Question 80 A market-research assistant cites a webpage for a numerical claim, but the cited page never mentions that number. What is the correct treatment? Keep the number because a URL is present. Remove or qualify the claim until an actually supporting source is found. Invent a quotation that would justify the value. Replace the citation with an unrelated reputable homepage. Explanation Correct Answer B:A citation is valid only when it substantiates the specific statement. Incorrect Answers A:A link is not automatically evidence. C:Fabricated quotes worsen the error. D:Reputation alone does not support the figure.
Question 81 A new agent repeatedly asks how to run the repository’s lint check. The team has a stable package command. What will most reduce repeated prompts? Record the exact lint command and its scope in persistent project instructions. Replace lint with a generic unrelated unit test. Ask a new specialist agent to guess the command on every run. Increase the chat length of every run. Explanation Correct Answer A:A short stable instruction is more reliable than rediscovering it each session. Incorrect Answers B:Different checks do not serve the same purpose. C:Repeated inference adds work and can be wrong. D:Longer history does not make a missing command explicit.
Question 82 A data team starts an asynchronous classification batch. The client disconnects before results are ready. Which state must persist? The batch job ID, input mapping and processing status for later polling or retrieval. Only a count of tokens sent. A guessed completion timestamp in place of a job ID. Only the terminal’s current scrollback. Explanation Correct Answer A:Durable identifiers make asynchronous recovery possible. Incorrect Answers B:Usage does not identify the job or its records. C:An estimate cannot retrieve the results. D:Transient shell output is not a durable processing checkpoint.
Question 83 A procurement workflow requires a signed manager approval for spending over $10,000. Which control best guarantees this policy even after prompt drift? Allow the agent to approve its own purchase requests. Put the policy only in a team README. Ask the assistant to apologize if it purchases without approval. Verify approval status and spend amount before the purchase API accepts the command. Explanation Correct Answer D:Deterministic validation at the irreversible action boundary enforces the limit. Incorrect Answers A:Self-approval removes the separation of authority. B:Documentation does not block a disallowed transaction. C:Apology after commitment cannot satisfy the approval requirement.
Question 84 A product agent chooses update_inventory instead of get_inventory because the names sound similar. Which description feature most helps routing? Include positive and negative usage criteria, side effects, and sample parameter meanings. Describe both as “general inventory utility.” Use identical descriptions to avoid bias. Require the model to infer side effects from tool registration order. Explanation Correct Answer A:Routing improves when the agent knows when to select a read rather than a mutation. Incorrect Answers B:Generic vocabulary provides no discriminating signal. C:Identical descriptions conceal the distinction. D:Ordering is not a reliable semantic contract.
Question 85 A developer has already approved a precise version update and its test command. Why might another plan-only cycle be unnecessary? The implementation is bounded and authorized; execution plus verification is the next useful step. Any code edit always requires an entirely new research project. The assistant should disregard the version requirement. Tests may be skipped once a plan exists. Explanation Correct Answer A:Repeating analysis without a new uncertainty adds overhead. Incorrect Answers B:The correct amount of planning depends on scope and risk. C:A clear user constraint remains binding. D:A plan does not prove edited code works.
Question 86 A large corpus consists of unrelated confidential customers. What batching constraint protects correctness? Concatenate all customers into one unsegmented prompt. Process each customer item with isolated instructions and data, and preserve stable customer IDs. Remove customer identifiers after submission. Reuse the previous customer’s output as the next customer’s ground truth. Explanation Correct Answer B:Isolation prevents one customer’s content leaking into another’s context. Incorrect Answers A:Shared context risks attribution and privacy errors. C:Results cannot be safely reconciled. D:That encourages cross-customer contamination.
Question 87 An API occasionally returns a 200 status with a business-level failure field. The model treats it as success. What hook should the integration add? Trust the HTTP status without inspecting payload semantics. Always translate 200 into “refund paid.” Interpret the payload status and expose the business failure as a typed error before reasoning. Hide failure fields to simplify tool output. Explanation Correct Answer C:HTTP success may still contain an application failure that downstream actions must respect. Incorrect Answers A:Transport success does not prove the requested business operation completed. B:That would fabricate successes for unrelated operations. D:Suppressing errors makes the next model decision less reliable.
Question 88 An agent plans a small edit to a known file but the target string occurs twice. What should it do? Apply a global replace without inspecting both matches. Use Glob to select a random occurrence. Delete the file and start from a template. Read surrounding context and narrow the Edit replacement to the intended occurrence. Explanation Correct Answer D:Scoped context prevents unintended duplicate replacements. Incorrect Answers A:A broad edit may change the wrong location. B:Filename matching cannot distinguish lines inside the file. C:Destructive reconstruction is unnecessary.
Question 89 An architect is comparing two candidate designs and wants no repository changes. What outcome should Claude Code produce? A tradeoff-oriented recommendation grounded in read-only inspection, not changed source files. A new branch with an unapproved implementation. A production deploy of the preferred candidate. A fabricated test run with no executed tests. Explanation Correct Answer A:The deliverable is a decision, not an implementation. Incorrect Answers B:Creating code exceeds the stated comparison task. C:An architectural comparison does not authorize release. D:Evidence should be reported only if actually obtained.
Question 90 A code reviewer flags public input parameters as SQL injection even though each query uses bound parameters. Which refinement fits? Ignore all database code because false positives occurred. Flag any function that accepts a string parameter. Define the injection criterion as data reaching dynamic SQL structure without safe parameterization. Label every parameterized query as critical until disproven. Explanation Correct Answer C:A source-to-sink rule distinguishes real unsafe query construction from harmless input handling. Incorrect Answers A:That would miss genuine query-construction flaws. B:A string input alone does not establish injection. D:The required evidence is missing when values are bound safely.
