Two true source passages do not automatically make their combined conclusion true. Multi-document knowledge answers often fail in the join: they connect people, dates, products, or policies that merely have similar names.
Reviewers need to inspect the reasoning boundary between documents. Counting citations is a weak proxy because an answer can attach several relevant links to a conclusion none of them establishes.
Map claims before judging the paragraph
Split the answer into factual claims and conclusions. For each claim, note the supporting span, document revision, applicable organisation, time range, and any conditions. Then identify the joins: which shared identifier or explicit relationship allows two passages to be combined?
If the join rests on an assumption, make it visible. “These appear to describe the same plan” is different from “this customer is covered.” The former reports uncertainty; the latter may invite an operational decision.
Review a fictional support entitlement answer
Consider three invented documents. A plan page says Premium includes weekend support. A customer record says Cedar is on Premium Legacy. A migration memo says some Premium Legacy accounts will move to Premium after renewal.
The question is whether Cedar has weekend support today. A response that combines the first two pages and says yes has ignored the distinction between plans. Adding the migration memo does not fix the answer unless Cedar's renewal and migration status are established.
Build a review table with rows for plan identity, customer entitlement, migration condition, renewal date, and the final yes-or-no conclusion. The answer should explain which facts are known and identify the missing entitlement evidence. If a source explicitly confirms current weekend access, the conclusion becomes supportable.
Now change only the customer name or renewal date. The conclusion should follow the changed evidence rather than remain attached to a familiar plan label. This counterexample tests whether the assistant used the relationship or merely repeated a plausible sentence.
Repair the unsupported join at the right layer
Missing identifiers may require better source metadata. Missing passages may require retrieval changes. An answer that sees the correct distinction but discards it needs a synthesis or instruction change. A reviewer should not prescribe reindexing until the trace shows where the evidence disappeared.
Requiring explicit joins can produce more qualified answers. That is a practical tradeoff when employees want a quick decision. Offer a concise conclusion when supported and a concrete next source to consult when it is not. Lengthy uncertainty language without an actionable gap is also a usability failure.
This proposed exercise complements citation review. The Nucleus project description supplies related portfolio context without implying that this fictional entitlement case occurred there.
Questions about multi-document evidence
Can an answer make any inference?
Yes. Label the inference and show the premises. Decisions that depend on an unverified identity or missing date should remain conditional.
Does every sentence need a citation?
Every material factual claim needs traceable support. One citation can support several closely related claims if its scope is clear.
What should an evaluation failure include?
Save the unsupported conclusion, the proposed join, and the smallest source change that would make it valid. That packet makes a knowledge review more productive than a screenshot alone.