Question 140 An application writes model-extracted invoice fields to a finance database. Which TWO safeguards are essential? Treat HTTP 200 as proof all extracted amounts are true Invent missing supplier IDs to satisfy required fields Enforce arithmetic and business rules plus authorized write permissions Validate structured field types and source identifiers before accepting extraction Allow model output to specify unrestricted SQL commands Explanation Correct Answers C:Correct schema does not prove total consistency or mutation authority. D:Parsing should reject malformed or misattributed records. Incorrect Answers A:Transport success does not validate financial semantics. B:Fabricated identifiers can link invoices to wrong suppliers. E:Generated text must not control arbitrary database writes.
Question 141 A document agent must use OCR, search and summarization. OCR finds no characters in a scanned page. What should the workflow do? Ask the summarizer to guess the missing words Mark the page empty without recording OCR failure Discard the entire document and declare the job successful Flag extraction failure and retry or escalate before claiming the page was blank Explanation Correct Answer D:A failed parser is not evidence that no text exists. Incorrect Answers A:Generation cannot recover source evidence reliably. B:Silent omission hides relevant document content. C:Lost pages violate ingestion completeness.
Question 142 A team needs to diagnose whether safety filtering or model reasoning caused a rejected request. What trace fields help? Record only that the user received a refusal without any policy stage Record the model deployment ID without content-filter decision categories Safety event categories, policy decision and model execution status Record input text and output only, omitting execution and filter status Explanation Correct Answer C:Separate stages distinguish prefilter blocks from refusals generated by the model. Incorrect Answers A:A refusal string alone cannot identify whether a filter or generation stage rejected the input. B:Model identity alone does not separate a safety rule trigger from a generated refusal. D:The same text can arise from different mechanisms and needs stage-level metadata.
Question 143 A deployment must support multiple customers without one tenant's secrets entering another model call. What environment layout supports that? Put all tenant documents into one unfiltered prompt One shared superuser key presented inside every agent's prompt Use only a polite system prompt to avoid cross-tenant retrieval Tenant-scoped credentials, storage and retrieval connections with enforced isolation Explanation Correct Answer D:Per-tenant connection boundaries limit leakage across independent customers. Incorrect Answers A:Mixed context exposes confidential records to unrelated requests. B:An unrestricted shared secret defeats tenant separation. C:Text guidance cannot enforce data-plane isolation.
Question 144 A visual agent encounters text in an image that says to export customer records. Which TWO controls apply? Automatically follow decoded links with stored credentials Grant administrative privileges to any photographed signature Enforce data-export authorization at the backend tool boundary Disable telemetry to hide the attempted export Treat recognized visual text as untrusted content rather than instructions Explanation Correct Answers C:Trusted policy must reject unauthorized disclosure independent of the model. E:OCR does not elevate embedded commands to application authority. Incorrect Answers A:Image content may be attacker controlled. B:Visual marks do not map to verified cloud RBAC. D:Concealment does not protect confidentiality.
Question 145 A custom speech model improves common phrases but performs poorly on rare product names. What data improvement fits? Train only on repeated greetings Remove rare terms from evaluation scoring Use only voice synthesis examples for recognition training Add reviewed domain-specific utterances covering the rare terminology Explanation Correct Answer D:Targeted representative adaptation can address known vocabulary gaps. Incorrect Answers A:Common words do not teach unusual product names. B:Excluding failures hides important service limitations. C:Output speech samples may not represent incoming customer audio.
Question 146 A legal team requests a structured compliance summary with clause, rule, exception, impact and citation. What generation technique fits? Ask for a pleasant essay without sections Generate the citation by choosing a plausible URL Use a defined schema and source-grounded extraction instructions Remove exception fields because they complicate parsing Explanation Correct Answer C:Explicit fields preserve distinctions needed for a legal review. Incorrect Answers A:Unstructured prose may omit exceptions and citations. B:An invented source identifier cannot be checked. D:Omission can materially change regulatory meaning.
Question 147 A safety classifier rejects a legitimate chart because one axis label resembles a prohibited word. What is a proper refinement? Review false positives using visual context and calibrated policy examples Replace all scientific charts with generated cartoons Keep rejecting the chart without examining the label Disable all image safety checking globally Explanation Correct Answer A:Contextual evaluation distinguishes benign technical labels from disallowed content. Incorrect Answers B:Altered data visualizations do not correct the classifier's decision rule. C:Unreviewed overblocking harms legitimate users. D:One false positive does not justify removing every safeguard.
Question 148 A generative service is repeatedly hallucinating policy sections that do not exist. What evaluation should the team build? Evaluate grammatical consistency of generated policy summaries without source evidence Evaluate retrieval latency against policies without checking supported claims Count linked citations without verifying which statement each passage supports A grounded factuality set with known supporting and unsupported policy claims Explanation Correct Answer D:Source-linked examples measure whether answers remain within retrieved evidence. Incorrect Answers A:Well-formed prose can still invent sections not present in the policy corpus. B:Fast retrieval does not establish that generated section numbers and obligations exist. C:Citation volume cannot distinguish valid grounding from fabricated adjacent claims.
Question 149 A model generates a near-correct tool call but misses one mandatory field. How should repair be controlled? Fill the missing account ID with a plausible guess Retry forever without explaining the missing field Return validation errors and allow a bounded correction followed by revalidation Send the invalid call directly to the backend anyway Explanation Correct Answer C:Targeted feedback can fix structural omissions without accepting malformed calls. Incorrect Answers A:Identifiers must come from authorized source data. B:No corrective signal encourages repeated failures. D:Missing required fields can cause unintended or rejected actions.
Question 150 A factory video includes workers changing stations, and the app must count actions per station. What pipeline feature helps? Choose random station IDs based on appearance Temporal object tracking and segmented action attribution Count all detected people in the opening frame only Use only a transcript of spoken instructions Explanation Correct Answer B:Counting requires linking observed activities to locations across time. Incorrect Answers A:Identifiers need consistent visual or operational grounding. C:Initial presence does not measure later station actions. D:Words do not prove which physical actions occurred.