Question 1 An agent has a strict 15-tool-call budget and has used all 15 without reaching an answer. What should the host do? Erase prior verified evidence and restart from zero Stop execution and report verified progress and unresolved work Silently allow unlimited calls because the task matters Generate a fake successful tool receipt Explanation Correct Answer B:A host-side limit must bound further external actions and completion claims. Incorrect Answers A:A bounded failure should preserve useful work for recovery. C:Unbounded loops violate the configured resource and safety policy. D:Fabricated results cannot justify a completed-state assertion.
Question 2 An incident response agent sees a suspicious file attachment and proposes isolating a workstation. What should govern action? An embedded attachment instruction to run an admin tool The model's tone of certainty as the sole signal Configured evidence thresholds and delegated containment authority A single guessed indicator automatically triggers global isolation Explanation Correct Answer C:Security response must balance confidence and operational impact under policy. Incorrect Answers A:Untrusted file content has no operational authority. B:Generated confidence cannot enforce response thresholds. D:Weak evidence can cause significant unnecessary disruption.
Question 3 A loan-application agent's explanation names an irrelevant protected attribute as a risk factor. What should the reviewer verify? Only the number of pages in the explanation Only grammatical correctness of the generated rationale Decision features, policy legitimacy and fairness against relevant protected groups Only how confident the model sounds Explanation Correct Answer C:Explainability must identify potentially impermissible predictors and causal assumptions. Incorrect Answers A:Length cannot prove proper feature selection. B:A fluent explanation can still justify an unfair or unsupported decision. D:Tone does not establish lawful use of customer attributes.
Question 4 An extraction pipeline processes digital PDFs and scanned photographs. What routing decision improves efficiency? Apply OCR to every binary file regardless of format Generate textual content whenever extraction fails Assume every PDF is text-searchable Use digital parsing when text is embedded and OCR/layout analysis for scanned sources Explanation Correct Answer D:Input-aware extraction reduces unnecessary OCR without ignoring image-only content. Incorrect Answers A:Unnecessary recognition can reduce fidelity and increase processing cost. B:Synthetic words are not recovered source evidence. C:Image-only pages often require OCR.
Question 5 A bank's summarizer is asked to report quarterly growth from two uploaded spreadsheets. What check is required before using generated percentages? Calculate the growth from validated cells and handle missing denominators explicitly Ask the model to use a plausible growth percentage Select the highest number observed anywhere in the spreadsheets Use only the shape of the final JSON object Explanation Correct Answer A:Deterministic arithmetic prevents a fluent model from inventing or miscomputing figures. Incorrect Answers B:An estimate is not a calculated result from the supplied files. C:An unrelated maximum cannot represent the period-over-period change. D:Valid structure does not establish numerical accuracy.
Question 6 A report-writing agent receives a successful empty search result. How should it distinguish that from search failure? Treat all errors as proof of no documents Invent a result to keep the report complete Treat all empty outputs as failures Use separate typed states for zero matches and execution errors Explanation Correct Answer D:A successful empty result can support a no-results statement within search scope. Incorrect Answers A:A failed search may not have examined the corpus. B:Fabricated findings misrepresent source availability. C:A valid query may return no records without error.
Question 7 A contact-center agent must understand a caller's spoken question. Which first stage is needed for a text-reasoning pipeline? Speech-to-text output synthesis without an input audio recognizer A vector search query with no transcript or audio features Speech-to-text transcription with language and acoustic configuration Text-to-speech synthesis before receiving audio Explanation Correct Answer C:The agent needs a recognized utterance before a text model can reason on words. Incorrect Answers A:Synthesizing speech from text produces audio but cannot decode the caller’s supplied waveform. B:Retrieval cannot identify the user's actual utterance without representation. D:Synthesis produces sound rather than interpreting the customer's speech.
Question 8 A document Q&A system must show which version of a policy was effective on a historical transaction date. What should queries include? Always select the newest document without regard to transaction date Ask the model to reconstruct old policies from memory Choose documents sorted only by filename Temporal filters and stored effective-date metadata Explanation Correct Answer D:Historical applicability can differ from today's current policy. Incorrect Answers A:Current rules may not have applied to older events. B:Generation cannot replace archived authoritative text. C:Names do not determine legal effective periods.
Question 9 A web application wants AI-generated FAQs but must avoid publishing draft answers that fail accessibility checks. What workflow fits? Generate drafts, run defined quality checks, then authorize publication separately Delete failed evaluation logs to simplify release Use a higher model confidence to replace accessibility validation Publish every model result automatically Explanation Correct Answer A:Content creation and release should remain distinct controllable stages. Incorrect Answers B:Lost failures make quality regressions harder to address. C:Self-reported certainty cannot test actual markup requirements. D:Unchecked answers can include errors or inaccessible formatting.
Question 10 A Content Understanding pipeline must produce agent-ready JSON from many mixed documents. Which TWO implementation controls help? Schema checks and per-file success/error reconciliation Silently fill unknown fields with generated plausible values Document classification and routing to appropriate configured analyzers Use one invoice-only schema for every document type Treat overall batch completion as proof every record succeeded Explanation Correct Answers A:A complete job must account for every uploaded input and field contract. C:Receipts, contracts and IDs require different extraction fields and expectations. Incorrect Answers B:Fabricated records undermine downstream factual accuracy. D:Noninvoice files may not contain mandatory invoice fields. E:Per-file errors can persist in finished batches.