Question 300 A CRM MCP server is healthy, but the caller supplied a negative invoice total outside allowed validation. Which category is appropriate? A successful empty invoice. A nonretryable invalid-argument failure with the violated constraint. A retryable network timeout. A permission-granted status with no data. Explanation Correct Answer B:The input must be changed; retrying unchanged values is not a transient recovery. Incorrect Answers A:A malformed request is not evidence of an invoice with zero lines. C:The server is reachable and has rejected invalid business input. D:Validation is unrelated to authorization success.
Question 301 A CI reviewer flags 500 style findings that a linter already handles, drowning out actual security risks. How should the prompt change? Define high-signal review criteria and explicitly exclude routine lint/style issues. Disable the linter and make the model handle whitespace only. Ask for as many comments as possible to demonstrate coverage. Randomize findings so every run looks different. Explanation Correct Answer A:Focused review reduces false positives and duplication with deterministic tools. Incorrect Answers B:Deterministic style checks are more appropriate for lint. C:Quantity can make the review less useful. D:Nondeterminism does not improve signal quality.
Question 302 A JSON schema describes allowed_status as enum open, paused, closed. The model wants to emit “almost closed.” What should the output contract do? Accept arbitrary status labels to improve expressiveness. Silently reinterpret “almost closed” as closed without authorization. Remove all status values from the record. Constrain output to the defined enum and represent ambiguity through a separate allowed field if needed. Explanation Correct Answer D:A typed enum provides a stable machine-readable state. Incorrect Answers A:Unknown values break downstream consumers. B:That can trigger the wrong business behavior. C:The consumer requires the field.
Question 303 A tool receives an authorized search request, but the data returned include secrets not needed for the requested answer. What PostToolUse control is appropriate? Tell the model to forget the secrets while still displaying them. Encrypt the final answer but leave tool logs public. Apply field-level minimization and redaction while retaining relevant factual fields. Copy all secrets into a general long-term summary. Explanation Correct Answer C:Removing unnecessary sensitive fields reduces prompt exposure without losing task evidence. Incorrect Answers A:A prompt reminder does not prevent exposure. B:Public logs would still expose the unnecessary secrets. D:Retention expands access and creates avoidable risk.
Question 304 A model produces an answer with a specific source quotation, but a source comparison shows several words differ. How should it respond? Remove the citation and leave quotation marks. Correct the quotation exactly or convert it into a clearly identified paraphrase with citation. Keep the inaccurate quote because its meaning is similar. Add invented quotation marks around the edited wording. Explanation Correct Answer B:Verbatim claims require faithful text; paraphrases must not masquerade as quotes. Incorrect Answers A:That makes attribution less verifiable. C:Quoted text must be exact. D:That worsens the false attribution.
Question 305 A team uses Claude Code to implement API handlers, but reviewers report repeated authorization omissions. Which improvement closes the feedback loop? Hide reviewer comments so the model is not distracted. Add more stylistic examples without testing authorization. Turn reviewed authorization requirements into explicit tests and reusable instructions. Assume every generated handler is secure by default. Explanation Correct Answer C:A recurring failure should become an enforceable check for future edits. Incorrect Answers A:Ignoring feedback preserves the omission. B:Style does not prove access-control behavior. D:Unverified security claims do not replace tests.
Question 306 A batch extractor passes JSON shape checks, but it assigns the same document ID to two unrelated contracts. Which feedback improves reliability? Remove the document identifier from every record. Compare extracted IDs with document-level provenance and flag mismatches for targeted correction. Use the ID that occurs more often regardless of source. Assume any matching JSON type rules out identity confusion. Explanation Correct Answer B:Source binding catches cross-document contamination. Incorrect Answers A:That prevents traceability. C:Frequency does not establish provenance. D:A valid string may refer to the wrong document.
Question 307 A team wants a deterministic one-tool classification step, followed by a flexible explanatory answer. How should tool selection be staged? Allow the explanatory phase to call an unrelated payment tool. Never call the classifier and infer categories from answer prose. Require the classifier at the first decision boundary and permit ordinary response generation afterward. Force repeated classification on every subsequent turn. Explanation Correct Answer C:Constrain only the step that needs a particular function. Incorrect Answers A:Unrelated privileges do not improve classification explanation. B:That defeats the deterministic classification requirement. D:Continued forced calls waste tokens and may loop.
Question 308 A software tool fails with a permission error halfway through an approved read-only inventory. What should the assistant communicate? Retry indefinitely without telling the user. Read a different customer’s account as a workaround. Claim the inventory is comprehensive. Report completed coverage, the denied scope and the specific access needed to finish. Explanation Correct Answer D:An explicit partial result allows safe correction without fabricated completeness. Incorrect Answers A:Authorization errors need intervention, not repetition. B:That crosses authorization boundaries. C:The denied resources were not inspected.
Question 309 A review bot must send structured comments to a PR API, but Claude Code sometimes emits a markdown essay. Which output contract is appropriate? Allow any text and let the PR API guess which parts are JSON. Disable all schema validation to reduce latency. Request machine-readable JSON and validate its schema before posting comments. Write the raw assistant transcript as a production environment variable. Explanation Correct Answer C:A structured CI contract prevents free-form prose from breaking downstream integrations. Incorrect Answers A:Implicit parsing is fragile. B:Invalid payloads can create malformed review comments. D:Unstructured transcripts do not satisfy a review-comment API.
Question 310 A compliance report requires every regulation to be mapped to controls. What decomposition makes completeness testable? Create a regulation-to-control matrix, verify each row, then synthesize gaps. Have each worker choose regulations without a shared index. Split solely by equal word counts. Write an executive summary first and declare it complete. Explanation Correct Answer A:Explicit coverage units make omissions visible before prose drafting. Incorrect Answers B:Uncoordinated choices can duplicate some and miss others. C:Text length does not correspond to legal control coverage. D:A polished summary cannot prove all regulations were mapped.
