Question 240
A nightly extraction job is rerun after a crash, and duplicate entries appear in storage. What should its reconciliation layer add?
- Use a new random key for each retry regardless of source identity
- Trust that asynchronous jobs never need retries
- Idempotent writes keyed by stable input or custom request identifiers
- Discard the source identifier after output retrieval
Explanation
Correct Answer
C:The same logical request may be processed more than once during recovery. Deduplication prevents repeated persistence of identical work.
Incorrect Answers
A:New keys make duplicates indistinguishable from new records. Reconciliation needs stable identity.
B:Jobs can fail or be interrupted. Recovery should account for repeated submissions.
D:Source provenance is needed to detect duplicates and connect records to original documents.
Question 241
A batch processor must report progress to operators while some requests remain outstanding. What should it track?
- Only a single Boolean labeled all done from the start
- Only the number of characters in completed answers
- Only the size of the original uploaded file
- Counts and IDs for completed, failed and pending requests
Explanation
Correct Answer
D:Operational status is a property of individual work items. These categories support retry policy and honest progress reporting.
Incorrect Answers
A:Premature completion conceals outstanding work. The system should derive completion from actual request statuses.
B:Text length does not indicate which source requests remain unfinished. Operators need per-item state.
C:Input size is unrelated to how many model requests succeeded. It does not provide reliable progress data.
Question 242
A team submits batch requests with inconsistent system instructions and sees unpredictable classification. What should be standardized?
- The order in which asynchronous responses finish
- Task prompt, category definitions and shared output schema
- The color of each source document's filename
- The number of workers instead of the classification contract
Explanation
Correct Answer
B:Batch processing changes scheduling, not semantic consistency. Shared contracts make outputs comparable and easier to validate.
Incorrect Answers
A:Completion order is not under semantic control and should not affect classifications. Correlate by IDs instead.
C:Filename color has no bearing on model criteria. The relevant variation is instruction and evidence.
D:Concurrency changes throughput but not whether labels mean the same thing. Standardize task definitions.
Question 243
An offline batch includes several files with confidential content. What design rule still applies despite asynchronous execution?
- Save every output in a publicly readable log to simplify audits
- Apply ordinary privacy, access and retention controls
- Embed private credentials into every request payload for convenience
- Assume delayed execution makes sensitive information public
Explanation
Correct Answer
B:Batch queues are not exempt from privacy policy. Only necessary content should be submitted and results should be protected.
Incorrect Answers
A:Broad log access can expose personal data. Audits need controlled records, not public disclosure.
C:Secrets in prompts expand exposure and do not replace secure authentication. Credential management belongs in the runtime.
D:Timing does not change ownership or confidentiality. Security requirements remain in force.
Question 244
A batch result can contain a per-item failure even when the overall batch has finished. What should the pipeline conclude?
- A finished batch automatically confirms extracted facts are true
- The application may delete its request mapping once completion is reported
- Every finished batch contains only successful model outputs
- Batch completion means processing ended, not that every request succeeded
Explanation
Correct Answer
D:The consumer must inspect each result type and apply retry or escalation policy to failures. A terminal batch status is not a quality certificate.
Incorrect Answers
A:Successful generation does not prove semantic correctness. Downstream evidence checks still apply.
B:The mapping remains necessary for source association and audit. It should be retained per policy.
C:Per-request validation and execution errors remain possible. They need explicit handling.
Question 245
A team wants to compare throughput and quality between direct calls and batches for a nightly job. What measurement would be fair?
- Assume one mode is more accurate solely because it is asynchronous
- Compare only how quickly each API accepts the initial request
- Compare valid results, latency and cost on matched inputs
- Ignore failed requests when computing throughput
Explanation
Correct Answer
C:Batch suitability depends on workload timing and quality, not just queue acceptance. Use the same dataset and criteria for both approaches.
Incorrect Answers
A:Scheduling architecture does not automatically improve reasoning accuracy. Quality needs empirical evaluation.
B:Submission latency does not equal time until usable results. It neglects retrieval and validation.
D:Counting only successes without the failure denominator can hide an unreliable pipeline. The comparison should include recovery work.
Question 246
A research agent writes a final policy summary and the team needs an independent check for unsupported claims. What workflow is appropriate?
- Have a review pass check claims against source evidence
- Ask the original writer to declare its answer flawless
- Remove citations to make the document easier to read
- Publish the first draft because multiple sentences sound authoritative
Explanation
Correct Answer
A:Independent evaluation can catch assumptions the writer overlooks. Findings should identify claims, sources and necessary revisions.
Incorrect Answers
B:Self-approval is not independent verification. The review should inspect actual evidence.
C:Provenance is necessary to evaluate claims. Simpler formatting should not erase support.
D:Fluent language does not prove factual support. The content must be checked.
Question 247
A code assistant finishes a patch; reviewers worry that an agent reviewing its own output may miss defects. Which strategy helps?
- Show the reviewer only the author model's praise
- Rely exclusively on the patch author's confidence statement
- Skip review whenever the diff is short
- Use a separate review context with the diff, requirements and tests
Explanation
Correct Answer
D:Independent context reduces anchoring on the original author's assumptions. It still needs executable verification where possible.
Incorrect Answers
A:Positive self-commentary biases review and omits the changed code. The reviewer needs the artifact and criteria.
B:The generating model can share blind spots with its review. Confidence alone is not a quality gate.
C:Small changes can have large effects. Risk and acceptance criteria matter more than line count alone.
Question 248
A writing pipeline has an evaluator that returns actionable defects to the generator. Which design prevents endless revision?
- Set acceptance criteria and cap evaluator feedback rounds
- Remove validation once the second draft is produced
- Allow indefinite revisions until the model feels satisfied
- Force the evaluator always to find at least one defect
Explanation
Correct Answer
A:A bounded evaluator–optimizer loop can improve output and stop on pass or escalation. It should not chase stylistic perfection forever.
Incorrect Answers
B:A fixed iteration count alone cannot establish correctness. The last candidate still requires checking.
C:Subjective satisfaction does not establish a stopping rule. The loop can waste resources or regress.
D:Mandatory criticism prevents termination even on satisfactory work. The evaluator needs a legitimate pass outcome.
Question 249
A regulator requires a report's numbers to match source spreadsheets. Which multi-pass division is strongest?
- Ask the writer to add more explanatory adjectives to each number
- Treat a well-formatted table as evidence that totals are accurate
- Draft first, then verify numerical claims against records
- Let the reviewer rely only on the report without source data
Explanation
Correct Answer
C:A specialized check can recompute figures and identify mismatches. Prose quality and numerical correctness are different evaluation dimensions.
Incorrect Answers
A:Additional prose does not verify arithmetic. The validator needs access to source values.
B:Formatting can be correct while data is wrong. Numeric assertions require calculation.
D:A reviewer cannot confirm agreement with records it cannot inspect. The check needs the authoritative inputs.
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.