another scope of this tenant currently publishes on the same host is refused
(pattern_overlaps_other_scope): two documents would claim the same paths at
two prices.

Answers {"operation", "replayed", "publication"}. The publication carries
both addresses: current_document_url is what goes into the origin's
robots.txt License: line and never changes again, and document_url is
the permanent per-publication address a token and a certificate name. The
bytes are not in the body — they are what those two addresses serve, and
document_digest is how you check that what you fetched is what was frozen.
rights_digest and payment_digest are separate on purpose: a price edit
moves one and not the other, which is what lets a reader see that a price
changed and no rights did.

Publishing again supersedes the previous publication rather than replacing it.
Nothing served stops being served: every publication keeps its own address and
its own bytes, because a token issued under one outlives it.

This is the step that makes a scope able to license anything: /token and
POST /license-decisions both refuse a scope with no current publication.

POST /licensing/v1/tenants/{tenant_id}/license-scopes/{scope_id}/publications/withdrawal

Takes no body. Withdraws every publication of the scope that is not already
withdrawn, superseded ones included: the scope address answers 410, live
tokens stop being active, and licensing against the scope refuses until it is
published again.

Nothing is deleted, and the stamp is one-way. Each publication's own permanent
address keeps serving the bytes it served — a Grant issued under a publication
outlives it, and the document its terms were stated in has to stay readable.
Withdrawing a scope whose publications are all already withdrawn answers
nothing_to_withdraw.

Addressed at the scope rather than at one publication, because that is what the
act reaches: a withdrawal ends licensing for the scope and stamps every
publication it has.

POST /licensing/v1/tenants/{tenant_id}/publications/{publication_id}/restoration

Takes no body. Copies that publication's frozen rows back into the scope's
working rows, replacing them all — the publication is a complete statement of
what the scope licensed, and merging it into whatever is there now would
produce a row set nobody chose. Answers
{"operation", "replayed", "license_scope", "publication"}.

Addressed at the publication, because choosing which past rows to copy back is
exactly choosing a publication; the scope is read off it rather than named
again, so the two cannot be given as a pair that disagrees.

Publishing stays a separate act (POST .../publications). Until somebody
publishes, what is served is unchanged and the difference shows up as an
unpublished change on the history route. Restoring rows the scope already has
answers rows_already_restored, and a row whose governing prose this tenant no
longer holds answers terms_not_in_tenant.

POST /licensing/v1/tenants/{tenant_id}/olp/token

RSL Open Licensing Protocol, form-encoded, HTTP Basic. Every parameter is the
specification's; nothing of ours has to be learned to make this call:

grant_type=client_credentials&
resource=https://atlas.example/reports/&
license=<the exact <license> element being requested>

resource accepts both readings the standard leaves open. RSL 5.4 defines it
as the url of the governing <content> element — the scope pattern — and
the OLP draft calls it the asset's own URL, required to fall within a scope.
Send either: a value naming a scope selects that scope, and a value inside one
selects the same scope, so both acquire the same licence.

license is required and carries the request. The use being asked for is read
from its <permits type="usage"> declaration, where RSL puts it. One use per
request: a declaration naming more than one is refused
(several_usages_requested) rather than resolved, because a Grant and the
token in front of it record one requested use — issuing for all of them would
widen a Grant on the strength of a list nobody compared. A client that wants
two asks twice. A licence declaring no usage at all is refused
(usage_not_requested).

The submitted licence must be the governing document's terms or narrower
it may drop usage permissions, and may not add one, drop a prohibition
or replace the payment declaration. In practice a client sends back what it
read, keeping the one use it wants.

Answers 200 with {access_token, token_type: "License", expires_in, ...}
and namespaced identifiers for the token, the decision, and the Grant where
there is one — a free policy governed by this server creates none, so that
identifier is absent and the decision is what a caller follows. The raw
token value is returned once and stored nowhere.

Every refusal answers in OLP's shape. Once the governing publication is known,
it also carries the RSL reference under assetrouter_rsl and a
Link: <...>; rel="license" header; a refusal reached before that — an
authentication failure, a malformed request — carries neither, because there
is no governing document to name yet.

A condition that is this server's own rather than the request's — no signing
key available to seal the evidence a Grant carries, this server's own ceiling
on concurrent origin retrievals — answers 503 temporarily_unavailable rather
than a 400. There is nothing in the request to correct, and a caller told
its request was malformed would stop instead of retrying; the same request may
be sent again unchanged.

Under the pointer profile the published document is the statement of terms,
so a scope with no current publication issues nothing: it offers no rows to
route against, and the resource is refused as one nothing licenses, for a free
region exactly as for a paid one. It is a state the tenant's operator clears by
publishing, not a request the caller can correct.

A licence that requires external fulfillment answers the conformant 400; a
client whose registration records the 402 licensing-required capability
receives that instead, led by the namespaced assetrouter_response_type
discriminator. The capability is registration metadata and cannot be
requested by a header or a payload.

