What you will be able to do
- Handle conflicting statistics from credible sources by recording both values with their sources instead of picking one
- Require publication or data-collection dates in subagent output so a difference over time isn't mistaken for a contradiction
- Structure a synthesis report that separates well-established findings from contested ones and presents each type of content in a fitting format
1.Conflicting statistics: annotate, don't pick
Anthropic's research-agent cookbook describes good research agents as "following promising leads, synthesizing conflicting sources, and knowing when they have enough information". Conflict between sources is normal. The design question is what happens to it.
The exam guide's answer is clear. When credible sources disagree, the subagent completes its analysis with both values included. Each value keeps its source attribution, and the conflict is explicitly annotated. The subagent does not choose. Reconciliation happens at the coordinator, which can see every subagent's results and decide whether to look into it further, before anything reaches final synthesis.
The provided sources don't spell out this exact procedure. They do support its parts. For complex research, the prompting guide says to "develop several competing hypotheses" and to "Track your confidence levels in your progress notes to improve" calibration. The hallucination guide gives the underlying reason: "Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information." A subagent that quietly picks one of two credible figures gives a false sense of certainty. The coordinator never learns the number was disputed.
Checkpoint 1 of 4· Check yourself
A document-analysis subagent finds that two credible sources report different values for the same statistic. Which output best supports correct reconciliation?
Keeping both values with attribution lets the coordinator, which has the full picture, decide. Choosing one value or dropping both hides the uncertainty instead of reporting it.
“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”Source: platform.claude.com
2.Dates turn apparent contradictions into trends
Many apparent conflicts are not conflicts at all. A figure from a 2019 study and a different figure from a 2024 study may both be correct. The world changed in between. If the subagent's output leaves out when each source was published or when its data was collected, the coordinator can't tell a real disagreement from a change over time. It may then report a contradiction that doesn't exist, or average two numbers that describe different years.
The fix belongs in the subagent's output contract. Each claim-source record should carry a publication date or data-collection date, next to the URL and excerpt. The provided sources don't mention dates specifically, but they give you the tools to require them. The prompting guide recommends that each document carry "<source> (and other metadata) subtags", and a date is metadata. In the Citations API, the context field holds metadata the model reads but never cites from. To make the field impossible to skip, use structured outputs. They exist to prevent problems such as "Missing required fields", and a date marked as a required field will always be present.
| Field in subagent output | What breaks downstream without it |
|---|---|
| Source URL or document name | The coordinator can't attribute the claim, and the final report can't cite it |
| Supporting excerpt (quote) | The claim can't be checked against the text that supposedly backs it |
| Publication or collection date | Differences over time look like contradictions between sources |
| Conflict annotation | A disputed value reaches synthesis looking settled |
Checkpoint 2 of 4· Check yourself
You want every claim a subagent returns to include its source's publication date, with no exceptions. What is the most reliable way to enforce that?
Structured outputs use constrained decoding to guarantee required fields, so no claim can come back without a date. A prompt instruction improves the odds but doesn't guarantee it.
“Structured outputs constrain Claude's responses to follow a specific schema, ensuring valid, parseable output for downstream processing.”Source: platform.claude.com
Checkpoint 3 of 4· Exam question
A market-research pipeline dispatches five research subagents, each of which reads several documents and returns a free-form paragraph summary to a coordinator agent. When the coordinator assembles the final report, reviewers find that several claims can no longer be traced to any specific source. Which change to the subagent output contract would best prevent this loss of provenance?
Correct answer: A — Require each subagent to pair every claim with its source URL or document name and a supporting excerpt
- A. Correct. A structured claim-to-source mapping ties each individual assertion to its origin, so when the coordinator merges outputs from multiple subagents it can preserve attribution instead of losing it during re-summarization.
- B. A trailing bibliography lists what was consulted overall but does not link any single claim to the specific source it came from, so attribution is still lost once claims are merged.
