Skip to content
Cheat sheets

CCAR-F · Domain 5: Context Management & Reliability · 15% of the exam

Task 5.6: Preserve information provenance and handle uncertainty in multi-source synthesis

Claim-to-source mappings that survive summarization, conflicting figures reported side by side with their sources, dates that keep old and new data from looking contradictory, and formats that suit each kind of content.

Study it

Sample questions

Try each one before opening the answer. Every option is explained, with the Anthropic documentation page that proves it.

Question 1 · choose 1

Two reputable industry reports, both covering 2025, put the share of small businesses using AI tools at 31 percent and 38 percent. They used different survey methods. Your research system's synthesis agent must present this finding. What should it do?

  1. AReport only the higher figure, since the larger survey is probably the more accurate of the two reports
  2. BLeave the statistic out because the two sources disagree
  3. CReport the average of the two figures, 34.5 percent, as the best estimate of the true value
  4. DReport both figures with their sources and methods, marking the point as contested
Show the answer and why

  • AReport only the higher figure, since the larger survey is probably the more accurate of the two reports

    Incorrect

    Picking one value without saying so drops credible evidence and its attribution.

  • BLeave the statistic out because the two sources disagree

    Incorrect

    Omitting it hides relevant evidence; the disagreement is itself useful to the reader.

  • CReport the average of the two figures, 34.5 percent, as the best estimate of the true value

    Incorrect

    Averaging invents a number neither source reported and hides the disagreement.

  • DReport both figures with their sources and methods, marking the point as contested

    Correct

    Annotating the conflict with attribution keeps the evidence honest and lets readers weigh it.

When credible sources conflict, keep both values with source attribution and methodological context instead of choosing one arbitrarily, and distinguish contested findings from well-established ones.

Anthropic documentation

Task 5.6Report a problem with this question

Question 2 · choose 1

Your research reports mix settled findings and disputed ones in the same paragraphs, written in the same confident tone. Readers act on contested claims as if they were established. How should the synthesis agent structure its reports?

  1. AKeep the paragraphs, but add "reportedly" to every claim
  2. BKeep the paragraphs, but give each one a confidence score from 1 to 10
  3. CPut well-established findings and contested ones in separate sections
  4. DOrder the claims by how many sources mention each one
Show the answer and why

  • AKeep the paragraphs, but add "reportedly" to every claim

    Incorrect

    A hedge on every claim makes settled and contested findings look the same.

  • BKeep the paragraphs, but give each one a confidence score from 1 to 10

    Incorrect

    One score for a paragraph that mixes settled and contested claims describes neither, and readers still cannot see which claims are disputed.

  • CPut well-established findings and contested ones in separate sections

    Correct

    Explicit sections let readers see at a glance what is settled and what is still disputed.

  • DOrder the claims by how many sources mention each one

    Incorrect

    Repetition across sources is not agreement; several sources can repeat one contested claim.

Structure synthesis output so readers can tell settled from disputed findings: separate sections for well-established and contested findings, rather than one uniform narrative in which every claim reads the same.

Anthropic documentation

Task 5.6Report a problem with this question

Question 3 · choose 1

Your research synthesis answers twelve standard questions about each company. For some questions the sources say little, yet the synthesis still writes confident conclusions, and analysts have acted on them. Analysts want to know when the evidence is thin. What instruction does the guidance support?

  1. AExplicitly allow Claude to say that the evidence is insufficient
  2. BLower the temperature so that conclusions are more conservative
  3. CRequire a definite conclusion for every question so that the report is complete
  4. DDrop the questions with thin evidence from the report
Show the answer and why

  • AExplicitly allow Claude to say that the evidence is insufficient

    Correct

    Giving explicit permission to admit uncertainty reduces false statements and tells readers where evidence is thin.

  • BLower the temperature so that conclusions are more conservative

    Incorrect

    Sampling settings do not reflect how strong the evidence is.

  • CRequire a definite conclusion for every question so that the report is complete

    Incorrect

    That forces conclusions the evidence does not support.

  • DDrop the questions with thin evidence from the report

    Incorrect

    Readers then lose the information that the evidence is thin.

Allow Claude to say "I don't know" or that the information is insufficient. Explicit permission to admit uncertainty is one of the simplest ways to reduce confident but unsupported output.

Anthropic documentation

Task 5.6Report a problem with this question

Question 4 · choose 1

Your research subagent uses the web search server tool, and the final answer must point readers to the pages each statement came from. A teammate proposes building your own tracker that maps statements to search results. What does the tool already provide?

  1. AOnly the list of search results; mapping statements to pages needs your own tracker
  2. BCitations, but only when you turn them on with a flag on the tool
  3. CCitations to the result pages it drew on, with their URLs, titles and cited text
  4. DCitations only for pages it opened with a separate fetch tool, not for search results
Show the answer and why

  • AOnly the list of search results; mapping statements to pages needs your own tracker

    Incorrect

    The response already includes citations that tie the text to the result sources.

  • BCitations, but only when you turn them on with a flag on the tool

    Incorrect

    Citations are always enabled for web search.

  • CCitations to the result pages it drew on, with their URLs, titles and cited text

    Correct

    Responses that use web search come with citations to the sources drawn from the results.

  • DCitations only for pages it opened with a separate fetch tool, not for search results

    Incorrect

    Search results themselves are cited; no separate fetch is required.

The web search tool's responses include citations for the sources drawn from search results. They are always enabled, and each carries the URL, title and cited text, so no separate tracker is needed.

Anthropic documentation

Task 5.6Report a problem with this question

Question 5 · choose 1

For an internal due-diligence report, every statement must come from the documents provided, and the model's general knowledge must not slip in, even where the documents are silent. Which instruction does the hallucination guidance describe?

  1. AUse general knowledge only where the documents are silent, and mark those statements
  2. BUse only the provided documents, not general knowledge
  3. CCite at least three documents for every statement
  4. DLower the temperature so that the model sticks closer to the documents
Show the answer and why

  • AUse general knowledge only where the documents are silent, and mark those statements

    Incorrect

    Marked or not, statements from general knowledge are exactly what must stay out.

  • BUse only the provided documents, not general knowledge

    Correct

    External knowledge restriction keeps answers grounded in the supplied material.

  • CCite at least three documents for every statement

    Incorrect

    A count of citations does not keep outside knowledge out.

  • DLower the temperature so that the model sticks closer to the documents

    Incorrect

    Sampling settings do not restrict which knowledge the model draws on.

External knowledge restriction: explicitly instruct Claude to use only information from the provided documents and not its general knowledge.

Anthropic documentation

Task 5.6Report a problem with this question

Practise domain 5 →Practise all domains →