Question 290
A customer service assistant must always escalate potential fraud accusations before taking adverse action. Which mechanism is best?
- Hide the source evidence from the reviewer to avoid bias
- A mandatory review gate triggered by explicit risk criteria
- Allow the assistant to bypass review for familiar customers
- Treat the model's fraud label as an automatic ban
Explanation
Correct Answer
B:A human checkpoint can assess evidence and consequences before account action. The agent can gather facts but not make irreversible accusations alone.
Incorrect Answers
A:Reviewers need relevant facts to assess the claim. Data should be scoped and presented with provenance.
C:Familiarity is not a documented exception and can create inconsistent treatment. The gate should apply according to policy.
D:An unverified classifier can produce false positives. Human review is the stated requirement.
Question 291
A security finding from an AI code reviewer lacks a reproducible exploit path. How should a human triage queue present it?
- Ask another model to repeat the same assertion without additional context
- Discard every unverified finding automatically
- Flag it as unverified, include affected code and request a targeted validation
- Publish the issue as confirmed critical without evidence
Explanation
Correct Answer
C:A clear uncertainty label avoids overstating risk while preserving a potentially valuable lead. Reproduction can change its status.
Incorrect Answers
A:Repetition is not independent evidence. The queue needs a reproducible test or stronger source.
B:Preliminary signals can still warrant investigation. Verification, not silent deletion, is appropriate.
D:Severity claims need supporting analysis. Overstating unverified findings harms prioritization.
Question 292
A release process uses two AI reviewers that agree a patch is safe, but mandatory integration tests are red. What should the release gate do?
- Block release until tests pass or an approved exception exists
- Hide the failing tests to avoid contradictory signals
- Approve because two models agreed
- Randomly pick one reviewer to make the final technical decision
Explanation
Correct Answer
A:Model consensus does not override deterministic acceptance criteria. Failed integration checks are direct evidence of a problem.
Incorrect Answers
B:Suppressing negative evidence defeats release assurance. The failure needs diagnosis.
C:Correlated model errors or missed edge cases remain possible. Measured tests are stronger for the specified gate.
D:The governing rule is the test and approval policy, not an arbitrary selection. Accountability must be explicit.
Question 293
A compliance team wants to determine whether its AI reviewer catches prohibited statements consistently. What evaluation method is best?
- Evaluate only the model reports that reviewers already approved
- Evaluate labeled positives, negatives and borderline cases
- Ask the model whether it is usually accurate
- Treat every disagreement with a human as proof the human is wrong
- Count only the number of warnings produced
Explanation
Correct Answer
B:Measured precision and recall reveal both missed risks and excessive flags. Results should guide prompt and routing adjustments.
Incorrect Answers
A:Excluding difficult and false-positive cases biases the results. Representative labeled examples are necessary for calibration.
C:Self-evaluation does not establish actual performance. Representative examples and outcomes are necessary.
D:Reference judgments may need adjudication, but automatic model superiority is unsupported. Calibration requires evidence.
E:High warning volume can reflect false positives rather than better detection. Labels are needed to assess quality.
Question 294
A research agent cites three articles, but all three repeat one press release. How should its synthesis reflect evidentiary independence?
- Remove source links and state the conclusion as proven
- Identify the shared origin rather than counting repeated claims
- Treat each website as a completely separate primary source
- Invent a fourth citation to increase apparent consensus
Explanation
Correct Answer
B:Source provenance matters more than citation count. Correlated reports do not establish multiple independent observations.
Incorrect Answers
A:Hiding citations weakens verification. The report should explain evidence limits.
C:Different hosts can repeat identical upstream material. Independence must be assessed from origin, not URL count.
D:Fabricated evidence destroys trust and does not improve the claim's basis.
Question 295
A report generator knows the numeric claim came from one dataset but does not know the date of collection. How should it present the figure?
- Remove the dataset reference so nobody asks about time
- Invent a recent collection date that seems plausible
- Present it as a current confirmed statistic
- Attribute the dataset while qualifying uncertain freshness
Explanation
Correct Answer
D:A figure without a known observation period may not support a current claim. Clear uncertainty allows readers to judge applicability.
