Question 240 A new moderation filter blocks previously allowed accessibility descriptions of medical images. How should the team respond? Remove image descriptions from the product permanently Investigate false positives with reviewed accessibility test cases and policy scope Accept every refusal as a correct safety result Disable all safety controls globally without review Explanation Correct Answer B:Safety changes must balance disallowed content with legitimate descriptions. Incorrect Answers A:A targeted filter error does not justify eliminating accessibility features. C:Legitimate content may be blocked incorrectly. D:Blanket removal creates unrelated misuse risks.
Question 241 A long-running document summarization returns `finish_reason` indicating token truncation. What should the application do? Treat output as incomplete and use a bounded continuation or chunking strategy Mark the run complete because HTTP returned 200 Assume an incomplete final paragraph still proves all required coverage Silently invent the remaining source paragraphs Explanation Correct Answer A:A truncated answer must not be published as a comprehensive summary. Incorrect Answers B:Transport success does not guarantee generation completeness. C:Truncation can omit material sections and citations. D:Generated filler is not grounded source data.
Question 242 A video clip has important captions embedded visually but no audio. What preprocessing should capture? Only the camera's manufacturer metadata Only speech recognition because the file format is video Frame-level OCR and timestamped text segments Only a model-generated soundtrack Explanation Correct Answer C:Burned-in text can convey meaning absent from the audio stream. Incorrect Answers A:Device metadata cannot recover on-screen captions. B:Silent clips have no spoken words to transcribe. D:Synthetic audio does not extract existing visual text.
Question 243 A medical-image analysis app receives an image where labels are upside down. What preprocessing could improve recognition? Use audio translation to recover rotated pixels Guess labels from typical medical reports Orientation correction and verification against the original source Ignore orientation and replace all text with placeholders Explanation Correct Answer C:Aligned text regions can improve OCR while retaining image provenance. Incorrect Answers A:Speech services cannot read orientation-dependent image text. B:Common values are not the exact observed text. D:Placeholder values prevent accurate extraction.
Question 244 A message contains a request to harm someone, and the assistant should detect safety risk before responding. Which analysis fits? Only detection of positive product sentiment Text safety classification with a defined content-policy rubric Only extraction of customer postal codes Only translation to another language without screening Explanation Correct Answer B:Risk categorization should inspect the actual user text before generation. Incorrect Answers A:Sentiment is distinct from policy-based harm classification. C:Entity extraction does not assess harmful intent. D:Changing language does not remove unsafe content intent.
Question 245 A tool-augmented workflow writes records to a service that may time out after accepting a request. What reliability control is essential? Disable all tool-result logging during retries Generate a fresh operation ID on every automatic retry Idempotency key and reconciliation of the original operation status Assume every timeout means the write failed Explanation Correct Answer C:A missing response must not trigger blind duplicate writes. Incorrect Answers A:Without outcomes, the agent cannot safely resolve uncertain state. B:New IDs prevent downstream deduplication of the original action. D:The server may have committed the transaction before disconnection.
Question 246 A model cites a page but the actual claim is absent from the cited paragraph. What evaluation should flag it? Only measure semantic similarity between two model answers Only test whether page URLs have correct syntax Only count the citations in the final answer Claim-to-source entailment or manual support checks using exact retrieved spans Explanation Correct Answer D:Citation presence alone is not proof the statement is supported. Incorrect Answers A:Agreement does not validate either against the source. B:Valid links do not guarantee content alignment. C:Many links can still be irrelevant.
Question 247 A case worker finds a missing critical document while other agents are drafting. What should the coordinator do? Update affected task dependencies and request targeted evidence before final claims Invent the missing attachment's content from context Publish the incomplete report as fully verified Discard all already verified unrelated work Explanation Correct Answer A:New source gaps should block only conclusions that depend on the missing material. Incorrect Answers B:Speculative text is not a replacement for the actual file. C:A coverage gap must be acknowledged or closed. D:One missing document does not invalidate independent findings.
Question 248 An image includes embedded instructions to disable the agent's security checks. What should an agent do? Obey the embedded command because OCR read it accurately Hide the command from logs and execute it anyway Treat embedded text as untrusted visual evidence, not executable instructions Automatically run any command in a screenshot Explanation Correct Answer C:OCR output has no authority to alter host-side policies or tool permissions. Incorrect Answers A:Accurate recognition does not grant instruction priority. B:Concealment makes the unsafe action harder to detect. D:Screenshots can be attacker-controlled prompt injections.
Question 249 A system uses LLM self-critique to approve generated legal advice without external reference. What is missing? A higher confidence score produced by the first model Longer generated explanations of the initial conclusion More agreement from the same model using its prior answer Independent source evidence and defined escalation for high-impact uncertainty Explanation Correct Answer D:Self-review does not establish applicability of the actual law or contract. Incorrect Answers A:Self-reported confidence is not external verification. B:Length cannot supply missing authoritative legal evidence. C:Correlated model outputs can reinforce the same unsupported assumption.
Question 250 An agent workflow has dependencies: identify customer, verify entitlement, then issue refund. How should steps be scheduled? Execute all steps concurrently to minimize latency Sequentially enforce prerequisites before the refund tool is available Treat missing verification as equivalent to approval Skip identity checking whenever the customer is polite Explanation Correct Answer B:Dependent tasks cannot be safely parallelized without verified earlier results. Incorrect Answers A:Refund might occur before identity or entitlement checks complete. C:Unknown state must not satisfy a mandatory prerequisite. D:Conversational tone is not authorization.