Question 110 A fraud analyst requires precise anomaly rules plus natural-language explanations. What hybrid design is appropriate? Ask a generative model to override all fraud limits privately Replace audit records with a natural-language confidence claim Deterministic rules for threshold decisions with a model explaining evidence Use only embeddings to approve all payments Explanation Correct Answer C:Business invariants remain testable while the model supplies understandable context. Incorrect Answers A:Probabilistic output should not supersede explicit controls. B:Confidence wording is not proof of enforced policy. D:Similarity vectors do not validate financial business rules.
Question 111 A critic flags a change as unsafe but provides no reproducible evidence. How should the team treat it? Delete every reviewed artifact and restart production Keep the concern nonblocking until validated against specific acceptance criteria Reject all future changes whenever a critic expresses discomfort Treat the critic as an external regulatory authority Explanation Correct Answer B:Substantive QA distinguishes a concrete defect from unsupported preference. Incorrect Answers A:A vague flag does not justify discarding completed valid work. C:Unfounded objections should not automatically block delivery. D:A model's opinion is not independent expert certification.
Question 112 A developer removes a model deployment while an agent still references it. What should an operational safety check catch? Confirm the deployment has a friendly name but skip its dependency graph Dependent client and tool configuration before destructive deployment deletion Verify that a replacement model is available in another region only Inspect only successful historical inference requests to that deployment Explanation Correct Answer B:Dependency inventory prevents leaving agents with unresolved model references. Incorrect Answers A:A display name cannot reveal agents, jobs and client endpoints that still reference its deployment ID. C:Replacement availability does not reconfigure every dependent agent or project connection. D:Past requests do not prove current clients will survive deletion of an active dependency.
Question 113 A self-critique pass always agrees with the first model's answer because it sees the original key. What review design is stronger? Judge solely by how long the reasoning text is Treat confidence percentages as proof of correctness An independent source-backed evaluation without anchoring to the preferred key Run the same answer through the same anchored prompt ten times Explanation Correct Answer C:Blinded critical checks can identify reasoning errors that repeated agreement misses. Incorrect Answers A:Length does not measure key exclusivity or factual support. B:Model scores may be poorly calibrated or circular. D:Correlated outputs do not constitute independent adjudication.
Question 114 A voice interface returns audio but no readable transcript for deaf or hard-of-hearing users. What accessibility improvement fits? Only change the audio codec Remove all textual controls from the interface Provide a synchronized textual alternative for the generated speech Increase audio volume alone Explanation Correct Answer C:Accessible interfaces should not require listening to retrieve content. Incorrect Answers A:Encoding does not supply a readable equivalent. B:That further reduces accessibility options. D:Volume does not replace caption or text access.
Question 115 A compliance assistant must compare two regulations and cite which date governs. What retrieval metadata matters most? Only the alphabetical order of results Effective dates, version status and source authority for each document Only number of tokens in each regulation Only the document retrieval similarity score for the two regulations Explanation Correct Answer B:Superseded texts can remain searchable but must not be mistaken for current rules. Incorrect Answers A:Search order is not regulatory precedence. C:Document length does not establish current applicability. D:Similarity ranking does not establish which regulatory version was effective on a specific date.
Question 116 An enterprise asks how to prove that an agent did not bypass required approval for a write. What evidence should be retained? Log the final tool outcome without the associated approval identity Store the agent role permissions but not the request-specific policy evaluation Archive only the agent's description of having obtained approval Policy checks, requested tool input, approval ID and resulting tool outcome Explanation Correct Answer D:An execution trace can show which authorization gate controlled the mutation. Incorrect Answers A:A successful mutation alone does not prove an authorization gate approved that exact action. B:Role configuration does not demonstrate which approval or denial governed this operation. C:Model prose is not a durable approval receipt tied to the actual tool arguments.
Question 117 A Content Understanding analyzer returns a field confidence score but the field conflicts with a verified database record. What should happen? Ignore the original document after extraction Investigate source and reference conflict before using the field in a mutation Delete the database record to remove the conflict Automatically accept any field with a high score Explanation Correct Answer B:A model confidence estimate does not override authoritative consistency checks. Incorrect Answers A:Source evidence is needed to adjudicate discrepancies. C:Destructive changes cannot resolve extraction uncertainty safely. D:High confidence does not guarantee the entity and value are correct.
Question 118 A customer service app should use a new prompt only after it passes policy-based regression tests. Where should configuration promotion occur? Use the model's own unverified success statement Publish only a screenshot of the draft prompt Replace prompts directly in production without a version CI/CD release gate on a versioned prompt with documented evaluation results Explanation Correct Answer D:The approved prompt revision must be the artifact deployed by the pipeline. Incorrect Answers A:Self-reporting cannot establish that test cases actually passed. B:A screenshot is not a deployable versioned configuration. C:Untracked hot edits impair rollback and reproducibility.
Question 119 A model keeps requesting the same invalid tool parameter despite receiving a type error. What feedback-loop improvement helps? Return only a generic 'failed' and keep retrying Allow unlimited retries until one randomly succeeds Expose field-specific validation errors and bounded retry limits Remove parameter validation from the backend Explanation Correct Answer C:Actionable error details let the agent correct input without infinite repetition. Incorrect Answers A:No corrective information means the same invalid request may recur. B:Unbounded invalid calls waste capacity and may trigger abuse limits. D:Malformed arguments can then reach unsafe business logic.
Question 120 A multi-agent app shows increased latency after adding a critic agent to every trivial turn. What operational response fits? Always launch three more critics Mark all critic responses as factual proof Measure critic utility and invoke it conditionally for material-risk tasks Suppress latency reports until users stop complaining Explanation Correct Answer C:Proportional review preserves useful assurance without unnecessary routine inference. Incorrect Answers A:Additional universal reviewers increase latency and cost further. B:Model agreement is not independent external validation. D:Removing observability does not reduce execution time.