Question 350 A conversation contains a requested purchase and a background research task. The model tries to place the order as soon as it finds a price. What host-side condition is essential? Assume any product mention grants purchase consent. Charge first and ask permission in the final message. Authorize if the tool schema validates the price field. Check explicit purchasing authorization and constraints before executing the order tool. Explanation Correct Answer D:Research access is not equivalent to permission to spend money. Incorrect Answers A:Interest in a product does not authorize a transaction. B:Post-action approval cannot protect the user against an unauthorized purchase. C:Schema validity checks form, not the user’s consent.
Question 351 Two independent runs both miss an error because the input was truncated before a critical appendix. What must the review protocol add? Check source coverage and input completeness before relying on reviewer agreement. Add more reviewers using the same truncated input. Treat repeated agreement as a proof of completeness. Drop all appendix references from the final report. Explanation Correct Answer A:Missing evidence can cause correlated blind spots despite separate model runs. Incorrect Answers B:The missing appendix will remain invisible. C:Agreement is compatible with common omissions. D:Concealing missing source coverage is misleading.
Question 352 An engineer asks for a safe preview of changes across 40 services but has not approved implementation. What deliverable best matches? A summary claiming all changes were deployed. Forty unreviewed commits to the production branch. A dependency-informed plan with risks, proposed edits, test strategy, and rollback approach. A blank file tree after cleanup. Explanation Correct Answer C:A plan supports review before many potentially disruptive writes. Incorrect Answers A:That misrepresents the actual authorization and status. B:Execution exceeds the requested preview. D:Deleting files is not a preview.
Question 353 A refactor spans API controllers, service classes and database queries. How should the agent gather context? Inspect only controllers and assume database behavior. Trace the call chain and relevant tests across layers, preserving interfaces between components. Change each layer independently without checking interface compatibility. Read unrelated UI snapshots until the context is full. Explanation Correct Answer B:Focused dependency traversal covers the behavior without reading the whole repository. Incorrect Answers A:That may miss authorization or persistence constraints. C:Cross-layer mismatches can break the refactor. D:Unrelated files do not inform backend control flow.
Question 354 An agent has reached max_tokens while drafting a long response and no final end_turn was received. What should the host conclude? Treat the answer as incomplete and apply a deliberate continuation or recovery strategy. Mark the task complete because any non-tool stop is final. Run every read-only tool again before resuming. Delete the previous messages and ask for a fresh answer. Explanation Correct Answer A:max_tokens indicates the generation stopped for a length boundary, not verified completion. Incorrect Answers B:A limit stop can truncate a response and should not be mistaken for a completed deliverable. C:Repeating tools wastes calls and does not address the truncated generation. D:Discarding evidence and prior constraints introduces avoidable inconsistency.
Question 355 A Claude Code user has a personal MCP search server but an audited project must use only a company-approved connector. What scope boundary matters? Enforce the project’s authorized server set instead of relying on personal global defaults. Treat every user-level server as implicitly approved. Remove the project’s audit trace to avoid configuration conflicts. Allow any connector that can return syntactically valid JSON. Explanation Correct Answer A:An audited workspace should not inherit unapproved data paths. Incorrect Answers B:Personal settings do not grant project authorization. C:Missing audit evidence worsens governance. D:Output shape is not a data-governance permission.
Question 356 A subdirectory-specific CLAUDE.md conflicts with a company rule forbidding credential commits. What should take precedence? The local file always wins even if it creates a secret leak. Whichever rule uses more exclamation marks. The higher-authority security constraint; the local instruction cannot authorize committing secrets. A random choice between the two instructions. Explanation Correct Answer C:Narrow scope does not override organizational security policy. Incorrect Answers A:A project exception is not permission to violate higher-level controls. B:Stylistic emphasis does not establish authority. D:Conflicts need explicit resolution, not chance.
Question 357 An architecture assistant is asked to review a migration for “risk” and produces vague concerns. What instruction makes output actionable? Remove all references to the migration plan. Demand a specific failure mode, triggering condition, business impact, and proposed mitigation. Require all findings to use exactly the same generic sentence. Ask for a longer introduction about general AI risk. Explanation Correct Answer B:Concrete criteria connect each finding to the actual migration decision. Incorrect Answers A:The answer would lack the evidence needed to assess risk. C:Uniform filler hides distinct mechanisms. D:More background does not make the case-specific findings testable.
Question 358 An agent repeatedly calls lookup_order with unchanged parameters after receiving the same result. What should the host implement? Raise the maximum output token setting. Replace the order API with a slower API. Allow infinite calls as long as the tool succeeds. Detect no-progress repetitions and apply a bounded recovery path. Explanation Correct Answer D:A repeated identical call with no new evidence signals a stuck loop rather than useful autonomy. Incorrect Answers A:Generation length does not solve a repeated tool-selection pattern. B:Latency changes do not address the missing progress check. C:Successful but unproductive calls still consume time and budget.
Question 359 A server tool is expensive and rate-limited but its description gives no indication. What useful guidance belongs in the interface? Keep error semantics secret so the model learns by trial. Describe it as unlimited even though quotas are enforced. Document call cost, pagination, filtering, and when a cached lookup is preferable. Ask the model to request every page by default. Explanation Correct Answer C:The agent can choose narrower queries and avoid wasteful repeated full scans. Incorrect Answers A:Trial and error is inefficient and may trigger avoidable load. B:False capacity guidance leads to throttling and failure. D:Unnecessary enumeration increases latency and cost.
Question 360 Two sources report different versions of an API. One is a current official API reference and the other an undated blog post. Which approach is best? Average the two version numbers. Trust the blog because it has more comments. Present both as simultaneous exact requirements without qualification. Use the versioned official reference for implementation while documenting the conflicting secondary account. Explanation Correct Answer D:Source authority and currency matter when technical contracts change. Incorrect Answers A:Version identifiers are categorical, not averaged. B:Popularity does not establish current API behavior. C:That may produce an invalid request.
