Question 340 An app needs to answer a question about changes between two images. What comparison method fits? Generate a third illustration and compare it to neither source Joint analysis of both images with image identity and difference grounding Analyze only the first image Use document word counts as a change metric Explanation Correct Answer B:The output must attribute which visible objects changed between sources. Incorrect Answers A:Synthetic pixels add no evidence about the requested difference. C:One image cannot establish changes relative to the second. D:Word counts cannot capture visual scene differences.
Question 341 An audio summarizer reports a speaker's statement that was actually background television. What should improve attribution? Remove all timestamps and speaker labels Speaker/source separation and diarization quality checks Treat every detected word as spoken by the primary caller Generate a cleaner voice and call it the source Explanation Correct Answer B:The system must distinguish participant speech from environmental audio. Incorrect Answers A:Less provenance makes attribution harder to correct. C:Background sources can contaminate the attributed transcript. D:Synthetic audio is not proof of who spoke originally.
Question 342 A model deployment fails with a regional capacity error despite available quota elsewhere. What should be evaluated? Change the application's CSS to lower token usage Supported region, deployment type, available capacity and allocated quota together Change the search index document count Publish the endpoint anonymously to bypass quota Explanation Correct Answer B:Quota numbers alone do not guarantee capacity in the selected deployment location. Incorrect Answers A:Visual styling cannot repair a provider capacity allocation failure. C:Index size does not grant model hosting capacity. D:Anonymous access does not create regional model capacity.
Question 343 A model-assisted review workflow is costly but important for high-risk decisions. Which TWO optimization practices make sense? Run the same anchored critic ten times as proof of independence Track marginal error detection against latency and token cost Use an explicit severity rubric and risk-based reviewer invocation Use response length as the main acceptance criterion Ignore all checks for risky financial operations to reduce cost Explanation Correct Answers B:Measured review value can guide a safe escalation threshold. C:Proportionate checks target consequential uncertainty rather than every trivial turn. Incorrect Answers A:Repeated model agreement may reproduce one underlying assumption. D:Verbose text does not prove correctness or safety. E:High-impact mutations still require appropriate controls.
Question 344 A synthetic public campaign must meet branding and disclosure policy. Which TWO safeguards fit? Assume synthetic images are exempt from trademark review Apply and verify required synthetic-media attribution before release Screen generated frames for prohibited marks and policy violations Evaluate only the input prompt, never the produced image Remove all provenance metadata to appear authentic Explanation Correct Answers B:Disclosure rules are only satisfied if the final artifact carries the mark. C:Output media may contain content not explicit in the prompt. Incorrect Answers A:Generated origin does not establish brand rights. D:Output pixels can differ from anticipated safe content. E:Hiding origin can mislead audiences.
Question 345 An architect must support image captions for accessibility and generated banner artwork. Should these use one undifferentiated endpoint? Use understanding and generation capabilities appropriate to their separate outputs Use the same OCR text as the only caption for all photos Use a text sentiment endpoint to create the banner Treat generated image pixels as ground truth captions Explanation Correct Answer A:Captioning requires faithful observation while artwork creation synthesizes new visual pixels. Incorrect Answers B:OCR misses nontextual objects and scene relationships. C:Sentiment classification cannot synthesize the required image. D:Generation can invent details not present in the original image.
Question 346 A brand requires consistent logo position across ten generated product images. What workflow control helps? Remove the brand asset from all inputs Use approved reference composition and placement validation on outputs Assume identical prompts force pixel-identical brand placement Judge only the length of each image description Explanation Correct Answer B:Explicit reference and visual checks reduce unintended logo movement. Incorrect Answers A:Without reference, logo fidelity may degrade. C:Generative sampling can vary even with similar instructions. D:Caption length does not establish visual layout consistency.
Question 347 A deployment health test shows 200 OK but the tool-enabled agent cannot complete orders. What must be added to the gate? Only inspect the agent's display name End-to-end tests for actual agent tool invocation and business result validation Only ping the model service again Only increase the maximum output length Explanation Correct Answer B:A ready endpoint does not establish downstream tool correctness. Incorrect Answers A:Names do not show authorization and tool-execution results. C:Repeated health pings do not exercise the order tool contract. D:More tokens cannot repair a failing API integration.
Question 348 A question-answering app requires exact product codes and semantic paraphrases. Which Azure AI Search configuration is relevant? Pure vector search with no lexical or identifier constraints Hybrid keyword/vector retrieval with ranking tuned to both requirements Keyword-only search ignoring semantic variation Video generation for each candidate answer Explanation Correct Answer B:Hybrid search handles exact identifiers and semantically related wording. Incorrect Answers A:Approximate neighbors may miss exact product codes. C:Lexical matches can miss synonyms and paraphrases. D:Producing visual assets cannot retrieve exact product records.
Question 349 A product support index needs exact part numbers and semantic symptom matches. Which TWO retrieval features help? Exact identifier filters or lexical matching for part codes Discard product-version metadata during ingestion Use only image generation to illustrate possible parts Vector/semantic retrieval for symptom paraphrases with evaluated ranking Accept any high-scoring vector as the requested part Explanation Correct Answers A:Structured equality avoids mixing near-identical product variants. D:Meaning-based search finds relevant descriptions with different wording. Incorrect Answers B:Missing versions allow irrelevant or superseded manuals. C:Generated images cannot identify exact catalog records. E:Similarity alone cannot guarantee part identity.
Question 350 A user asks for an extended accessibility description of a dashboard with multiple plots. What level of detail is appropriate? Replace the charts with generic motivational language Only return a one-line caption without explaining the requested data relationships List random exact numbers not legible in the image Include relevant chart types, trends, labels and relationships in a structured description Explanation Correct Answer D:Complex informational visuals need more than a decorative one-line caption. Incorrect Answers A:Unrelated prose does not describe the supplied dashboard. B:A short decorative caption cannot convey the axes, trends and plot comparisons. C:Invented figures undermine accessible understanding.