generated: '2026-09-19'
method: searched
source: >-
  openapi/afmr-ai-discovery-api-openapi.yml + live response headers and bodies observed on
  2026-09-19 (GET /, /openapi.json, /nope, /mcp; POST /api/rpc, /a2a) + https://afmr.ai/llms-full.txt
  + https://afmr.ai/changelog.json (which documents the CORS / cache / Link / problem+json policy).
description: >-
  Cross-cutting behaviour of the AFMR Discovery surface. It is a READ-ONLY static-document API
  plus two JSON-RPC protocol endpoints (MCP, A2A) that are themselves read-only; there is no write
  surface, so idempotency, reversibility and dry-run are all not-applicable rather than absent.
base_url: https://afmr.ai
api_style: REST-style GET of static JSON/text/HTML/XML documents; JSON-RPC 2.0 over POST for MCP (/api/rpc, alias /mcp) and A2A (/a2a)
authentication:
  scheme: none — every surface is public read-only
  detail: authentication/afmr-ai-authentication.yml
write_surface: none
idempotency:
  supported: na
  coverage: na
  scope: []
  note: >-
    No mutating operation exists. All 20 GET operations are inherently idempotent; the two POST
    protocol endpoints are read-only JSON-RPC calls, and every MCP tool declares
    annotations.idempotentHint true / destructiveHint false / readOnlyHint true. No Idempotency-Key
    header exists because nothing needs one; NO Idempotency pointer is emitted.
reversibility:
  status: na
  write_surfaces: []
  note: Read-only API — there is nothing to cancel, refund, void or restore. Recorded as na so the dimension leaves the denominator rather than scoring zero.
dry_run_mode:
  status: na
  note: Read-only API; no operation has side effects to rehearse.
pagination:
  style: none
  note: Documents are returned whole. The only paged-shaped input is the MCP tool lookup_failure_mode (limit 1-10, default 5) which caps matches rather than paginating.
field_expansion:
  supported: false
metadata:
  supported: false
request_tracing:
  request_id_header: null
  note: No request-id header is returned; Cloudflare's cf-ray is the only per-request identifier observed.
caching:
  cache_control: 'public, max-age=3600, must-revalidate on documents; public, max-age=60 on problem responses'
  last_modified: exposed (Thu, 30 Jul 2026 00:00:00 GMT on the discovery documents)
  exposed_headers: ETag, Last-Modified, Link
  note: The provider explicitly designs for revalidation (changelog 0.1.1-review).
content_negotiation:
  note: Each document has a fixed media type (application/json, application/schema+json, text/plain, text/html, application/xml, application/atom+xml); there is no Accept-driven negotiation. MCP requires Accept including application/json (text/event-stream is accepted but no SSE stream is offered).
cors:
  access_control_allow_origin: '*'
  access_control_allow_methods: GET, HEAD, OPTIONS
  access_control_allow_headers: Accept, Content-Type
link_relations:
  on_html: [canonical, alternate (llms.txt / llms-full.txt / resources.json), describedby (afmr.json), service-desc (openapi.json), service-meta (agent-card.json), license]
  on_problem_responses: [describedby (afmr.json)]
versioning:
  scheme: per-component semantic; standards profiles versioned in the URL path (/standards/<profile>/<version>/)
  header: none
  detail: lifecycle/afmr-ai-lifecycle.yml
  changelog: changelog/afmr-ai-changelog.yml
error_envelope:
  media_type: application/problem+json
  rfc9457: true
  shape: '{ "type": "about:blank", "title", "status", "detail", "instance", + extensions (discovery, resource_index, authority_note, protocol_version, manifest) }'
  jsonrpc_errors: JSON-RPC 2.0 error objects inside the MCP/A2A envelopes (e.g. -32004 on /a2a for an unsupported method)
  detail: errors/afmr-ai-problem-types.yml
rate_limits:
  documented: false
  headers_observed: none (no RateLimit-*, X-RateLimit-* or Retry-After on any response)
  detail: rate-limits/afmr-ai-rate-limits.yml
claim_boundary: >-
  The provider repeats one convention across every document: discovery is not admission, admission
  is not trust, trust is not transaction authority. Agents should treat every response as a bounded
  fact about a signed record or a published document, never as a safety or capability endorsement.