Question 300
A retrieved policy points to a separate eligibility table needed for the answer. What should the agent do?
- Assume the table matches common practice.
- Retrieve the referenced table before drawing the dependent conclusion.
- Ignore the reference and answer definitively.
- Treat the reference label as the full table.
Explanation
Correct Answer
B:Retrieve the referenced evidence or identify the unresolved dependency; the first document's pointer does not itself supply the eligibility values.
Incorrect Answers
A:General expectations cannot establish this specific referenced rule.
C:That omits material evidence explicitly required by the source.
D:A pointer identifies content but does not contain its details.
Question 301
Search returns the same announcement mirrored on five pages. How should evidence aggregation treat them?
- Count each mirror as an independent confirmation because it has a different hostname.
- Assume five URLs guarantee five independent sources.
- Recognize their shared origin rather than count them as five independent confirmations.
- Choose the copy with the most advertisements.
Explanation
Correct Answer
C:Mirrors share a common origin, so count them as corroborating copies rather than independent verification of the underlying announcement.
Incorrect Answers
A:Different locations can reproduce one original claim, so URL diversity does not establish independent evidence.
B:Different locations can reproduce the same underlying statement.
D:Advertising volume does not improve source independence or authority.
Question 302
A document has the highest semantic similarity score. What does that score establish?
- That every statement in the document is verified.
- That the caller is authorized to access it.
- Relevance under the retrieval method, not guaranteed factual authority.
- That no other evidence should ever be checked.
Explanation
Correct Answer
C:A close textual match can still be outdated, untrusted, or factually wrong. Similarity measures relevance under the retrieval method, not truth, authority, recency, or the caller's access rights.
Incorrect Answers
A:Similarity is not a truth certification.
B:Ranking does not replace access control.
D:High ranking does not eliminate the need to assess completeness or conflicts.
Question 303
A source says a feature is not supported for archived records. A summary removes “not.” What kind of error results?
- An authorized relaxation of the product limitation.
- A material semantic reversal of the restriction.
- Only a harmless punctuation change.
- A guaranteed improvement in readability and accuracy.
Explanation
Correct Answer
B:Compaction must preserve negation and conditions; a shorter summary that changes the rule is not faithful context reduction.
Incorrect Answers
A:Summarization does not grant authority to change the underlying rule.
C:Negation changes meaning, not merely style.
D:Readability cannot compensate for reversing the source’s meaning.
Question 304
A user corrects a previously stored shipping preference. What should memory management do?
- Ignore corrections because the first note is permanent.
- Apply the correction to every user globally.
- Append the correction without resolving which value is current.
- Update or supersede the old preference with clear provenance.
Explanation
Correct Answer
D:Update or supersede the stale preference with provenance, ensuring subsequent retrieval does not keep presenting the old value as current.
Incorrect Answers
A:Stored context is not immutable truth.
B:A personal update does not authorize changing unrelated users’ records.
C:Unresolved contradictions can cause later tasks to use the obsolete preference.
Question 305
Retrieved passages are reranked before generation. What must citation metadata preserve?
- Only the highest-ranked document’s name for all passages.
- The association between each passage and its original source identifier.
- Regenerate citation identifiers from the new list positions without updating references.
- The original list positions as if order never changed.
Explanation
Correct Answer
B:Bind citations to stable source identities and passage metadata, not positions that change when the retrieval list is reordered.
Incorrect Answers
A:That falsely attributes content from other sources to one document.
C:Position-based renumbering can disconnect generated citations from their original source passages after reranking.
D:Position-based references can become incorrect after reranking.
Question 306
An agent summarizes two groups for later combined-average calculations. It saves each group's mean but discards the group sizes. What must the summary preserve?
- The count and total, or count and mean, for each group.
- The most recent group's mean as a replacement for both groups.
- Only the average of the two group means.
- The number of groups, without their individual sample counts.
Explanation
Correct Answer
A:A combined mean weights each group by its size. Without the counts, a later stage cannot distinguish a mean based on two records from one based on thousands.
Incorrect Answers
B:Recency does not make one group's statistics representative of the combined population requested by the later task.
C:An unweighted average is correct only when the groups have equal sizes. That condition is not established and cannot be checked after discarding counts.
D:Knowing there are two groups does not provide their relative weights. Unequal group sizes require counts to combine means correctly.
Question 307
A user requests a comparison of three named policies. All top 20 retrieved chunks come from the first policy, although the other two are indexed. How should context assembly address this?
- Treat the number of retrieved chunks as each policy's relative importance.
- Ask the model to infer the two missing policies from the retrieved one.
- Increase the number of chunks from only the first policy.
- Retrieve relevant evidence for each required policy before comparing them.
Explanation
Correct Answer
D:The top chunks may all be relevant yet fail the task's coverage requirement. Source-aware retrieval or separate queries can supply evidence for the missing comparison dimensions.
Incorrect Answers
A:Chunk count reflects retrieval concentration and document segmentation, not importance under the user's comparison request.
B:One policy does not supply evidence for the others. A comparison of named documents needs support from each relevant source.
C:More evidence from the already represented source does not fill the gaps for the other two required policies.
Question 308
A prompt’s rule labels overdue accounts as urgent, but several examples label them routine. What should be corrected first?
- Add more contradictory examples.
- Remove the evaluation cases that reveal confusion.
- Increase the answer length.
- Resolve the conflict between the rule and demonstrations.
Explanation
Correct Answer
D:Contradictory supervision makes the desired mapping unclear; consistent examples should reinforce the intended rule.
Incorrect Answers
A:Volume does not resolve inconsistent labels.
B:Hiding failures does not repair the prompt.
C:More prose does not clarify which instruction should govern classification.
Question 309
A prompt says “give a useful result,” but consumers disagree on required fields. What should be specified?
- Only an example with no statement of required versus optional fields.
- Only a larger context allowance.
- The required information, output structure, and missing-data behavior.
- A promise that consumers will infer every field.
Explanation
Correct Answer
C:Define required fields, permitted values, and missing-data behavior so both generation and downstream validation share a concrete success contract.
Incorrect Answers
A:An isolated sample can leave consumers unsure whether absent fields are allowed or represent generation failures.
B:Capacity does not supply an output contract.
D:Different consumers may interpret free text inconsistently.
Question 310
A task requests exhaustive detail and a strict fifty-word limit. What should prompt design clarify?
- Add another equally conflicting length instruction.
- Assume the model can fit unlimited detail into any length.
- Repeat both requirements in uppercase.
- Which constraint takes priority and what information must be preserved.
Explanation
Correct Answer
D:Explain which constraint takes priority and how to summarize when both cannot be satisfied, preventing incompatible instructions from remaining unresolved.
Incorrect Answers
A:Additional contradiction makes compliance less determinate.
B:A finite output budget imposes real information tradeoffs.
C:Emphasis does not resolve their practical conflict.