- C. Temperature affects the randomness of generated text, not whether structured source metadata is captured, so this would not restore per-claim provenance.
- D. Shortening the paragraph limits how much a subagent can say but does nothing to attach individual claims to their sources, so provenance can still be lost when paragraphs are combined.
3.Structuring the report: established, contested, and fit-for-purpose formats
The final report is where everything kept so far either shows up or gets flattened. The exam guide calls for explicit sections that separate well-established findings, backed by several agreeing sources, from contested ones, where credible sources disagree. Within each section, keep how the sources themselves describe their findings and the methods behind them. For example, "a small pilot study suggests" should not be rewritten as "research shows".
The sources support the discipline behind this structure, though not the section layout itself. The hallucination guide's verification step applies directly. After drafting, check each claim against a supporting quote, and if none exists, "remove that claim from the press release and mark where it was removed with empty [] brackets." In other words, a claim the evidence can't support is either removed or visibly marked, never left in as if it were settled. Tracking confidence while researching, as the prompting guide recommends, is what gives you the information to sort findings into the established and contested groups.
The exam guide also asks you to present each type of content in a format that fits it: financial data as tables, news as prose, technical findings as structured lists. Don't force everything into one format. The provided sources don't address output formatting for synthesis reports, so this lesson can only state the requirement. The reason behind it matches the rest of this page, though. A uniform prose summary can bury the structure of a financial comparison in the same way it buries a source citation.
Checkpoint 4 of 4· Check yourself
During the final verification pass on a synthesis report, the agent finds a claim that no source excerpt supports. Following Anthropic's hallucination guidance, what should happen?
The guide's verification technique says to retract any claim without a supporting quote and mark the removal, so the reader can see something was taken out.
“If it can't find a quote, it must retract the claim.”Source: platform.claude.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.When two credible sources disagree, the subagent should pick the more authoritative value so the coordinator gets one clean answer.Why is that wrong?
Picking one value hides real uncertainty. Report both with their sources, annotate the conflict, and let the coordinator decide. Research guidance favors keeping competing hypotheses open and admitting uncertainty.
Covered in Conflicting statistics: annotate, don't pick
2.A claim the agent believes is true can stay in the final report even without a supporting source, as long as the wording is hedged.Why is that wrong?
Under the verification technique, a claim with no supporting quote is retracted, and the spot where it was removed is marked.
Covered in Structuring the report: established, contested, and fit-for-purpose formats
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/cookbook/claude-agent-sdk-00-the-one-liner-research-agentSecondary source
“following promising leads, synthesizing conflicting sources, and knowing when they have enough information”
↩︎ Conflicting statistics: annotate, don't pick - 2.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Encourage source verification: Ask Claude to verify information across multiple sources.”
↩︎ Conflicting statistics: annotate, don't pick“Track your confidence levels in your progress notes to improve”
↩︎ Conflicting statistics: annotate, don't pick“<source> (and other metadata) subtags”
↩︎ Dates turn apparent contradictions into trends“Track your confidence levels in your progress notes to improve”
↩︎ Structuring the report: established, contested, and fit-for-purpose formats - 3.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Conflicting statistics: annotate, don't pick“remove that claim from the press release and mark where it was removed with empty [] brackets.”
↩︎ Structuring the report: established, contested, and fit-for-purpose formats“If it can't find a quote, it must retract the claim.”
↩︎ Structuring the report: established, contested, and fit-for-purpose formats“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Exam trap 1“If it can't find a quote, it must retract the claim.”
↩︎ Exam trap 2 - 4.
“title and context are optional fields that are passed to the model but not used toward cited content.”
↩︎ Dates turn apparent contradictions into trends - 5.
“Missing required fields”
↩︎ Dates turn apparent contradictions into trends“Structured outputs constrain Claude's responses to follow a specific schema, ensuring valid, parseable output for downstream processing.”
↩︎ Checkpoint