Direct answer: When evidence is ambiguous, the system should preserve the competing interpretations, identify the missing fact and stop any affected decision that cannot be made safely. It may request a clearer source, master-data correction, Rulepack decision or engineer review. It should not hide uncertainty by selecting the most fluent interpretation. Unaffected obligations may continue only if their independence is explicit.
How the controlled workflow works
- Detect conflicting tags, crossings, states, part identities, interface candidates or geometry references.
- Attach each candidate interpretation to its source evidence and state which downstream obligations it affects.
- Apply a declared policy: resolve from authoritative data, ask for review, fail closed or return cannot verify.
- Record the human decision and its provenance before resuming dependent execution.
What evidence should a reviewer inspect?
| Review item | What it can establish |
|---|---|
| Ambiguity record | Names candidates, supporting evidence and the reason no unique answer is justified. |
| Impact graph | Identifies which placements, routes, mates or verdicts depend on the unresolved fact. |
| Resolution record | Shows the authoritative source or named engineer decision used to continue. |
Where this answer stops
Confidence thresholds alone do not make an engineering decision safe. The refusal policy must reflect consequence, evidence quality and the customer’s acceptance rules.
Human engineering authority
The customer identifies who may resolve each class of document, part, interface and design ambiguity.
Related ways people ask this question
- What should AI do when a P&ID is unclear?
- How should a CAD agent handle ambiguous geometry?
- Should engineering AI guess missing information?
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 Happens When P&ID or CAD Evidence Is Ambiguous?.” Engineering Questions Q19. Reviewed 2026-08-15. https://mst-us.ai/questions/ambiguous-pid-cad-evidence/
NEXT STEP · NON-CONFIDENTIAL FIRST
Test a fail-closed edge case
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.