openapi: 3.1.0 # RST ATLAS — the four accountability operations, as deployed. # # Written FROM the deployed router (bastion_atlas_accountability_router_20260920.py), not from memory or from a # design document. Every status code, header and field below is one the live service actually returns. A contract # that describes an intended API is worse than no contract: an integrator discovers the difference at runtime, # and a coding agent discovers it in production. # # WHY THIS FILE EXISTS. Enterprise buyers do not publish doors to vendors — verified twice this week against two # of the largest US insurers, whose only published routes are procurement databases. Developers and agents, # however, find services by reading contracts. This is the door that does not require someone's permission. # # WHAT THIS IS NOT. Atlas is not an A2A-conformant agent and this file does not claim to be an Agent Card. A2A # requires a server implementing message:send and the task lifecycle; Atlas exposes four REST operations. An # /.well-known/agent-card.json declaring a transport Atlas does not speak would send every A2A client into a # failure on first contact. That adapter is real work and is not pretended here. info: title: Reliant Scale Atlas — Agent Accountability version: "1.0" summary: Record what an agent did, what it relied on, what it was allowed to do, and what happened afterward. description: | Atlas records consequential agent activity and makes it accountable after the fact. Four operations and a scoped bearer credential. The surface is small deliberately: a service holding other organisations' accountability records should not have a large attack surface. Behaviour worth checking, because it is the behaviour that matters when something has gone wrong: - **Revocation takes effect on the very next request.** There is no session and no cache; every request re-reads the credential's state, so a revoked key stops working immediately rather than at the end of a token lifetime. - **A denial is accountable work.** An action refused for being outside its scope still produces a receipt and is still metered. A prevented unauthorised action is the service doing its job. - **The record outlives the credential.** Receipts survive revocation and termination. - **Retrying is safe and is not billed twice.** The same `Idempotency-Key` returns the same receipt. - **Reads are never charged**, so inspecting your own record does not create a reason to inspect less of it. - **A receipt belonging to another organisation returns 404, not 403**, so a stranger learns nothing from the difference. Replay is FORENSIC RECONSTRUCTION of the stored decision-time envelope. It does not re-execute a model and cannot tell you what a model would output today. The response says so in its own `replay_class` field. Atlas is accountability infrastructure. It is not a compliance certification for any regulation. contact: name: Reliant Scale Technologies url: https://reliantscale.com/atlas/integrate email: partnerships@reliantscale.com servers: - url: https://api.reliantscale.com description: Production. Controlled enterprise pilot; no public signup or self-service credential. security: - bearerAuth: [] tags: - name: accountability description: The four operations. paths: /v1/atlas/actions: post: tags: [accountability] operationId: submitAccountableAction summary: Submit an accountable action description: | One agent action through the governed lifecycle: the evidence it relied on, the action proposed, the authority evaluation, the outcome, and a receipt. `Idempotency-Key` is REQUIRED — without it a retry could be billed twice, so the service refuses rather than guessing. A repeated key returns the existing receipt with `billed: false`. A denial returns **403 with a full receipt body**, not an error. The refusal is the accountable event. parameters: - name: Idempotency-Key in: header required: true schema: { type: string } description: Required. A retry with the same key returns the same receipt and is never billed again. requestBody: required: true content: application/json: schema: type: object required: [action, workspace_id] properties: action: { type: string, description: The action class being proposed. } workspace_id: { type: string } evidence_ref: { type: string, description: Reference to the evidence relied on at decision time. } observed_outcome: { type: string } responses: "200": description: Allowed. A receipt was created and the action is accounted for. content: application/json: schema: { $ref: "#/components/schemas/Receipt" } "400": description: Missing `Idempotency-Key`, missing `action` or `workspace_id`, or a body that is not JSON. content: { application/json: { schema: { $ref: "#/components/schemas/Error" } } } "401": description: Credential not recognised. An unknown key and a revoked key are refused identically. content: { application/json: { schema: { $ref: "#/components/schemas/Error" } } } "403": description: | Denied — and this carries a FULL RECEIPT, not an error. The action was outside the credential's scope, which is itself accountable work and is metered. content: { application/json: { schema: { $ref: "#/components/schemas/Receipt" } } } "409": description: That idempotency key belongs to another organisation. content: { application/json: { schema: { $ref: "#/components/schemas/Error" } } } "413": description: Request body too large. content: { application/json: { schema: { $ref: "#/components/schemas/Error" } } } /v1/atlas/actions/{key}: get: tags: [accountability] operationId: readReceipt summary: Read the receipt description: | The accountability record for one action: decision, reasons, evidence reference, observed outcome, and which principal and credential performed it. **Reads are never charged.** parameters: - name: key in: path required: true schema: { type: string } description: The idempotency key the action was submitted with. responses: "200": content: { application/json: { schema: { $ref: "#/components/schemas/Receipt" } } } description: The accountability record. "401": { description: Credential not recognised. } "404": description: | Not found — returned both for a key that does not exist AND for one belonging to another organisation. Deliberate: a 403 would confirm the record exists to someone not entitled to know. /v1/atlas/actions/{key}/replay: get: tags: [accountability] operationId: replayAction summary: Replay the decision-time envelope description: | What was knowable when the decision was made, and what has changed since. The decision is reported as it was and is never re-decided against today's state. The response carries `replay_class: FORENSIC_RECONSTRUCTION`. The classification is in the payload rather than in the documentation, so neither an integrator nor we can mistake what is being offered. Where nothing has changed, replay says so — "no later change has been admitted" is a finding. parameters: - name: key in: path required: true schema: { type: string } responses: "200": content: { application/json: { schema: { $ref: "#/components/schemas/Replay" } } } description: The reconstructed decision-time envelope. "401": { description: Credential not recognised. } "404": { description: Not found, including across an organisation boundary. } /v1/atlas/whoami: get: tags: [accountability] operationId: whoami summary: Ask what a credential may do description: | The principal, its organisation, and the scopes on the presented credential. **Scopes are an upper bound, not a grant** — what is permitted and what is authorised for a given action are evaluated separately. responses: "200": content: { application/json: { schema: { $ref: "#/components/schemas/WhoAmI" } } } description: The presented credential's identity and scopes. "401": { description: Credential not recognised. } /v1/atlas/health: get: tags: [accountability] operationId: health summary: Liveness security: [] responses: "200": { description: The service is reachable. } components: securitySchemes: bearerAuth: type: http scheme: bearer description: | A scoped, revocable credential presented as a bearer token, stored as a hash and re-read on every request. Not OAuth2 and not a JWT. Credentials are issued by Reliant Scale with the customer; there is no public signup and no self-service issuance. schemas: Receipt: type: object properties: idempotency_key: { type: string } receipt_id: { type: string } decision: { type: string, enum: [ALLOW, DENY] } reasons: type: array items: { type: string } description: Why the authority evaluation reached this decision. evidence_ref: { type: string, description: The evidence available at decision time. } observed_outcome: { type: string } billed: type: boolean description: False on a retry returning an existing receipt, and false for reads. replay: { type: string, description: Path to the replay for this action. } Replay: type: object properties: replay_class: type: string const: FORENSIC_RECONSTRUCTION description: | Stated in the payload on purpose. Atlas reconstructs the stored decision-time envelope; it does not re-execute a model, and makes no claim about what any model would produce today. decided_at: { type: string, format: date-time } knowable_at_decision_time: { type: object } changed_since: type: array items: { type: string } description: Named explicitly. An empty list means no later change has been admitted, which is itself a finding. WhoAmI: type: object properties: principal_id: { type: string } principal_type: { type: string } organization_id: { type: string } scopes: type: array items: { type: string } description: An upper bound on what this credential may attempt, never a grant. Error: type: object properties: error: type: string description: | Deliberately uninformative on authentication failure: an unknown credential and a revoked credential are refused with identical text, so neither teaches an attacker which keys once existed.