Direct answer: `CANNOT_VERIFY` is a bounded verdict: the system cannot establish pass or fail for a declared obligation using the available authorized evidence and supported checks. It is different from failure, which means evidence shows the obligation was not met. The verdict should name the missing evidence, unsupported test or conflicting state and identify the next responsible reviewer or input.
How the controlled workflow works
- Name the exact obligation under review and the evidence that would be required for pass or fail.
- Check whether the evidence is missing, unauthorized, inconsistent, outside tool scope or not testable by the configured verifier.
- Return a structured reason code and safe next action without converting uncertainty into a positive result.
- Re-evaluate only after a controlled source, rule, tool or engineer decision changes the evidence state.
What evidence should a reviewer inspect?
| Review item | What it can establish |
|---|---|
| Obligation and reason code | Makes the non-verdict specific, searchable and operationally actionable. |
| Evidence-gap statement | Explains what fact or supported test is missing and why it matters. |
| Disposition and owner | Routes the case to new input, engineering review, unsupported scope or product repair. |
Where this answer stops
CANNOT_VERIFY must not be used to hide a known failure, and a high rate of avoidable non-verdicts may reveal poor master data or product limitations. Both require measurement.
Human engineering authority
The customer defines whether a non-verdict blocks the workflow and who may supply or approve the missing evidence.
Related ways people ask this question
- What does CANNOT_VERIFY mean?
- Why would an AI engineering system refuse to answer?
- Is cannot verify the same as failure?
References and claim boundary
- NIST AI 100-1 — Artificial Intelligence Risk Management Framework
- W3C Recommendation — PROV-DM provenance data model
- MST Engineering AI — Evidence model and public status
Limit: External sources support the named standards, platform features or governance concepts. They do not endorse MST or independently validate MST product performance. MST implementation statements remain bounded by the current public evidence status.
MST Engineering AI. “What Does CANNOT_VERIFY Mean in Engineering AI?.” Engineering Questions Q20. Reviewed 2026-08-15. https://mst-us.ai/questions/cannot-verify-engineering-ai/
NEXT STEP · NON-CONFIDENTIAL FIRST
Review a cannot-verify policy
Describe the workflow and decision boundary without uploading protected P&IDs, CAD, BOM, PDM data, credentials or customer identities. Confidential exchange begins only after identity, NDA and scope review.