Question 230
A parser repeatedly fails on the same malformed document after two informed retries. What should the workflow do next?
- Return a controlled validation failure for human correction
- Publish a fabricated complete record to meet the deadline
- Continue retries without an upper limit
- Reset all error counters after each attempt
Explanation
Correct Answer
A:Bounded retries avoid runaway cost. Persistent failures may require better source quality or human judgment.
Incorrect Answers
B:False data creates downstream business risk. Honest failure is preferable to invented success.
C:Repeated model calls can loop indefinitely on irreparable input. The system needs a stop condition.
D:Counter reset defeats the retry policy and obscures repeated failures. Persistent state is required.
Question 231
A tool returns a corrected structured object, but the application does not rerun validation before saving. What should change?
- Validate repaired output against schema and business rules
- Skip checking source evidence after the first failure
- Assume a second model attempt always resolves the issue
- Save both candidates as one merged JSON string
Explanation
Correct Answer
A:A repair attempt is not automatically successful. The system should verify the new artifact against the original acceptance conditions.
Incorrect Answers
B:A format repair may alter values as well as shape. Source consistency still matters.
C:Retries can introduce different errors. Validation must be repeated on the actual candidate.
D:Combining incompatible records can create invalid data. One checked result should be committed.
Question 232
A model returns a date in the wrong locale despite an otherwise valid record. What targeted feedback should a validator provide?
- Replace all dates with today's date
- Ignore the issue if the month name is recognizable
- Delete the field because locale is complicated
- Name the field, observed value and expected date format
Explanation
Correct Answer
D:Field-level feedback enables a narrow correction without recomputing unrelated values. The host can also normalize dates deterministically when safe.
Incorrect Answers
A:A default date is not the original document's fact. It would silently corrupt historical records.
B:Downstream parsers may require an exact format. Human readability is not sufficient for machine ingestion.
C:The business process still needs the date. Conversion or clarification is more appropriate.
Question 233
A content moderation pipeline reports borderline cases as certain violations without supporting evidence. Which reliability loop is appropriate?
- Apply explicit policy criteria; escalate unresolved cases
- Ignore all borderline cases as unquestionably safe
- Silently block every borderline user to avoid risk
- Ask the same model to insist more strongly on its first conclusion
Explanation
Correct Answer
A:A second assessment with clear criteria can reduce overconfident false positives. High-impact ambiguities need human judgment.
Incorrect Answers
B:Uncertainty is not evidence of safety either. The workflow needs a documented review threshold.
C:A blanket policy may violate the application's defined thresholds. Escalation offers accountable handling.
D:Confidence language does not add evidence. The result should be checked against policy and context.
Question 234
An extractor repairs a missing key but unexpectedly deletes a previously correct address. How can the loop avoid this regression?
- Ask for a more verbose explanation without inspecting the record
- Make all fields optional so deletions no longer fail
- Stop checking other fields once the first error is fixed
- Revalidate the full record after each repair attempt
- Commit whichever candidate has the most fields regardless of validity
Explanation
Correct Answer
D:Local edits can introduce new defects. Whole-object validation and source checks catch collateral damage.
Incorrect Answers
A:Long explanations do not ensure data integrity. The structured output must be tested.
B:Weakening requiredness hides regressions. The schema should reflect business needs.
C:Partial validation can accept new omissions. The original contract applies to every field on every iteration.
E:Field count is not a correctness measure. Every revised candidate must satisfy the full schema and evidence checks.
Question 235
A tool result contains a transient connection error during extraction. Which retry rule makes the workflow safer?
- Retry instantly forever to minimize delay
- Convert connection errors into empty extraction records
- Treat every tool failure as a reason to repeat all writes
- Retry read-only calls with bounded backoff and error evidence
Explanation
Correct Answer
D:Temporary failures can justify retry, but the policy must respect idempotency and resource limits. Successful recovery should be verified.
Incorrect Answers
A:Unbounded hot loops can aggravate outages and exhaust quota. Backoff and stop conditions are required.
B:A blank success hides unavailable evidence. The workflow must distinguish failure from a genuinely empty document.
C:Write operations may have committed before an error response. Blind retry can duplicate side effects.
Question 236
A company must classify thousands of independent support tickets overnight rather than interactively. Which API processing pattern is suitable?
- Run every ticket in a live interactive agent session
- Use asynchronous batches with per-request reconciliation
- Assume outputs arrive in the same order as inputs without identifiers
- Wait for all tickets to be merged into one giant user prompt
Explanation
Correct Answer
B:Batch processing suits independent offline work. The application must track request identifiers and handle partial failures on retrieval.
Incorrect Answers
A:Interactive turn-by-turn control adds unnecessary overhead to independent offline classification. The workload can be queued.
C:Batch completion order may differ. Identifiers are necessary for reliable mapping.
D:A monolithic prompt can exceed context limits and blurs per-ticket validation. Independent requests are easier to reconcile.
Question 237
A batch result file contains records in a different order from the original requests. How should consumers match them?
- Join each batch result by its unique custom ID
- Discard request IDs to make files smaller
- Sort results alphabetically by answer text
- Map the first response to the first submitted ticket by position
Explanation
Correct Answer
A:Out-of-order results are normal in asynchronous processing. Stable IDs prevent attaching one customer's result to another case.
Incorrect Answers
B:The identifiers are essential operational metadata. Removing them undermines reconciliation.
C:Answer wording does not encode request identity. Sorting cannot restore the correct correspondence.
D:Position is not a safe correlation contract when work completes asynchronously. This can misassign sensitive records.
Question 238
A nightly batch of independent document extracts completes with some failed requests and results in an unexpected order. Which TWO capabilities should the consumer implement? Choose TWO.
- Assume the first output corresponds to the first submitted request
- Correlate each result to its original document using a stable request identifier
- Discard all successful items whenever a small portion fail
- Convert tool or processing failures into valid empty documents
- Track per-item status and retry or report only the failed inputs
Explanation
Correct Answers
B:Asynchronous completion order is not a reliable mapping. IDs prevent applying one document's extraction to another record.
E:A finished batch can contain partial failures. Targeted recovery preserves validated successes and makes completeness measurable.
Incorrect Answers
A:Queue execution can reorder results. Position-based mapping risks serious data mix-ups.
C:Confirmed results can be retained after validation. Global replay wastes time and may create duplicates.
D:Errors are not the same as zero-field extractions. Silent substitution hides missing work.
Question 239
A workload needs answers immediately while the customer waits on a web page. Why might batch processing be the wrong default?
- Batch processing cannot handle independent requests
- Batch processing always returns in under a second
- Batch outputs automatically become visible to browser users
- Batching favors offline throughput, not immediate replies
Explanation
Correct Answer
D:The user-facing interaction has a latency requirement. A direct request path is more appropriate for time-sensitive replies.
Incorrect Answers
A:Independence is a good batch use case. The mismatch here is response urgency, not task structure.
B:An asynchronous queue does not guarantee immediate completion. It is not a low-latency chat transport.
C:Application code must retrieve and deliver results. Queue completion does not supply a live web response by itself.
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.