Sources and conflicts
A research system gathers facts from many sources, then summarises them into one report. Somewhere in that summarising, the link between each claim and its source can quietly fall away, and two sources that disagree get settled by picking one. The trap options all produce a tidy report. The right one keeps every claim checkable and shows a disagreement as a disagreement.
D5 Context Management & ReliabilityExam guide 5.6About 15 minutes
Start here: two newspapers and a rally
And a crowd of 8,000 in 2019 and 20,000 in 2024 is not a disagreement at all. It is two different years. Without the dates, you can't tell the difference.
A multi-agent research system has the same risk at every hand-off. Subagents read sources and pass findings to a synthesis step, which merges them into one story. If findings travel as prose, the merge compresses them, and the sources drop out. The fix is a claim-source mapping: each finding travels as data (the claim, its source and its date), so every step can pass the links on. When figures conflict, the report shows each one with its source. Dates sit next to every figure, so a change over time isn't mistaken for a conflict. And each kind of content is shown in its natural form: figures in a table, news in prose.
- Claims lost their sources somewhere: pass a claim-source mapping (claim, source, date) as data through every step, so citations are carried, not reconstructed.
- Two sources give different figures: show both, each with its source, date and method. Don't pick one, and don't average them.
- The figures come from different years: require publication and collection dates, so a change over time isn't read as a conflict.
- Mixed content: render figures as tables and news as prose.
Words in the question that tell you which side you're on
Keep the claim-source mapping
- “lost track of which source”
- “can't cite”, “cited the wrong paper”
- “the summary dropped the references”
- “a reviewer needs to verify”
Show the conflict, with dates
- “two reports disagree”, “different figures”
- “kept only one figure”
- “an older study”
- “published in different years”, “a decade apart”
Try it: write the report
The question is: what share of the (made-up) town of Riverton's workers work from home at least two days a week? Pick the sources your subagents found, then pick how synthesis turns them into a report. The simulator writes the final report and marks each line: can a reader check it? The diagram above highlights where your choice acts.
The sources going in
The final report
The town, sources and figures are made up for this lesson (illustrative). The exam guide's point is that attribution is lost when summaries compress without claim-source mappings, that conflicting figures are annotated with their sources rather than settled by picking one, and that dates keep a change over time from reading as a contradiction.
Things to try: different figures + plain summary to see a number nobody measured, then + keep the most authoritative to watch 41% vanish, then + claim-source mapping. Then different years + show side by side: the dates are printed, yet it still calls a trend a conflict. Last, sources agree + keep the most authoritative: when it does no harm.
The four traps
| Trap | What the wrong option looks like | Why it loses |
|---|---|---|
| Picks one | “Use the most recent figure and drop the older one”, “keep the stronger study, footnote the other” | It works when the kept figure is right, but it hides a real disagreement the reader should see. |
| Invents a figure | “Average the two conflicting figures”, or a summary that blends them | It gives one number, but that number appears in neither source. |
| Too late | “Ask the synthesis agent to add citations at the end”, or have the report step match sentences to sources | Once summaries have compressed the content, the links between claims and sources are already lost. Matching afterwards is guessing. |
| Loses provenance | “Have each subagent return a prose summary of its findings” | Prose reads well, but without structured source fields, attribution drops out during synthesis. |
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 5.6: “Source attribution is lost when summarization compresses without claim-source mappings.” “Preserve structured claim-source mappings through synthesis; annotate conflicting statistics with source attribution (don't arbitrarily pick one).” “Require publication/collection dates so temporal differences aren't misread as contradictions; render content types appropriately (financial as tables, news as prose).”