Purpose of the Template
You'll soon encounter client evidence labeled "formally verified with AI assistance." A prime example is Google DeepMind's work on Signal. AlphaProof Code verified Signal's post-quantum PQXDH handshake and Triple Ratchet, with a team of five working for four weeks. However, they noted that the result assumes perfect cryptography and doesn't guarantee computational secrecy of bytes. Human reviews of the specification and flagged sections were also conducted.
This blend of strong claims and narrow scope is what you need to evaluate. This template is an evidence assessment worksheet for Assurance Engagements where AI-assisted formal verification supports a control. It complements your methodology by providing a consistent set of questions for your workpapers.
Prerequisites
Before using the worksheet, ensure you have:
- The control or criterion the client is mapping the verification to (e.g., a SOC 2 common criterion on system operations or change management, or an ISO/IEC 27001:2022 Annex A Technological Control such as 8.24 on cryptography or 8.28 on secure coding).
- The client's verification report and the formal specification itself, not just a summary.
- Named contacts who can explain who wrote and reviewed the specification.
- The exact version of the code that was verified, plus the production release history.
- Your firm's guidance on reliance on specialists, if you plan to involve a cryptography or formal methods expert.
The Template
Copy this into your workpapers. Complete one per verification claim.
Section A: Claim Definition
- Control or criterion supported: ______
- Verification claim as stated by the client (quote exactly): ______
- Component(s) in scope (protocol, library, module): ______
- Component(s) explicitly out of scope: ______
Section B: Specification Integrity 5. Who authored the specification? (Name, role, date): ______ 6. Who reviewed it, and what evidence shows the review happened? ______ 7. Does the specification capture the security goals the control cares about (confidentiality, authentication, forward secrecy, etc.)? Yes / Partial / No. Rationale: ______ 8. Was the specification written by humans, generated by AI, or both? If both, where is the human approval recorded? ______
Section C: Assumptions and Boundaries 9. List every stated assumption (e.g., "assumes perfect cryptography"): ______ 10. For each assumption, does it hold in the client's production environment? ______ 11. What properties does the verification NOT establish? ______ 12. Does the client's control narrative overstate what was proven? Yes / No. Detail: ______
Section D: Currency and Coverage 13. Verified code version/hash: ______ 14. Deployed code version/hash: ______ 15. Changes since verification and how they were assessed: ______ 16. Is re-verification triggered by change management? Reference procedure: ______
Section E: Human Oversight of AI Output 17. Which parts of the AI-generated proof or analysis were flagged for human review? ______ 18. Evidence the flagged items were resolved: ______ 19. Independent checker used (e.g., a proof kernel) and its version: ______
Section F: Auditor Conclusion 20. Does this evidence support the control, supplement other evidence, or fall outside the control's scope? ______ 21. Additional procedures performed (reperformance, expert consultation, inquiry): ______ 22. Conclusion and reviewer sign-off: ______
Customizing the Template
Section A, Scope. Ensure there's no gap between what was verified and what the control covers. A proof about a handshake protocol doesn't address key storage or access to signing systems. Add rows if one verification maps to several controls.
Section B, Specification. In the Signal work, security goals were turned into a human-reviewed mathematical specification before agents generated proofs. The hard part is deciding what to prove. Treat the specification as your primary audit object. If you can't read it, engage a specialist and document your reliance.
Section C, Assumptions. Tighten this section for cryptographic claims. If the client says "verified secure," question 11 should clarify the distinction between protocol-level guarantees and implementation-level or computational ones.
Section D, Currency. Verification applies to a specific artifact. For SOC 2 Type II periods, compare the verified version against what ran during the observation window. For ISO/IEC 27001, tie this to your review of change management and Control of Documented Information.
Section E, Oversight. The Signal team audited the generated Lean code and flagged parts for human review. Ask your client for the equivalent. If nothing was escalated to a person, ask why.
Scaling the Worksheet. For large codebases, don't complete this per function. Complete it per verified component and keep a register linking each entry to the relevant control.
Validation Steps
Before relying on the completed worksheet, test it:
- Reperform One Link. Pick one assumption from Section C and confirm independently that it holds in production.
- Trace the Version. Match the hash in question 13 to a release record. If you can't, don't rely on the evidence.
- Interview the Reviewer. Ask the named specification reviewer to explain, in plain language, what the specification does and doesn't say. Hesitation here tells you a lot.
- Check the Narrative. Compare the client's control description to your answers in Section C. Any overstatement should be communicated with management and may inform a Nonconformity or an exception, depending on your framework's rules.
- Document Consultation. If an expert assisted, record their qualifications and conclusions in line with ISO/IEC 27021 competence expectations where relevant to your engagement.
- Decide the Weight. AI-assisted verification can strengthen a design assertion. It doesn't replace operating effectiveness testing. State this in your conclusion.
Limits Worth Stating Plainly
AI may reduce the cost of verification compared to efforts like the seL4 microkernel proof, which took 20 person-years. But cheaper proofs make scope discipline more important. If proofs are easy to produce, your job shifts to asking whether the right thing was proven. This worksheet is built for that shift.





