ai-safe2-framework

Deployment Proof & Readiness

Human name: Deployment Proof & Readiness

Technical name: DeploymentReadinessAuthority

Profile: CP.5.APAY production-admission extension

Status: Draft reference standard; not a production assurance claim

Purpose

Component tests and a successful sandbox run do not, by themselves, prove that the exact production deployment is safe to activate. Deployment Proof & Readiness converts authenticated operational evidence into one short-lived, signed, fail-closed decision that a human or agent can verify without trusting the agent requesting activation.

The authority does not deploy software, move money, or turn a rail on. It sets the maximum permitted mode. A failed assessment permits only disabled; a complete assessment may permit enforced, subject to separate deployment change control.

Normative requirements

ID Required outcome
DPR-01 Evidence binds the exact deployment, environment, software, configuration, policy-parity, and sandbox-profile digests.
DPR-02 The complete execution plane reports deployment assurance with no missing or reference-only component.
DPR-03 Sandbox Rail Readiness independently permits enforced mode and its exact report is bound into the evidence set.
DPR-04 Every required control has exactly one explicit passing result; missing, duplicate, conflicting, malformed, or false evidence fails readiness.
DPR-05 Evidence is fresh, authenticated by a deployment-assured verifier, and produced by allowlisted assessors.
DPR-06 At least two distinct assessors contribute; assessors cannot also be deployment approvers.
DPR-07 Accountable and security owners are named and separated; data ownership and jurisdiction are declared.
DPR-08 Operational proof covers state and evidence restore, evidence completeness, revocation, ambiguous recovery, bypass prevention, monitoring, and incident escalation.
DPR-09 Evidence lifecycle proof covers export, deletion, and legal hold.
DPR-10 Fresh authenticated approval binds an allowlisted change approver to the exact profile and enforced mode; an independently protected authority signs the result.
DPR-11 A refusal is signed as evidence; an unavailable signer never yields an unsigned readiness decision.
DPR-12 Decisions are short-lived and expose digests and findings, not payment payloads, credentials, or account details.

Required control identifiers

DEFAULT_DEPLOYMENT_CONTROLS defines the baseline set. Operators may replace it with a stricter governed set, but must not silently remove controls from a previously approved profile. Profile changes create a new profile digest and invalidate old evidence.

The baseline covers execution readiness, live policy parity, Key Guardian isolation, independent runtime verification, state and evidence recovery, evidence completeness, revocation and ambiguous-state drills, sandbox readiness, network bypass prevention, monitoring and escalation, evidence export/deletion/legal hold, data residency, and accountable ownership.

Verification flow

  1. Construct a versioned DeploymentReadinessProfile with content-addressed identities, owners, jurisdiction, trusted assessors, and approvers.
  2. Collect one authenticated DeploymentControlEvidence item per required control. Evidence contains results and digests only.
  3. Supply the current GatewayReadiness and SandboxReadinessReport objects.
  4. Obtain DeploymentApprovalEvidence through a separately authenticated human change-control path. It must bind the exact deployment, environment, readiness-profile digest, enforced mode, validity window, and approver.
  5. Call DeploymentReadinessAuthority.assess(...) with that approval evidence.
  6. Independently verify the returned decision digest, proof, expiry, profile digest, mode, and findings before any activation mechanism can proceed.
  7. Reassess after software, configuration, policy, sandbox profile, ownership, jurisdiction, trust roots, or required controls change.

Security, privacy, governance, and sovereignty

The profile makes agent self-approval, approver-name replay, evidence replay across deployments, single-assessor promotion, and substitution of a different gateway or sandbox report deterministic failures. It leaves the operator in control of assessors, approvers, owners, jurisdiction, retention, trust roots, and activation.

Reports intentionally retain only identities, timestamps, content digests, control findings, and a decision proof. Payment payloads, unrestricted keys, credentials, and account details remain outside the readiness artifact. A deployment must separately protect evidence at rest, signing keys, verifier channels, and the change-control mechanism.