Question 370 A supervisor agent delegates insurance analysis to a junior subagent, which asks to issue a payout. What should happen? Allow the junior agent to use any visible payout tool Treat its high confidence as sufficient authorization Keep payout execution with authorized supervisor or policy tool gating Hide the payout in the final response Explanation Correct Answer C:Delegated analysis does not imply delegated financial disbursement authority. Incorrect Answers A:Tool presence alone does not establish permission for side effects. B:Confidence cannot grant a business signing limit. D:Concealment does not protect the transaction boundary.
Question 371 A model's safety performance is measured on only simple offensive-language prompts. What additional cases should be added? Only latency measurements under load Indirect injection, multi-turn coercion and context-sensitive legitimate requests Only requests with no risky content More identical one-turn offensive examples only Explanation Correct Answer B:Adversarial and false-positive slices provide a more representative safety picture. Incorrect Answers A:Speed does not measure policy boundary behavior. C:Benign cases cannot test harmful-output controls. D:Repetition does not cover distinct attack mechanisms.
Question 372 An agent handles a multi-step registration workflow but the user disconnects after creating an account. What state enables continuation? Store only the initial welcome message Assume every step completed because account creation succeeded Persist account ID, completed step and next required action Always restart registration by creating a new account Explanation Correct Answer C:Durable status prevents recreating accounts after connection loss. Incorrect Answers A:Greeting content does not record the completed mutation. B:Later verification or consent stages may remain unfinished. D:Duplicate accounts can result from repeated initial steps.
Question 373 A model-based app must return exactly three risk categories from an approved list. What is the most reliable implementation? Assume any free-form phrase is one of the categories Use image generation to express the category as a picture Ask for creative new category names every run Constrain the category enum and validate returned values server-side Explanation Correct Answer D:Enumerated contracts prevent unapproved labels controlling business workflows. Incorrect Answers A:String fluency does not establish enum membership. B:Pictures cannot replace deterministic machine-readable labels. C:Unlimited labels undermine consistent routing and reporting.
Question 374 An image analyzer must identify objects and support review of disputed detections. Which TWO fields improve traceability? A generic statement that everything was recognized correctly Per-object class and confidence or uncertainty metadata Random coordinates generated after recognition Only the image filename without content identity Stable image identity with object region coordinates or polygons Explanation Correct Answers B:Reviewers can prioritize ambiguous regions for verification. E:Location-level grounding ties each label to visible evidence. Incorrect Answers A:Self-assurance is not inspectable evidence. C:Fabricated positions cannot locate the claimed object. D:Names can collide and do not identify a region.
Question 375 A team changes embedding encoders and notices the same query returns unrelated passages. What should it examine? Index/query embedding-model parity and vector migration strategy Only the date on the document's cover page Only regenerate the query embedding with the new model while keeping old document vectors Only the language style of final responses Explanation Correct Answer A:Different embedding spaces can invalidate similarity ranking across old and new vectors. Incorrect Answers B:Freshness alone cannot fix incompatible vector encoders. C:Query/document encoder mismatch remains when the index has not been rebuilt or migrated. D:Generated tone does not change stored embeddings.
Question 376 A marketing video must change a background scene while leaving the speaker's audio intact. What should the editing plan separate? Apply a text-only style rewrite to the video file Ignore frame and timestamp relationships Regenerate all speech from guessed transcript Video-region edits from unchanged audio tracks, with synchronized output checks Explanation Correct Answer D:Separate modality handling protects speech timing and sound during visual changes. Incorrect Answers A:Text formatting cannot alter video pixels while preserving audio. B:Temporal continuity is essential to a coherent edited clip. C:New audio may misstate or desynchronize the speaker's original words.
Question 377 An agent is assigned to assess code safety and is given production deployment credentials by default. What should the architect do? Skip review logs so access goes unnoticed Automatically deploy any fix the reviewer suggests Remove deployment access and retain only required read/test tools Keep administrator credentials hidden in the system prompt Explanation Correct Answer C:Code analysis should not carry unrelated external publication authority. Incorrect Answers A:Lack of traceability does not reduce operational risk. B:A suggested patch is not authorization for production release. D:Hidden access still leaves an unnecessary capability available.
Question 378 A text analysis assistant sees a customer's ambiguous relative deadline. Which TWO practices maintain accuracy? Convert all 'next week' references to the same calendar date Remove the timeframe from the summary without explanation Record the original phrase and the relevant message timestamp/time zone Use explicit unresolved/clarification state if an exact date cannot be established Assume the current device locale equals the source author's locale Explanation Correct Answers C:Relative dates need a reference instant before normalization. D:Missing context should not become an invented deadline. Incorrect Answers A:Relative expressions depend on when they were written. B:An unresolved commitment can still matter operationally. E:User settings may differ from message origin.
Question 379 A map screenshot shows route arrows but the model guesses travel time without a scale. What should it do? Describe visible route structure while withholding unsupported travel-time estimates Generate a plausible time and label it verified Treat every screenshot as a live traffic feed Infer exact minutes from arrow thickness Explanation Correct Answer A:Pixels without grounded distances and traffic cannot prove an ETA. Incorrect Answers B:An estimate cannot be asserted as measured without evidence. C:A static map image may be stale or lack traffic data. D:Drawing style is not a travel-duration metric.
Question 380 A retrieval app must answer from images as well as PDF text. What representation choice matters? Index only image filenames Ignore images with no embedded text Image descriptions/embeddings alongside text chunks and modality metadata Treat every image as a paragraph generated from the title Explanation Correct Answer C:Visual facts need searchable representations that retain their image origin. Incorrect Answers A:Names may not reveal visual content. B:Some images convey purely visual evidence. D:Guessed descriptions may not match the actual pixels.