Question 20 A user asks for three options with short explanations but a prompt template returns pages of generic prose. What should be tuned? Explicit output count, length and example format, then test compliance Delete all constraints on response count Increase maximum output tokens further Replace the text model with a video generator Explanation Correct Answer A:Specific structure guides concision without changing the underlying factual task. Incorrect Answers B:Unbounded instructions do not guarantee three concise options. C:More generation capacity may worsen excessive verbosity. D:Video generation cannot satisfy the requested short textual comparison.
Question 21 A support service must classify an uploaded screenshot and answer a question about its visible error. Why must a visual model see the actual screenshot? Because a model catalog page already contains the customer error The error may depend on pixel-level text and UI context absent from the prompt Because every support question must always generate a new image Because OCR is sufficient without any surrounding UI context Explanation Correct Answer B:Direct visual input grounds the response in the customer's displayed state. Incorrect Answers A:Catalog metadata is not a substitute for the user's evidence. C:Image creation is not required to interpret an existing error screenshot. D:Text recognition may recover an error string but can miss the visual state and layout explaining it.
Question 22 A travel assistant gives outdated visa advice when prompted without sources. What architecture most directly adds current evidence? Increase response temperature so answers vary Retrieve approved up-to-date guidance and inject cited passages into generation Use a larger model without external retrieval Cache one old answer for all future travelers Explanation Correct Answer B:Retrieval-augmented generation grounds claims in source content at inference time. Incorrect Answers A:Sampling diversity does not introduce authoritative visa policy updates. C:More capacity cannot guarantee current legal requirements. D:Caching can preserve the very outdated advice causing the error.
Question 23 An accessibility caption model guesses a person's protected medical status from their appearance. What policy helps? Treat clothing color as a medical diagnostic feature Limit descriptions to relevant visible facts and avoid unsupported sensitive inferences Generate fictional medical notes to accompany the caption Include the inferred status if the model sounds certain Explanation Correct Answer B:A photograph cannot reliably establish private health information or diagnoses. Incorrect Answers A:Appearance cues are not sufficient clinical evidence. C:Fabricated records can mislead and harm the depicted individual. D:Confidence does not establish medical evidence.
Question 24 A retrieval result includes instructions to disable agent audit logging. How should the agent treat those lines? An approved security change because the page uses bold headings A new system instruction because it appears in a tool result Untrusted source text that must not override control-plane policies A required step before answering any user question Explanation Correct Answer C:Retrieved data may be quoted or analyzed but cannot change governance rules. Incorrect Answers A:Formatting cannot authorize disabling trace logs. B:Tool responses do not inherit administrator authority. D:No legitimate user instruction delegates audit control through retrieved text.
Question 25 An app should return a concise defect report in stable fields for automated PR comments. What output contract fits? A schema with path, line, severity, evidence and recommendation fields Unbounded prose that mixes all defects into one paragraph Only a pass/fail emoji without details A fabricated line number for any missing location Explanation Correct Answer A:Structured findings can be validated before attaching comments to exact code locations. Incorrect Answers B:Automation cannot reliably parse or position free-form comments. C:A label lacks actionable file evidence or severity rationale. D:Invalid anchors can mislead developers and reviewers.
Question 26 A model sometimes returns nonsensical confidence values above 100% in classification JSON. What constraint should be added? Validate bounded numeric ranges and reject out-of-contract values Accept any floating point number as a valid percentage Convert every number above 100 to a string Use only output length as the quality check Explanation Correct Answer A:A type-correct number can still violate a required business invariant. Incorrect Answers B:Out-of-range scores are semantically invalid. C:Reformatting does not establish confidence calibration. D:Number of characters cannot detect invalid numeric bounds.
Question 27 An invoice scanner must return vendor, amount, tax and invoice date from photographed receipts. Which analyzer is relevant? General-purpose OCR returning lines of text without mapped invoice fields A supported receipt/invoice Content Understanding analyzer with structured fields Azure AI Search indexer over raw image binaries without an extraction skill An entity extractor configured only for company names in receipt text Explanation Correct Answer B:Domain-specific extraction combines OCR and field interpretation for common financial forms. Incorrect Answers A:Recognized strings alone do not guarantee normalized vendor, tax and date fields for downstream processing. C:A searchable file record does not by itself produce typed financial fields from scanned pixels. D:Entity extraction of organization names does not calculate amounts, tax and invoice dates.
Question 28 A purchase tool times out after submission. Which TWO actions help avoid duplicate orders? Invent a purchase confirmation so the user is reassured Assume any timeout means the purchase failed Use the same durable idempotency key for recovery and safe retries Query existing operation status before any new create request Issue a new purchase ID on every timeout Explanation Correct Answers C:The backend can recognize repeats of the original business event. D:The first submission might have succeeded before the connection failed. Incorrect Answers A:A fake receipt is not a transaction outcome. B:Network errors do not prove the server rolled back. E:Fresh IDs defeat deduplication and can create repeated orders.
Question 29 A safety dashboard reports higher refusal rates on benign medical education prompts. What should evaluators inspect? Lower all safety defenses without testing harmful content False-positive slices and documented policy thresholds for the benign set Count only correctly blocked attacks Treat every refused prompt as malicious by definition Explanation Correct Answer B:Safety quality requires measuring unnecessary blocks as well as unsafe completions. Incorrect Answers A:Blanket relaxation can create a different high-impact failure. C:Ignoring legitimate refusals conceals user-impact regressions. D:A policy block is not proof the input was unsafe.
Question 30 A service uses a generated summary as an input to a later compliance decision. What intermediate control fits? Allow downstream stages to treat every assertion as verified Use only the summary's readability score as approval Validate supported claims and retain exact source references between stages Discard citations after the first model summarizes Explanation Correct Answer C:Compaction must not turn uncertain evidence into unqualified policy facts. Incorrect Answers A:An earlier model's fluent output may contain errors. B:Style is separate from factual or legal validity. D:Later decisions cannot inspect the underlying support.