I would ask to see the two actual records. “Indexed most recently” is the time our search pipeline processed a record. It is not the time the policy was approved, published, or became effective. A nightly sync of an old handbook can give it the newest index timestamp in the system.

Suppose the policy service says the new refund rule takes effect on 1 July. A support wiki copied the draft last week and says 5 July. It is tempting to sort by timestamp and show one. The real task is to establish which record is allowed to speak for the policy, which revision it represents, which customers or regions it applies to, and whether its effective interval covers the user's question.

I would model at least these fields on a retrieved record: source system and owner, stable policy ID, revision, scope, publication or approval status, effective start and end, extraction time, index time, and derivation link if it was copied from another record. The authority rule is part of the product's governance contract. For this policy family, the approved record in the policy service may outrank the support wiki. For a different family, the support wiki may be the only maintained source. The system should not learn that hierarchy solely from whichever text sounds more official.

That is also why deduplicating the two passages by embedding similarity is dangerous. The 1 July and 5 July versions are almost identical in meaning to a vector search system. Their disagreement is exactly the fact the user cares about. Keep both identities through ranking and context assembly. A copied draft should carry its source lineage so the assistant can say “this is a copy of revision 7” rather than present it as an independent second vote. The W3C PROV vocabulary is a useful model for recording derivation and primary source relationships. It does not decide which corporate policy has authority. The organization must do that.

My answer path would fetch candidate policy records by exact policy ID, region, and applicable time, then check approval state and authority. If the authoritative revision is available and valid, answer from that revision and cite it. If two approved authoritative records conflict for the same scope and time, stop and show the conflict to the policy owner. A fluent compromise date would be a fabricated policy.

The interviewer now removes the policy service for ten minutes. We need to distinguish failure to query the source live from lack of any verified source. If a signed or otherwise trusted last-known revision with an acceptable freshness bound is stored locally, and there is a reliable invalidation path, I may answer with that revision and say when it was last verified. The exact freshness limit depends on how often emergency policies change and the consequence of being wrong. If we cannot establish that the cached revision remains valid, the wiki copy does not gain authority merely because it is reachable. For a high-impact decision I would say the current policy cannot be confirmed and route to a human or a live policy channel.

They add one more fact: the source system often publishes a correction with an effective date in the past. Now a single “latest” field is still insufficient. A question asked today about a transaction on 2 July may need the revision that was effective for that transaction, along with the correction's publication time. I would store both the time the rule applied to the world and the time our system learned it. Historical answers can be replayed against the knowledge that existed then, while current answers use the corrected record if the policy owner says the correction is retroactive. That retroactivity is a domain decision, not a search ranking trick.

I would test the whole path with a duplicated draft indexed later than an approved policy, an emergency revocation, two overlapping approved revisions, and a source outage. The expected output for some cases is “I cannot determine the current rule.” A system that always chooses one citation may look decisive in a demo and be worse for the person making a real policy decision.