Question 250
A customer support response passes factual review but contains a disallowed promise of guaranteed compensation. What second review criterion is relevant?
- Declare success because no factual errors were found
- Ask a model to make the promise more persuasive
- Remove the customer request so the reviewer sees less context
- Check prohibited commitments apart from factual accuracy
Explanation
Correct Answer
D:A response can be true about one fact yet violate service commitments. Distinct checks make the failure diagnosable.
Incorrect Answers
A:Factual correctness is necessary but does not exhaust the business policy. The promise still needs correction.
B:Persuasive phrasing compounds the policy breach. The prohibited commitment must be removed.
C:The request is needed to judge whether the response overcommits. Hiding it weakens review.
Question 251
A three-pass extraction pipeline uses one model to parse, one to check schema, and one to verify sources. What improves reliability most?
- Keep each pass's contract distinct and pass structured findings into targeted repair
- Discard validation failures and average the three responses
- Have all three passes generate independent free-form answers with no reconciliation
- Treat schema compliance as sufficient proof of source accuracy
Explanation
Correct Answer
A:Separation of concerns allows each checker to test a measurable property. The repair stage should address only verified defects.
Incorrect Answers
B:Averaging inconsistent values does not establish evidence. The pipeline must inspect actual source support.
C:Multiple outputs do not automatically yield one valid record. Conflicts need a defined decision process.
D:The passes check different properties. A syntactically valid record can still misread a value.
Question 252
A reviewer and author repeatedly disagree over a citation because each sees different source versions. What should the workflow standardize?
- Hide timestamps to make the documents appear consistent
- Choose whichever opinion is stated more confidently
- Ask more agents to vote without sharing source versions
- Reuse versioned evidence and record source identifiers
Explanation
Correct Answer
D:Version alignment prevents spurious disagreements caused by changed inputs. The reviewer can then evaluate the actual claim against the cited version.
Incorrect Answers
A:Removing metadata conceals the root cause of the mismatch. Version information is essential.
B:Confidence is not a substitute for consistent evidence. The disagreement may be caused by input drift.
C:Additional votes cannot resolve inconsistent evidence. Source harmonization must precede consensus.
Question 253
A product team uses two independent model runs to assess a risky architectural decision. How should disagreement be treated?
- Accept the answer from the more verbose model automatically
- Assume agreement between two models proves the design is safe
- Investigate the supporting evidence, not merely reviewer votes
- Average incompatible architecture recommendations into one vague choice
Explanation
Correct Answer
C:Different conclusions can expose ambiguous requirements or weak evidence. A responsible synthesis resolves the underlying uncertainty.
Incorrect Answers
A:Output length does not establish soundness. The reasoning must fit the constraints.
B:Correlated model errors remain possible. Independent tests and security review still matter.
D:Averaging decisions may satisfy neither requirement. Tradeoffs need explicit evaluation.
Question 254
A final report contains duplicated paragraphs after a model revises several sections independently. What postprocessing helps?
- Review section consistency, redundancy and contradictions
- Remove all section headings and hope duplicates become less visible
- Publish all paragraphs because each was generated by a separate pass
- Replace all evidence with generic summaries
Explanation
Correct Answer
A:Local revisions can introduce global editorial defects. A final synthesis pass can consolidate without changing validated facts.
Incorrect Answers
B:Formatting cannot eliminate repetitive content. The underlying passages need editorial resolution.
C:Independent origin does not justify duplication. The finished artifact needs coherent organization.
D:Removing detail may conceal conflicts and weaken support. The integration should preserve traceable facts.
Question 255
A research team wants confidence in a cross-country trend report, but all reviewers used the same stale dataset. What is the primary limitation?
- Additional reviewers guarantee that data becomes current
- Repeated model agreement is not independent source verification
- The report is invalid solely because it has several reviewers
- Polishing narrative language removes the need to update numbers
Explanation
Correct Answer
B:Multiple passes over identical outdated evidence can repeat the same mistake. Data freshness and provenance must be checked directly.
