Question 200
A code analysis model flags comments containing a deprecated API name even though no code calls it. What instruction is most effective?
- Count every textual match as runtime usage
- Ask the model to provide longer warnings for every match
- Require execution evidence, not mere mentions in prose
- Ignore all deprecated APIs to eliminate warnings
Explanation
Correct Answer
C:A precise finding needs the relevant syntax and call location. Mentions in documentation do not establish active dependency.
Incorrect Answers
A:Raw search results are candidates for inspection, not proof of execution. The context must be classified.
B:More words do not correct an invalid trigger condition. The underlying rule needs refinement.
D:That suppresses genuine issues. The requirement is improved specificity, not zero detection.
Question 201
A document review process must flag invoices missing purchase-order identifiers but only when the vendor contract requires one. How should the task be specified?
- State the conditional rule and proof needed for omission
- Infer the rule from the invoice's visual design
- Skip all missing-field checks whenever a contract is long
- Treat every absent identifier as a breach
Explanation
Correct Answer
A:Conditional criteria prevent labeling legitimate exceptions as defects. The model should check the applicable agreement before asserting noncompliance.
Incorrect Answers
B:Layout is not a contractual requirement. The conclusion must be tied to a trusted rule.
C:Length does not determine whether a requirement applies. Efficient retrieval should locate the relevant clause.
D:Universal assumptions ignore contract differences. The review needs the governing policy context.
Question 202
An AI reviewer correctly finds defects but reports vague critiques without file references. What acceptance criteria should be added?
- Require location, impact, evidence and reproducible fix
- Ask for more dramatic language to emphasize severity
- Remove the requirement to cite changed lines
- Score reviews only by number of paragraphs
Explanation
Correct Answer
A:Structured findings make review actionable and independently checkable. Unsupported general comments should not count as defects.
Incorrect Answers
B:Emphasis cannot replace evidence. It may overstate risks and reduce trust.
C:Omitting locations makes reproduction harder. The complaint is already about insufficient grounding.
D:Length does not measure usefulness. The deliverable requires specified evidence fields.
Question 203
A policy classifier labels any angry customer as a fraud risk without transactional evidence. What change best controls that bias?
- Treat negative sentiment as conclusive proof of suspicious activity
- Use verified transaction signals, not customer tone
- Ask for a fraud verdict before retrieving account facts
- Increase the classification threshold only by arbitrary token length
Explanation
Correct Answer
B:Emotional style is not sufficient evidence of abuse. The classifier needs observable criteria and explicit noncriteria.
Incorrect Answers
A:Sentiment and transaction misconduct are different constructs. The rule must avoid conflating them.
C:A premature judgment can anchor the model on incomplete evidence. Verification should precede classification.
D:Message length is not a fraud signal. This may create different but still unjustified decisions.
Question 204
A compliance reviewer must identify one of three specific violations, not write a general quality essay. What prompt design is best?
- Define each category and an explicit none-found result
- Request an unrestricted critique with no category definitions
- Ask the model to choose a label at random when uncertain
- Force at least one violation to be reported for every artifact
Explanation
Correct Answer
A:A bounded taxonomy aligns output with the audit task and handles clean cases. Evidence for each category reduces speculative flags.
Incorrect Answers
B:Open-ended feedback can introduce irrelevant issues. It does not meet the targeted violation taxonomy.
C:Random classification is not reliable judgment. Uncertain cases should be marked for clarification or review.
D:Mandatory positive findings create false positives on compliant material. A no-violation path is necessary.
Question 205
A data quality auditor needs to distinguish unsupported assertions from claims that merely lack a citation in the final text. What criteria should it use?
- Classify any cited sentence as verified fact
- Score only the stylistic elegance of the paragraph
- Check evidence, citations and claim truth independently
- Classify every uncited sentence as false
Explanation
Correct Answer
C:A statement can be true but uncited, or cited to a source that does not support it. Independent criteria prevent conflated labels.
Incorrect Answers
A:A citation can be irrelevant or misinterpreted. The actual claim-source relation must be inspected.
B:Style does not establish factual support or provenance. The audit needs evidence-focused rules.
D:Missing attribution does not by itself prove factual incorrectness. It is a separate evidence-quality issue.
Question 206
A document extractor returns addresses in inconsistent formats despite clear written instructions. Which few-shot examples would help most?
- Repeat the same trivial address ten times
- Pair realistic inputs with normalized outputs, including edge cases
- Remove all examples because any sample necessarily causes memorization
- Provide only an unrelated sample of a product review
Explanation
Correct Answer
B:Examples demonstrate expected formatting and decision boundaries. Diverse cases teach the pattern more effectively than repeated easy instances.
Incorrect Answers
A:Redundant examples teach little about edge cases. Variety in realistic formats better exposes the rule.
C:Well-chosen examples can improve consistency. The issue is quality and diversity, not the mere presence of examples.
D:Irrelevant examples do not illustrate address normalization. They may distract the model from the required structure.
Question 207
A support classifier has examples only of successful refunds and falsely approves unusual returns. What should be added?
- A rule to always approve whenever an example resembles a request
- Examples that are all unrelated general customer greetings
- More duplicates of straightforward approved refunds
- Include borderline denial and escalation examples
Explanation
Correct Answer
D:Contrastive cases teach what falls outside eligibility. This helps distinguish normal exceptions from eligible requests.
Incorrect Answers
A:Similarity alone is not policy compliance. The assistant must evaluate the actual eligibility conditions.
B:Greetings do not define refund decision criteria. Samples must represent the target task.
C:More positives may reinforce the current bias without explaining the rejection boundary.
Question 208
An API extraction prompt includes an example whose output violates the declared JSON schema. What should the team fix?
- Remove schema validation so the incorrect example is accepted
- Add unrelated examples that use different output formats
- Use the runtime schema consistently in every example
- Tell the model to ignore the example while leaving it visible
Explanation
Correct Answer
C:Contradictory examples undermine the format contract. The model should see valid keys, types and optional-field behavior.
Incorrect Answers
A:Weakening the contract hides the discrepancy. The example should be corrected.
B:Multiple incompatible shapes increase ambiguity. All instructional examples should reinforce one contract.
D:Conflicting prompt material remains a potential influence. Removing or fixing the example is clearer.
Question 209
A purchase-order extractor misses cases where the order ID appears in the footer. What example strategy addresses the failure?
- Use only samples with the ID in the first line
- Drop order IDs from the required fields
- Increase final answer verbosity without changing examples
- Demonstrate extraction when the identifier is in a footer
Explanation
Correct Answer
D:A targeted edge case teaches the model not to rely solely on headers. Validation should confirm it still handles normal layouts.
Incorrect Answers
A:Header-only examples reinforce the existing blind spot. They do not illustrate the footer variant.
B:That changes the business requirement rather than solving the extraction problem.
C:More text does not change where evidence is found. The input pattern needs coverage.
Question 210
A contract reviewer must distinguish obligations from optional recommendations. Which example selection is strongest?
- Contrast mandatory and advisory terms in near-match clauses
- Mix contradictory labels for similar obligations
- Use only long samples without labels
- Provide samples from an unrelated image captioning task
Explanation
Correct Answer
A:Contrastive examples highlight the linguistic distinction that changes classification. Labels should explain why each clause belongs in its category.
Incorrect Answers
B:Inconsistent demonstrations create ambiguity and unreliable classifications. The labels must follow the defined criterion.
C:Unlabeled text does not show the desired decision. The model needs example outputs and category meanings.
D:Different modalities and goals do not teach contract obligation semantics. Examples should mirror the use case.