FS AI RMF change-control checklist

Twelve questions on the SDLC slice FS AI RMF reviewers ask when Copilot, Cursor, or agent tools merge to protected branches. Email-gated results — voluntary readiness, not certification. Feedback welcome.

Voluntary self-assessment for SDLC and audit-trail themes only — not all 230 FS AI RMF control objectives.

1. 1. Can you attribute each merged PR to a declared human or agent actor?

   No systematic recordManual notes or git blame onlySigned metadata at commit time

2. 2. Are agent-assisted changes to protected branches blocked when provenance is missing?

   No CI gateWarn-only or partial coverageEnforced Scroll Gate on protected branches

3. 3. Can an assessor verify change evidence offline without trusting your platform?

   Screenshots or platform logs onlyExport exists but not independently verifiableOffline Ed25519 verify on exported pack

4. 4. Do you maintain a registry of approved agents/tools that may author production code?

   No registryInformal listDocumented registry in evidence export

5. 5. For payments or ledger paths, are changes gated before merge?

   Not applicable / no special controlsSame as general reposDedicated financial-infra policy preset

6. 6. Has a bank, auditor, or enterprise counterparty asked about AI-authored code evidence?

   Not yetAnticipated / planningYes — active questionnaire or sampling

7. 7. Can you export a machine-readable audit ledger for a review period?

   NoManual assembly from multiple toolsStructured JSON evidence pack

8. 8. Are post-commit edits that break envelope integrity detected in CI?

   NoSometimes / spot checksEvery PR in protected branch range

9. 9. Is agent tool identity (Cursor, Copilot, Claude Code, etc.) captured at commit?

   NoCo-author trailers or comments onlyDeclared tool field in signed envelope

10. 10. Do you map change evidence to FS AI RMF or PCI change-control language for procurement?

    NoAd hoc in security questionnairesStructured compliance_mappings in export

11. 11. For agent payment or spend mandates, is human approval recorded with offline verification?

    Not applicablePlatform logs onlyPiloting signed approval records

12. 12. Could you run warn-mode provenance on two protected branches within 30 days?

    Unknown / blocked by processPossible with engineering buy-inYes — ready to pilot