Question 210 A model quality dashboard shows rising refusal rates after a filter-policy update. What should the team compare? Compare only the model deployments, assuming safety thresholds stayed constant Review total refusal volume without grouping cases by the changed policy Check only inference response times before and after the filter release Safety filter versions, event categories and matched evaluation cases Explanation Correct Answer D:Correlating policy changes with refusal examples identifies overblocking or actual risk shifts. Incorrect Answers A:A moderation update can shift outcomes without changing the serving model. B:Aggregate counts hide whether benign examples were wrongly blocked by a specific rule. C:Latency trends do not establish which content category caused refusals.
Question 211 An agent asks to use a tool outside its approved workspace after encountering a permission error. What is the correct response? Search for stored administrator secrets to bypass the guard Run the same call with random resource IDs Report the blocked scope and request authorized access rather than bypassing policy Disable the guard because the task appears urgent Explanation Correct Answer C:Failure should not trigger privilege expansion by guesswork. Incorrect Answers A:Unauthorized credential access violates the original task boundary. B:Guessing targets can cross tenant or customer boundaries. D:Urgency does not replace explicit authorization.
Question 212 An extracted document contains monetary totals that do not match line items. What should downstream processing do? Accept totals merely because they are numeric Silently adjust line items to match the stated total Recompute sums from source-backed fields and flag inconsistent results Increase model temperature to seek a more plausible amount Explanation Correct Answer C:OCR confidence and JSON parsing do not prove arithmetic correctness. Incorrect Answers A:Correct types do not establish valid accounting values. B:Invented changes conceal source discrepancies. D:Sampling cannot substitute for deterministic reconciliation.
Question 213 An evaluation pipeline stores failure cases but never revisits them after a model upgrade. What is missing? Deleting old failure examples to make the dashboard green Treating model version increases as guaranteed safety improvements Versioned regression tests for previously identified harmful and incorrect behaviors Running only the newest happy-path prompt Explanation Correct Answer C:Known failures should become repeatable checks against new releases. Incorrect Answers A:Removing evidence conceals regressions instead of fixing them. B:New versions may introduce different failure patterns. D:One easy case cannot protect against historical defects.
Question 214 An AI support app must detect and report when its index is missing a document that the customer explicitly references. What flow fits? Check source inventory and return a typed missing-document condition before answering Treat a search service outage as an empty authorized index Invent the referenced document title and its contents Use the nearest unrelated chunk as the missing document Explanation Correct Answer A:A failed lookup should not be masked by semantically similar unrelated content. Incorrect Answers B:Infrastructure failure does not prove the document is absent. C:Generated evidence cannot replace absent source material. D:Similarity does not establish exact document identity.
Question 215 A team must compare two recognition systems for medical dictation. What metrics are most useful? Only the voice quality of synthesized responses Word/term error rates on representative recordings plus downstream clinical impact Only number of speakers supported in marketing material Only total GPU hours consumed Explanation Correct Answer B:Domain errors matter more than generic clean-speech demonstrations. Incorrect Answers A:Output audio fidelity does not measure speech recognition. C:A specification cannot replace measured transcription quality. D:Cost alone cannot show dictation accuracy.
Question 216 A synthetic product video is to be used publicly. What provenance control may be appropriate? Apply approved synthetic-media disclosure or watermark requirements Strip all origin labels to imply actual events Use the highest frame rate as a disclosure substitute Treat photorealism as proof it was filmed Explanation Correct Answer A:Clear provenance helps prevent presenting generated scenes as documentary footage. Incorrect Answers B:Concealing synthetic origin can mislead viewers. C:Smooth motion cannot communicate whether the clip is synthetic. D:Realistic synthesis does not establish real camera capture.
Question 217 A staging AI endpoint must not incur production spending. What operational boundary helps? Merge staging and production rate limits without labels Disable token accounting entirely Dedicated budgets, monitoring and quotas aligned to the staging deployment Only lower output length limits without enforcing usage budgets Explanation Correct Answer C:Environment-specific limits make test capacity and costs observable and bounded. Incorrect Answers A:Shared untagged usage defeats attribution and predictability. B:Lost telemetry hides unexpected charges instead of limiting them. D:Shorter answers may reduce tokens but do not bound test request volume and cost.
Question 218 An agent asks a human to approve a specific invoice, then changes the amount before payment. How should approval be handled? Bind approval to immutable invoice, amount and payee details; reapprove changed actions Ask the model whether the change feels small Use the old approval for any amount from the same supplier Hide the changed amount and execute anyway Explanation Correct Answer A:Approval for one financial payload does not extend to a modified transaction. Incorrect Answers B:Model judgment does not replace delegated approval. C:Material change breaks the original authorization scope. D:Concealment violates transaction integrity and consent.
Question 219 An agent must send a sensitive document externally, but the recipient changed after approval. Does the earlier approval suffice? Yes; any previous send approval covers all future addresses No; revalidate the final recipient and document against the approved scope Yes if the agent's confidence is high Yes if the document is shorter than a token threshold Explanation Correct Answer B:An approval tied to one destination does not automatically authorize another. Incorrect Answers A:Permission should be bound to the actual external action and target. C:Model confidence cannot extend human or policy authorization. D:Document length does not determine allowed recipients.
Question 220 A source assistant has two tools: `search_policy` and `update_policy`. The user asks what the current rule says. What choice is proper? Create a new policy document automatically Invoke only the read-only policy lookup tool Invoke both tools to be thorough Update the policy to match the user's question Explanation Correct Answer B:An informational request does not authorize changing official policy. Incorrect Answers A:Content creation is not warranted by the lookup task. C:Unnecessary writes could alter current company guidance. D:Reading a rule is not a request to modify it.