Question 40 An agent summarizes a disputed technical incident from incomplete records. Which TWO practices best preserve reliability? Convert all inferences into settled facts for a simpler story. Label each conclusion as confirmed, inferred or unresolved with supporting evidence. Suppress contradictory records that make the summary less concise. Record timestamps and source systems for critical indicators. Attach plausible but unvisited URLs to make the report appear sourced. Explanation Correct Answers B:Readers can distinguish observation from interpretation. D:Provenance allows independent reconstruction and checks for conflicting records. Incorrect Answers A:That overstates evidence. C:Relevant conflict must be addressed. E:Unverified citations are misleading.
Question 41 A review bot runs on untrusted pull requests. What isolation is most important for its agent? Skip provenance and audit records entirely. Avoid privileged secrets and write permissions when reading untrusted repository changes. Expose production release credentials to every PR job. Execute arbitrary commands from diff comments. Explanation Correct Answer B:Malicious PR text can attempt prompt injection or command misuse. Incorrect Answers A:Traceability helps identify misuse and errors. C:Untrusted code should not receive deployment authority. D:Comments are not trusted instructions.
Question 42 A report writer may search public material, but confidential customer records are needed only for a compliance subtask. What arrangement fits? Expose private-record tools only during the authorized compliance branch. Paste private records into every system message. Remove access logs from private tools to reduce latency. Let all report-writing steps access the full customer database. Explanation Correct Answer A:Dynamic scoping preserves required work without exposing confidential data to unrelated tasks. Incorrect Answers B:Global prompts broaden sensitive-data exposure. C:Auditing remains important and is not the cause of unnecessary exposure. D:This violates data minimization for public research.
Question 43 A subagent is asked to summarize a huge log file but its output is larger than the parent’s available context. Which handoff is best? Drop all errors and rare events to fit the window. Paste the entire log into every subsequent worker prompt. Return bounded findings with precise line references and an external artifact link. Return only “looks fine” with no evidence or uncertainty. Explanation Correct Answer C:The coordinator can reason on selected evidence while retrieving details when needed. Incorrect Answers A:Potentially important anomalies must be retained or explicitly flagged. B:Repeating raw data causes avoidable context pressure. D:The coordinator cannot assess coverage or validate the claim.
Question 44 A nightly classification export must prove that all input records have outputs. What reconciliation should run? Compare only file byte counts. Assume the job completion status guarantees 100% item success. Spot-check the first result only. Compare submitted IDs against successful, failed and still-pending result IDs. Explanation Correct Answer D:An exhaustive set reconciliation detects omissions and duplicates. Incorrect Answers A:Similar file sizes do not prove one output per request. B:A completed job may contain item-level failures. C:That cannot account for the complete input set.
Question 45 A research summary has enough evidence for most findings but cannot corroborate one claimed benchmark. What should it do? Remove source citations so readers cannot check the claim. Withhold every confirmed finding until all questions are resolved. Deliver supported findings and clearly label the benchmark as unverified rather than inventing a number. Estimate the benchmark and present it as measured. Explanation Correct Answer C:Truthful partial completion separates facts from unresolved claims. Incorrect Answers A:That reduces transparency. B:The remaining uncertainty does not invalidate supported facts. D:A guess is not evidence.
Question 46 A finance reviewer should only read invoices, while a billing executor can post adjustments after approval. Which distribution follows least privilege? Expose posting tools to the reviewer but ask it not to use them. Separate read-only review tools from approved mutation tools by agent role. Share an unrestricted billing superuser token between both. Give the executor only web search and no billing access. Explanation Correct Answer B:Each specialist receives capabilities matching its delegated authority. Incorrect Answers A:Prompt-only restrictions are weaker than capability boundaries. C:Unnecessary privilege and common credentials raise risk. D:It cannot perform its authorized adjustment task.
Question 47 A documentation skill is intended for an entire team but needs a careful invocation boundary. Which TWO attributes should it specify? An unlimited production-deploy permission. A requirement to change all files, regardless of requested scope. A limited set of tools consistent with reading and editing documentation. A clear command or skill description with expected inputs and output format. Credentials copied into the markdown body for ease of use. Explanation Correct Answers C:Least privilege keeps a documentation skill from making unrelated changes. D:Users and the agent need to know when and how the reusable workflow applies. Incorrect Answers A:Documentation work does not require release authority. B:Broad mutation contradicts a targeted documentation task. E:Shared skill content should not carry secrets.
Question 48 A contract extractor outputs a numeric total that differs from the sum of its listed line items. Which check is necessary? Round every field until the numbers appear close. Accept the total because it has the correct numeric type. Recompute the arithmetic from authoritative fields and reject inconsistent totals. Ask for a more confident narrative explanation instead. Explanation Correct Answer C:Deterministic calculation catches a semantic error that schema validation may miss. Incorrect Answers A:Changing source values hides the discrepancy. B:Type checks do not prove arithmetic correctness. D:Confidence wording cannot repair a calculation.
Question 49 A host logs only the final assistant text for a long-running analysis. An incident investigator cannot explain why a report used a particular record. Which trace element is missing? The model’s maximum output-token configuration alone. The ordered tool-call and tool-result history associated with the run. A count of total final-answer characters. A screenshot of the chat window’s theme settings. Explanation Correct Answer B:The final prose alone does not show what evidence the agent observed or what actions it took. Incorrect Answers A:A limit setting cannot reconstruct the actual information flow. C:Output length is not an audit trail of tool observations. D:Visual theme does not reveal evidence provenance or execution order.
Question 50 An assistant has changed an implementation and sees the relevant unit tests pass, but the public API contract may have changed. What next? Delete the API specification because the implementation changed. Inspect the public interface and run focused integration or contract checks for affected consumers. Declare all external clients compatible without checking them. Replace integration tests with a summary of code style. Explanation Correct Answer B:Local tests do not prove compatibility across module boundaries. Incorrect Answers A:The published contract remains a relevant requirement. C:Compatibility cannot be inferred from internal unit tests alone. D:Style is not interface behavior.