Incorrect Answers
A:Suppressing provenance conceals the very uncertainty that needs attention. The reference should remain available.
B:A made-up date is false metadata. Missing information should be stated honestly.
C:Unverified recency can mislead decisions. The source date is a material limitation.
Question 296
A published research brief compares public statistics that differ by collection year and derives an uncertain trend. Which TWO practices make the synthesis defensible? Choose TWO.
- Invent the most recent dates where source metadata is missing
- Identify the source, revision and observation period behind each statistic
- Suppress citations so the report feels more coherent
- Average incompatible periods without describing the mismatch
- Clearly separate measured findings from inferred trends and unknowns
Explanation
Correct Answers
B:Provenance exposes whether apparent disagreement reflects different years or methods. Readers can trace each claim to evidence.
E:Data can support a hypothesis without proving it. Explicit qualification prevents speculation from being presented as verified fact.
Incorrect Answers
A:Fabricated recency is false provenance. Unknown dates should be disclosed rather than guessed.
C:Removing attribution hides the source-quality problem. Coherence does not justify unverifiable assertions.
D:Combining incomparable data can produce misleading figures. The analysis must consider scope and methodology.
Question 297
A policy summary includes a quoted requirement but cites only an unrelated overview page. What should the checker flag?
- Replace the actual policy quote with an invented authoritative paraphrase
- Treat any accessible URL as proof of every nearby claim
- A source-to-claim mismatch even if the citation URL is valid
- Ignore the quoted requirement because citations are only decorative
Explanation
Correct Answer
C:Citation existence is insufficient. The linked material must actually support the specific requirement and its scope.
Incorrect Answers
A:Fabrication is not a remedy. The author should locate the proper supporting source.
B:A working link says nothing about evidentiary relevance. The checker must inspect the passage.
D:In compliance summaries, accurate attribution is substantive. Mismatched support needs correction.
Question 298
A research report must distinguish measured results, model inferences and speculation. What output convention helps?
- Suppress all uncertainty markers to improve readability
- Write every sentence with equal certainty
- Claim the model personally verified every cited source whether or not it did
- Label claim types and link each to evidence or assumptions
Explanation
Correct Answer
D:Readers can then assess confidence and intended use. A clear distinction prevents speculation from becoming an asserted fact.
Incorrect Answers
A:Removing qualifiers can distort the substantive conclusion. Clarity should not come at the expense of honesty.
B:Uniform confidence masks differences in evidence quality. It can exaggerate weak hypotheses.
C:Verification must reflect actual work performed. Unsupported review claims undermine provenance.
Question 299
A document summarizer quotes text from an old revision of a contract after a newer revision supersedes it. What should the source policy require?
- Treat any old contract text as permanently current
- Choose whichever version contains a shorter clause
- Track revisions and use the currently governing source
- Combine contradictory revisions into one invented clause
Explanation
Correct Answer
C:Version-aware synthesis avoids treating obsolete clauses as active obligations. The report can mention historical differences when relevant.
Incorrect Answers
A:Legal obligations can change through amendments. Source age and authority must be checked.
B:Brevity does not determine precedence. The governing document and effective date matter.
D:Merging incompatible text falsifies the source. Differences should be attributed explicitly.
Question 300
A research agent cannot find a primary source for an important assertion but has several secondary summaries. How should its final report handle the gap?
- Describe the claim as independently verified at the source
- Qualify secondary evidence and request primary verification
- Cite a made-up primary document name
- Discard the entire report without mentioning which claim is unsupported
Explanation
Correct Answer
B:The absence of primary support affects confidence. The user can decide whether secondary reports suffice for the intended decision.
Incorrect Answers
A:That would overstate the investigation. A secondary summary is not direct examination of the primary record.
C:Inventing provenance is unacceptable. Unknown source details should remain unknown.
D:A targeted limitation is more useful than losing all verified findings. The gap should be explicit and scoped.