Incorrect Answers
A:Review volume cannot update a dataset. New authoritative records are required when freshness matters.
C:Multiple reviews can improve quality when they test distinct criteria. The problem is shared stale input.
D:Style changes do not fix outdated evidence. Factual validity depends on source recency and accuracy.
Question 256
A support agent handles a conversation spanning many turns and forgets a customer's verified delivery address after context compaction. What state strategy is most reliable?
- Assume compacted summaries always retain every earlier fact
- Ask Claude to guess the original address from later replies
- Persist the verified address in scoped durable case state
- Disable verification because the user already mentioned an address
Explanation
Correct Answer
C:Compaction may omit earlier conversation details. Critical facts should be retained in authoritative external state with source and validity information.
Incorrect Answers
A:Summaries prioritize information and can lose details. An exact operational address deserves explicit persistence.
B:Inference cannot restore a verified personal record. Guessing risks misdelivery and privacy mistakes.
D:Past mention does not prove ongoing correctness. The system should restore and, where needed, revalidate the trusted fact.
Question 257
A long-running research agent repeats the same web searches after context pruning. What should its recovery design preserve?
- Only the most recent generated sentence
- A research ledger of queries, sources and unresolved gaps
- A confident claim that no further research is necessary
- Every raw webpage in the active prompt forever
Explanation
Correct Answer
B:Durable checkpoints prevent duplicated work while keeping evidence navigable. The agent can resume from remaining questions rather than re-search everything.
Incorrect Answers
A:The latest prose does not represent which sources have been checked. Repetition is likely without a progress record.
C:Confidence without task-state evidence cannot establish completeness. The ledger should record specific coverage.
D:Unbounded raw content crowds out current instructions and tasks. Preserve concise references and findings instead.
Question 258
A document assistant's context is nearly full with verbose tool outputs. What should it keep before continuing?
- An invented compressed record of details it cannot verify
- All intermediate tool logs verbatim regardless of relevance
- Only the agent's last friendly message
- A summary of decisions, open tasks, facts and source links
Explanation
Correct Answer
D:Context management should retain the information needed for future decisions. Irrelevant debug logs can be dropped or archived externally.
Incorrect Answers
A:Compression should reflect observed history. Fabrication turns missing context into false inputs.
B:Raw outputs can consume context while adding little value. Long-running tasks need prioritization.
C:A social response omits technical status and evidence. Continuation requires actionable state.
Question 259
A contract review agent must remember a customer's signed exception across long conversations. Where is the strongest location for that authorization?
- A guessed approval status inferred from customer sentiment
- An informal conversational memory that says approved once
- A summary that removes the approver and time to save tokens
- Durable approval with identity, scope, expiry and audit ID
Explanation
Correct Answer
D:The model's context should reference the approval but not serve as the authority. Expiring permissions require live verification at execution.
Incorrect Answers
A:Approval is a factual policy artifact, not an emotional judgment. Trusted records are required.
B:A chat statement may become stale or be dropped. It cannot replace an auditable authorization record.
C:Missing scope and expiry can lead to invalid later actions. Critical approval metadata must survive compression.
Question 260
A multi-day research agent must survive context compaction and process restarts while avoiding repeated external writes. Which TWO state records matter most? Choose TWO.
- An unlimited transcript with no indexing or selection strategy
- Every greeting and stylistic aside in the active context
- A durable ledger of completed work, remaining tasks and source references
- A transaction execution record with identifiers and verified outcomes
- A model-generated guess about which operations succeeded
Explanation
Correct Answers
C:The ledger preserves progress and evidence beyond prompt truncation. It lets the resumed agent choose the next necessary investigation.
D:Conversation history alone may not show whether a write committed before interruption. Durable transaction state prevents duplicate effects.
Incorrect Answers
A:Raw history may overflow context while hiding critical facts. Curated external state is more useful.
B:Conversational filler does not establish progress or transaction outcomes. It is low-priority state for technical recovery.
E:Speculation is not a trustworthy execution ledger. Recovery must consult actual backend evidence.