Skip to content
You are reading the Ratify Protocol v2 beta documentation. Protocol v1 remains available at the site root.

Deployment profiles

Every SDK defines the same storage and service interfaces:

Component Contract
Challenge issuance store Persist before expose; never overwrite; retain through expiry plus skew
Claim store and authoritative clock Atomic claim, evaluation progress, finalization, recovery, release record; one clock for the replay domain that fails closed if it moves backward
Rate reservation store All-or-nothing max_rate reservations committed or released with the terminal result
Receipt outbox and publisher Seal the receipt body at commit; append and sign in order; never re-sign a different body
Independent checkpoint Acknowledge log entries in order from an independent failure domain
Security journal store Load and compare-and-swap against an independently anchored head
Trust, status, and audience providers Caller-authenticated inputs; a failure never passes a check

For one device without connectivity, such as an edge controller or a vessel, that must keep releasing decisions. The device hosts the claim store, the receipt log, and the independent checkpoint, and the release rule is unchanged: commit, append, acknowledge, and only then act.

  • The checkpoint must sit in its own failure domain on the device: its own key, its own rollback-protected storage, and an anchored digest of the primary’s complete security state. A secure element or TPM-backed component can qualify if it authenticates both the sequence and that digest. A second process, container, VM, backup, plain counter, or log file never does.
  • Signed status is staged into the anchored journal before disconnecting. Offline, the device stops allowing at the first deadline: challenge expiry, certificate expiry, status next_update, the allow horizon, full checkpoint storage, or a clock fault.
  • After reconnection, receipts are uploaded in log order. Upload never releases anything.

Claim it as “conforms to the Ratify Protocol v2 single-device offline profile” (SPEC-v2 §7.13).

For actions that need approval from two independent roots, such as two organizations, or an organization and a separately pinned person. One verification trusts one root, so the profile composes two:

  1. The application records an approval set: the action’s request hash, the agent key, the verifier, and the two pinned root keys.
  2. It derives the expected session context from the set and issues two challenges bound to it.
  3. Verification A pins only root A and verification B only root B. Each must reach a released allow.
  4. An application ledger fills each root’s slot only after checking the released decision, the bundle, the action binding, and the root, and marks the action executable once both match.

One root never counts twice, and one released decision never counts toward two sets. The composite result is application state, not a Protocol receipt (SPEC-v2 §7.14).

For platforms, relays, or gateways that hold agent keys and sign presentations for the agents they run. Because the signer also routes the proof, every presentation must be session-bound: session_context covers the verifier, tenant, agent, session, invocation, and the exact operation, and the verifier compares it with the value stored at challenge issuance before claim.

The profile stops a signed proof from being moved to another tenant, session, action, or receiver. It does not protect against compromise of the custodian, which can sign new proofs; custody controls, short lifetimes, and revocation cover that (SPEC-v2 §7.15).

A deployment may claim a profile only when it meets every requirement of it. An SDK may claim support only when its tests run that profile’s shared vectors. The reference components in the SDKs are test-only models.