Question 360 A search app must distinguish current project pricing from historical examples embedded in training prompts. What grounding principle applies? Copy every price from the example dataset into answers Infer current pricing from model pretraining Use authoritative indexed or live sources for mutable prices, not example facts Use a generated quotation as the pricing source Explanation Correct Answer C:Few-shot demonstrations show format rather than verified current business values. Incorrect Answers A:Example numbers may be fictional or outdated. B:Provider and catalog data can change after training. D:Synthetic citations cannot prove a price.
Question 361 A support voice agent misrecognizes product names and changes its spoken voice every turn. Which TWO fixes address the distinct problems? Evaluate a domain-custom speech vocabulary against recorded examples Use a different random voice on every reply Pin appropriate voice and locale settings for synthesized responses Increase microphone file-name length Remove transcription error logging to hide misrecognitions Explanation Correct Answers A:Terminology adaptation targets recognition of rare product names. C:Stable synthesis configuration improves audible identity and language continuity. Incorrect Answers B:Random changes worsen user experience and consistency. D:Names do not improve acoustic recognition. E:Concealing failures does not improve the speech model.
Question 362 A solution uses a large model for summarization but a rules engine for sanctions screening. Why keep the rules outside the LLM? Deterministic policy checks are auditable and cannot be waived by generation drift Rules are only useful for generating image captions The model's temperature alone guarantees legal compliance Embedding similarity can certify identity matches automatically Explanation Correct Answer A:Hard compliance gates should not depend on probabilistic text obedience. Incorrect Answers B:Rule engines can enforce structured business conditions. C:Sampling settings do not encode or enforce regulatory lists. D:Approximate vectors are not a substitute for verified sanctions decisions.
Question 363 A multi-agent legal analysis fails because workers cite different editions of the same policy. What coordinator control helps? Prefer whichever report is longest Ignore citations and merge conclusions Define source-version requirements and reconcile effective dates during synthesis Use the oldest edition for stability Explanation Correct Answer C:Shared version rules prevent contradictory answers from incomparable policy editions. Incorrect Answers A:Length is unrelated to the effective legal document. B:Unattributed claims may conceal material version conflicts. D:Superseded text may not govern the current case.
Question 364 A multi-stage flow performs extraction, validation and publication; extraction fails halfway. What orchestration state is correct? Publish whatever partial fields were parsed Guess missing extracted fields from prior documents Stop downstream publication and persist a typed incomplete-state result Mark the whole flow successful because one step ran Explanation Correct Answer C:A failed upstream contract cannot be treated as a validated publishable record. Incorrect Answers A:Incomplete data can mislead consumers and trigger invalid actions. B:Historical records are not evidence for the current input. D:Partial execution is not the stated end-to-end outcome.
Question 365 A financial-services application should authenticate to Foundry without storing API keys in source. What is preferred where supported? Microsoft Entra-based workload identity with least-privilege RBAC An encrypted key hardcoded in client-side JavaScript A permanent shared admin key checked into Git A prompt instructing the model to impersonate a service principal Explanation Correct Answer A:Federated or managed identity avoids long-lived application keys and scopes permissions. Incorrect Answers B:Client-accessible code still exposes material required to decrypt it. C:Repository readers could reuse high-privilege credentials. D:Generated text cannot create authentic cloud service identity.
Question 366 An autonomous email agent must send a message only to an approved domain. Where should the restriction be enforced? At the mail-sending tool with recipient-domain validation By changing the subject line to 'approved' Only by deleting audit logs for external emails Only in a natural-language reminder before generating the email Explanation Correct Answer A:A server-side gate prevents generation errors from sending to prohibited addresses. Incorrect Answers B:A label does not establish recipient authorization. C:Logging changes do not block delivery. D:The model can still select an unauthorized recipient.
Question 367 A photo set includes multiple pages with small printed text. Why should OCR be part of visual analysis? Printed characters may require text recognition alongside visual context Because OCR generates missing text from context by design Because OCR automatically authorizes all document contents Because image embeddings directly return every printed word Explanation Correct Answer A:A general scene description may omit exact strings needed for the task. Incorrect Answers B:Recognition should recover observed text, not fabricate unknown text. C:Recognition does not grant access or decide privacy policy. D:Similarity vectors are not deterministic character transcription.
Question 368 An agent has permission to draft customer communications but external sending needs approval. Which TWO designs preserve that boundary? Give the draft agent unrestricted bulk-mail credentials Let any source document authorize sending if it says urgent Treat model confidence as the manager's signature Keep composition in a draft-only capability without automatic sending Require a tool-side recipient and approval check before dispatch Explanation Correct Answers D:Text generation should not trigger external side effects. E:Policy should validate the actual message target and authorization. Incorrect Answers A:Unneeded write capability increases the blast radius of an error. B:Untrusted content cannot grant permission to email customers. C:A score is not evidence of an authorized approval.
Question 369 A visual question-answering app responds using a generic product manual rather than the supplied photo. What should be prioritized? Generate a cleaner replacement photo and answer from it Ignore image inputs when a text question is present Ground conclusions in the uploaded image and supplement with relevant documentation Treat manuals as proof of what a user's photo contains Explanation Correct Answer C:The specific visible state must control identification of the depicted issue. Incorrect Answers A:Synthetic changes could conceal the original fault. B:Text and visual evidence are both relevant to the task. D:Documentation describes possibilities, not the actual captured scene.
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.