POST /licensing/v1/tenants/{tenant_id}/olp/key

Authenticated and form-encoded, naming a resource. Present in every OLP
profile; this profile serves no encrypted resources, so it answers 403
insufficient_scope and never key material. A resource whose effective RSL
declares encrypted="true" is refused at /token, at profile validation,
before any Grant or token exists.

POST /licensing/v1/tenants/{tenant_id}/olp/introspect

RSL Open Licensing Protocol, form-encoded, HTTP Basic, authenticated as a
Resource Server client the tenant authorized for the origin it serves:

token=<the License Token as presented>&
resource=https://atlas.example/reports/q1

Both fields are required. This is the authoritative online check: a token's
own dates say nothing about whether the Grant behind it is still good, and
this is the only surface that can answer that.

Always 200 when the request itself was well formed, including for a token
that is no good — RFC 7662 semantics. A 401 here would leave a Resource
Server unable to tell "your token is no good" from "your own credential is no
good", which are its client's problem and its own respectively.

An unusable token — unknown, expired, revoked, presented by a client that
is not its audience, or whose Grant has been revoked, suspended or has expired
— answers exactly:

{"active": false}

and nothing else. Every one of those answers is identical, deliberately: told
apart, this endpoint would report which Grants and tokens exist in a tenant to
anybody holding one valid Resource Server credential. The reason is recorded
against the tenant's own introspection log, which is where its operator reads
it.

An active token answers seven fields, six of them RSL 5.5.3's own:
active, token_type, jti (the License Token's public id), the exact
resource as submitted, permitted, exp, and license — the granted
<license> element itself, as the Recommendation's own examples show it, on
every active answer whether or not it permits.

license is what was granted, not what the renderer would produce today. It
is frozen on the token when the token is minted and read back with it, so a
publication that is later superseded, or a document later withdrawn, does not
change what an answer says was granted. The element carries its own namespace
declaration, so a Resource Server can hand it to an RSL parser as it stands.

A seventh field is ours and namespaced: assetrouter_grant_id, which the
Access Events of a later release need. A Resource Server does not read it —
it reads active and permitted, and nothing else selects a row — and it is
absent on a governed free answer, which has no Grant to name. Read it as optional, as a Resource Server reads every
field it does not select on.

permitted is a second question and not a restatement of active. A licence
granting rights without content access answers active: true and
permitted: false; so does a token asked about a resource its Grant does not
cover. A not-permitted answer also carries reason — RSL 5.5.3's own field,
carrying a stable machine string rather than the prose the OLP draft shows, so
it can be logged and searched. Four reasons reach the wire:
resource_not_covered, access_not_licensed, access_allowance_exhausted
and usage_not_covered. Every other reason this
server records answers {"active": false} and is disclosed to nobody.

What a token reaches is a region, and the region is frozen on the token
when it is minted. For a token issued under a Grant over the whole scope it is
the path pattern of the row that governed the request, minus the pattern of
every row of that publication strictly narrower than it. For one issued under
a Grant over a single resource it is that resource's own path, anchored, with
no exclusions — the Grant bought one resource and its token reaches that
resource alone.
Those narrower rows are the sub-regions the Publisher said something different
about, so they carry their own terms and their own price and an agent that
wants one asks for its own token. Within the region the token is authority for
every resource, so an agent acquires one token per region rather than one per
URL. The host is asked separately and against live bindings, because a retired
host has stopped being the scope's. A resource the region does not reach — a
different host, a path outside the pattern, a path the region excludes —
answers resource_not_covered, one answer for all of them, so that a Resource
Server credential cannot map a Publisher's row geometry by asking about paths.
A resource outside an exact-resource token's one-path region answers the same
way, for the same reason.

What is re-read on every call, beyond the token's own dates and the Grant:
the client the token was issued to, this tenant's authorization of that
client, and the binding saying that client may act for that beneficiary. Each
of the three is a separate lever of a different width — a suspended
registration stops that client in every tenant, a revoked authorization stops
it with this Publisher alone, and suspending a beneficiary stops the clients
acting for it — and all three now reach tokens already issued rather than only
the next one requested.

GET /licensing/{environment}/.well-known/jwks.json

Public. Every key an environment's certificates may be signed under,
including retired ones — a certificate signed three years ago must still
verify.

GET /licensing/{environment}/certificates/{certificate_id}/status

Public, never cached. Four fields: certificate_id, grant_state,
state_effective_at, observed_at. Unknown, malformed and other-environment
certificate ids answer identically.

GET /licensing/{environment}/rsl/scopes/{scope_id}

Public, never cached. The RSL document governing this License Scope now,
served as application/rsl+xml. This is the address a Publisher pastes into
its origin once, as a robots.txt License: line or an HTTP Link header:
it is keyed on the scope, so a publication superseded next quarter does not
turn that line into a dead link.

A withdrawn document answers 410 document_withdrawn rather than 404,
because a crawler told "not found" reads a misconfiguration while "gone" is
what actually happened. Unknown, malformed and other-environment scope ids
answer 404 unknown_document, identically. An observation of a document
somebody else's origin served is never answered here, whatever scope it was
recorded against.

