Skip to the content.

Policy Commitment Attestation

A specification for verifiable AI agent commitments to governance statements.

flowchart LR
  Agent((AI agent)):::agent
  Policy[Governance statement]:::policy
  Cred[PolicyCommitmentCredential<br/>T1-T6 + evidence + scope]:::cred
  Verifier:::verifier

  Agent -- "commits at tier N" --> Cred
  Policy -- "version + hash" --> Cred
  Cred -- "signed W3C VC<br/>resolvable evidence" --> Verifier

  classDef agent fill:#fff4e0,stroke:#d97706,stroke-width:2px,color:#451a03
  classDef policy fill:#e8f4ff,stroke:#2563eb,color:#0b1a36
  classDef cred fill:#e6ffed,stroke:#16a34a,stroke-width:2px,color:#052e16
  classDef verifier fill:#f3e8ff,stroke:#7c3aed,color:#2e1065

What PCA is

The Commitment Maturity Ladder

Tier Name Meaning Floor evidence
T1 Read Agent knows the statement exists and is bound by it Agent DID + timestamp
T2 Understood Agent can paraphrase, cite, surface in reasoning Self-explanation; similarity ≥ 0.8
T3 Adopted Statement lives in the agent’s working memory Memory file, AGENTS.md fragment
T4 Codified Durable repository artifact carries the commitment ADR, skill, PR commit
T5 Bounded Scope constraint / refusal rule applied ODRL prohibitions
T6 Enforced Runtime guardrail blocks violations Hook, middleware, tool allowlist

Tiers are cumulative — T6 includes T5 includes T4… all the way down.

Get started

Reference implementation

smoke-test.mjs in the Dictiva repo is a reproducible end-to-end issue + verify round-trip. Clone, cd scripts/attestix-spike, npm install, npm run smoke.

Status

Version: v0.1 draft · Stage: pre-AAIF project proposal · License: Apache 2.0 · Stewardship: Dictiva, targeting Linux Foundation AAIF contribution.

Feedback welcome

File issues on the repository. Review especially wanted from the W3C VC community, in-toto maintainers, the MCP working group, and AAIF TSC members.