Question 190 A model self-critique identifies a likely error but proposes an untested correction. What should the release system do? Treat self-critique as independent human approval Discard the passing previous tests after one suggestion Apply the targeted change and rerun affected deterministic tests and evaluations Release the critique text as a substitute for a fixed product Explanation Correct Answer C:A critic suggestion is a hypothesis until revised behavior is checked. Incorrect Answers A:Model review cannot be misrepresented as external SME signoff. B:Unnecessary evidence loss does not strengthen confidence. D:Observations alone do not correct application behavior.
Question 191 A compliance assistant should ground extracted statements in tables, images and paragraphs. Which processing approach is suitable? Only exact keyword search over document titles Multimodal content enrichment with layout, visual descriptions and source references Flat OCR alone with no structure or images Text generation that invents unlabeled figures Explanation Correct Answer B:Different evidence modalities require a shared traceable representation. Incorrect Answers A:Titles cannot support detailed cross-modal claims. C:OCR text may miss chart meaning and table relationships. D:Synthetic data cannot substitute for retrieved visual evidence.
Question 192 A marketing assistant performs content creation and evaluation with one model. What is a stronger review criterion? Disable all independent automated checks to reduce cost Only compare output character counts Accept every response that sounds confident Evaluate against predeclared brand, factual, accessibility and safety requirements Explanation Correct Answer D:Explicit rubrics distinguish actual deficiencies from generated self-praise. Incorrect Answers A:Unexamined regressions can reach customers. B:Length does not establish usefulness or brand safety. C:Confidence is not evidence of factual or policy compliance.
Question 193 A review agent needs internal specifications but must never invoke order-creation tools. Which tool-list decision supports least privilege? Expose order creation but write NEVER USE in the prompt Expose read-only knowledge tools and omit mutation capabilities Give the agent a write-capable browser session Use an admin API key shared with all specialist agents Explanation Correct Answer B:Tool availability itself constrains which external effects a reviewer can attempt. Incorrect Answers A:Text restrictions alone are weaker than removing an unnecessary capability. C:A browser with mutation privileges violates the review-only assignment. D:Shared authority greatly increases the impact of misrouting or injection.
Question 194 A picture includes a QR code that might route to a malicious site. What should a safe agent do? Automatically open the URL and submit saved credentials Treat decoded content as untrusted data and inspect its destination without automatic navigation Delete the image before recording the destination Treat any QR code as an administrator instruction Explanation Correct Answer B:Recognized visual text or codes do not authorize external actions. Incorrect Answers A:Image content may contain an attack-controlled destination. C:Erasing evidence impairs safe analysis and incident reporting. D:A printed code has no instruction authority.
Question 195 A production workflow repeatedly copies a 10,000-token instruction prefix. What cost optimization should be assessed? Double the prefix to emphasize its importance Supported prompt caching and elimination of redundant context Remove all system requirements regardless of safety Move all instructions into user messages without scoping Explanation Correct Answer B:Reused prefixes may be cached or shortened without losing necessary instructions. Incorrect Answers A:More repeated input increases token processing cost. C:Critical behavior constraints must not be discarded for savings. D:Unstructured prompts may increase token use and reduce authority clarity.
Question 196 A Content Understanding analyzer extracts object labels but the app needs a tailored JSON field schema. What should be configured? Only increase the model response length Only adjust the model’s output token limit without an analyzer field schema Only change the app's logo A suitable custom analyzer with explicit requested fields and output types Explanation Correct Answer D:Field schemas make visual extraction usable by downstream applications. Incorrect Answers A:Longer prose is not a structured field contract. B:More tokens cannot create the required typed JSON contract for visual fields. C:Brand styling does not define analyzer outputs.
Question 197 A Foundry workload is hitting both RPM and TPM limits. Which mitigation is most directly relevant? Disable source citations in all responses automatically Replace all model prompts with larger contexts Increase the number of database columns in the search index Control request concurrency, token budgets, backoff and approved capacity expansion Explanation Correct Answer D:Rate limiting tracks request and token consumption across the deployed workload. Incorrect Answers A:Citations are not the root cause of request-volume saturation. B:Larger prompts can worsen token-per-minute pressure. C:Index schema size does not raise model rate quota.
Question 198 An enterprise deployment has unexplained cost spikes because many agents repeat the same context. What should be measured? Only total model requests without measuring repeated prompt prefixes Only the number of Git commits this month Only the total staff directory size Prompt token volumes, repeated-prefix caching and tool-loop frequency Explanation Correct Answer D:Token analytics reveal avoidable repeated input and excessive iteration. Incorrect Answers A:Call counts cannot identify redundant token-heavy context driving cost. B:Version-control activity does not measure inferencing costs. C:User count alone does not show per-request token inefficiency.
Question 199 A corporate RAG service has private documents that vary by user entitlement. What retrieval control should it apply? Permission-aware query filtering under the requesting principal's effective rights Trust a user-supplied role name with no server check Store one unfiltered full index snapshot in every user session Retrieve all documents and ask the model to self-censor Explanation Correct Answer A:A matching chunk must remain inaccessible if the caller lacks source access. Incorrect Answers B:Unverified claims do not establish document authorization. C:Duplicate exposure increases the risk of cross-user leakage. D:Sensitive content would already enter the unauthorized prompt context.
Question 200 A customer asks a grounded bot about last week's pricing, but the index holds today's price list only. What should the bot require? Historical versioned documents or a verified time-series source for the requested date Use the nearest semantic hit without checking effective date Assume today's prices always applied last week Invent a likely discount amount from marketing copy Explanation Correct Answer A:Current pricing alone cannot establish past transaction terms. Incorrect Answers B:Similarity does not verify the requested time period. C:Pricing policies may change between dates. D:Generated estimates do not establish historical facts.