A multi-agent conclusion should be traceable to its evidence, not merely to the last agent that repeated it. Preserve the lineage of consequential claims so reviewers can find unsupported upgrades and shared-source errors.
Conclusion provenance connects claims to sources
A conclusion audit records how evidence and intermediate interpretations support the final statement. Relevant dimensions include source identity, version, claim scope, verification status, and any transformation through summaries or specialist analysis.
For a proposed evidence graph, use minimal structured references rather than copying every message. Distinguish original observations, retrieved policy, caller reports, tool results, and agent inference. The same source repeated by three agents remains one source.
Do not treat a chain of internal citations as proof of correctness. Review whether the primary evidence actually supports the claim and whether the approved authority rule applies to the question.
Audit a fictional service-eligibility conclusion
A final answer says an address is covered. One agent cites a summary from another, which cites an old service-area document. The current approved policy excludes that district. The audit should reveal the stale source and the point where version context was lost.
A corrected process links the final eligibility claim to the current policy decision and relevant address interpretation. If the address itself is ambiguous, the final answer should retain that uncertainty rather than rely on a clean-looking evidence chain.
| Provenance question | Evidence needed |
|---|---|
| Where did the claim originate? | Primary source or observation reference |
| Which version applied? | Effective policy or record version |
| What was inferred? | Explicit interpretation category |
| Did agents share one source? | Source identity preserved across handoffs |
| What remained uncertain? | Unresolved condition carried forward |
Test with a valid chain, a stale source, an inaccessible source, and a citation that does not support the claim. Evaluate final meaning and the reviewer’s ability to reconstruct the basis.
Keep audits useful and private
Lineage metadata can expose relationships between customers, documents, and staff. Apply task-scoped access and avoid broad raw-context archives. A reviewer should retrieve only the evidence necessary for the authorised investigation.
Full provenance for every conversational phrase can be expensive. Prioritise claims affecting eligibility, authority, commitments, and external actions. Use sampled review for lower-impact explanation while keeping authoritative operation references directly linked.
This audit design is proposed. The voice integration case study describes retrieval and tool experience without claiming this evidence graph. Pair it with evidence sharing between agents.
Conclusion audit FAQ
Does agreement from several agents create independent evidence?
Not when they rely on the same source or repeat one another.
Should hidden model reasoning be the audit record?
Use observable inputs, source references, decisions, and operation outcomes instead.
What should happen when a source is missing?
State the evidentiary limit and avoid presenting the conclusion as fully verified.