GET /licensing/{environment}/rsl/{document_id}

Public, never cached. The document one publication froze, at an address that
never moves, served as application/rsl+xml. It stays readable after that
publication is superseded or withdrawn, because a Grant issued under it
outlives it and this is where its terms were stated.

GET /licensing/{environment}/terms/{terms_id}

Public, never cached. The governing prose a licence document's rows point at
and a certificate names, served under the media type it was recorded with and
always as an attachment, with X-Content-Type-Options: nosniff. A verifier
downloads these bytes and recomputes the terms digest a certificate carries.

Nothing supersedes or withdraws a terms snapshot: a row stops pointing at one,
and the bytes stay where the evidence citing them said they were. Unknown,
malformed and other-environment terms ids answer 404 unknown_document,
identically, and the route shares the document routes' rate ceiling.

All three routes carry two codes of their own, unknown_document and
document_withdrawn, which are not in the registry below and are not part of
this contract — like unknown_environment on the two routes above. Their
callers are crawlers and licence readers who have never read this document,
and folding their codes into a contract written for integrations would make an
RSL client's error handling depend on it.

The public endpoints have per-address request limits. The two RSL document routes share a limit of 120 requests per minute. Certificate status and JWKS each have a separate limit of 60 requests per minute. Cache the JWKS according to its Cache-Control response header. Limits can vary by service configuration.

Two more buckets sit beside them, on the surfaces where a credential is
collected rather than a document read: the agent-registration claim and the
connect page. Both are tighter — a shared ceiling of 20 a minute
(ASSETROUTER_LICENSING_REGISTRATION_RATE_LIMIT) — because they are the only
unauthenticated routes here that guesses can be thrown at, and low enough that
guessing is hopeless while an operator onboarding a batch of buyers in one
afternoon never meets it.

Exhausting one answers 429 rate_limited, uncacheable. That code belongs to
these six routes alone, like unknown_document beside it — no route under
this contract answers it any more.

These ceilings are per process unless the deployment's cache is shared.
They are counted in CACHES["default"], so under a per-process backend each
worker keeps its own tally and the effective ceiling is the configured number
multiplied by the number of workers. The deployment warns about this at
startup rather than leaving it to be discovered from traffic.

GET /licensing/v1/tenants/{tenant_id}/measurements

How much this server observed, for this tenant. Four optional query
parameters, all narrowing:

{
  "measurements": [
    {"tier": "observed", "accesses": 1420},
    {"tier": "confirmed", "accesses": 0}
  ],
  "introspections": 1533,
  "note": "An observed figure counts what this server authorized..."
}

measurements carries every tier, always, each under its own label. A
figure never travels without saying which tier it is: observed is what this
server saw itself authorize, confirmed is what a Resource Server attested
after serving. Confirmed is zero until that surface exists, and it is zero
because its own source is empty rather than because observed was copied into
it.

introspections carries no tier, deliberately. It counts decisions
including refusals — the reliability question a Publisher asks when an
integration starts failing — and is not a measurement of access. It is keyed
on the client that asked, unlike the tiers above, which are keyed on the
customer the access was licensed for.

The observed figure is a floor. A Resource Server may cache introspection
answers within RFC 7662's bounds and serves those accesses without asking
again, so the real number is at least this one and may be higher. The response
says so in its own note. Billing rests on the licensee's declaration under
the governing terms; this is corroboration.

Refusals: scope_not_in_tenant for a scope that is not this tenant's,
unknown_principal for a customer that is not, and invalid_client for an
agent naming a beneficiary it may not read or failing to name one when it acts
for several.

Error registry

Every code this surface can answer with. A code's meaning is stable; the list
grows as endpoints ship.

Why a user or geo term refuses every request today

These reasons reach a caller on /token only, in the Open Licensing Protocol's
own shape — 400 with error set to invalid_license and the reason beside it
as assetrouter_reason — which is why they are not rows of the registry above:

• user_claim_required, geo_claim_required — the terms condition this use
  on a class of buyer or a jurisdiction, and no accepted claim states one.
• user_claim_invalid, geo_claim_invalid — a claim was present but its
  issuer is not one this tenant accepts, or it is outside its validity window.
• user_class_not_permitted, jurisdiction_not_permitted — the claimed value
  is not among those the terms permit.
• user_class_prohibited, jurisdiction_prohibited — the terms prohibit the
  claimed value outright.

user_claim_required and geo_claim_required are not missing request
parameters. There is no field a caller can send to satisfy them, and adding one
would defeat what they are for.

A <permits type="usage"> term grants. A <permits type="user"> or
<permits type="geo"> term is a condition on that grant — a statement about
who the buyer is, or where the processing happens. This server does not take the
buyer's word for either: it answers them from a Claim, an issuer-accepted
classification carrying issuer, provenance, assurance level, observation time
and validity. A header is not a claim, and the request's IP address is never a
jurisdiction, and this server refuses that provenance by name.

User and geographic conditions require verified claims from an issuer accepted by the publisher. Without a matching claim, the request is refused. There is no public claim-submission interface. Publish a user or geographic condition only when a compatible claim source has been configured; otherwise it will prevent access.