Question 290
A system hits its budget cap even though request rate is low. What distinction matters?
- A low request count guarantees low token cost.
- Treat the budget failure as a network timeout and retry indefinitely.
- Every limit is solved by millisecond jitter.
- Cumulative spending limits differ from instantaneous throughput limits.
Explanation
Correct Answer
D:Low traffic does not undo accumulated spend; recovery must address the actual budget condition. A few large requests can consume a financial allowance without approaching a requests-per-minute limit, so diagnose which quota was actually reached.
Incorrect Answers
A:Large or expensive requests can consume substantial budget.
B:A financial allowance is not restored by network retry, and repeated attempts can worsen cost or remain consistently blocked.
C:Delay variation cannot reset an exhausted spending allowance.
Question 291
Several products share model infrastructure. What supports meaningful cost attribution?
- Count only successful final responses.
- Guess product shares from team size.
- Associate usage with application, task, and configuration identifiers.
- Store only one global total indefinitely.
Explanation
Correct Answer
C:Attribution connects resource consumption to the work responsible for it and supports targeted optimization. Tag usage with the responsible product or job and include retries and failed attempts where billed, preserving attribution through shared infrastructure.
Incorrect Answers
A:Failed attempts and intermediate calls can also incur cost.
B:Headcount is not measured model usage.
D:A total cannot identify which workload or release caused increased spending.
Question 292
A tool list’s order changes on every otherwise identical request. Why can that undermine prefix caching?
- The serialized prefix may no longer match exactly.
- Unused tools execute automatically when reordered.
- Reordering definitions guarantees identical serialized request prefixes.
- Caching requires every answer to be identical.
Explanation
Correct Answer
A:Stable content representation matters for exact prefix reuse; unnecessary ordering changes can invalidate an otherwise reusable prefix.
Incorrect Answers
B:Reordering definitions does not itself invoke them.
C:Order changes can alter the prefix representation used for caching even when the set of tools is semantically unchanged.
D:Prompt-prefix caching is not simply storing one fixed generated answer.
Question 293
A job must never exceed a fixed budget. Every operation has a known enforced maximum charge, several operations may run concurrently, and no other charges occur. Which admission design enforces the cap?
- Atomically reserve each operation's maximum charge before admitting it.
- Let each worker independently read the balance before starting.
- Admit operations whenever the model estimates that they will be cheap.
- Subtract actual charges only after each operation has finished.
Explanation
Correct Answer
A:Admission succeeds only if the reservation fits the remaining budget. Reserving before concurrent work begins prevents double-spending the balance; reconcile actual charges and release unused reservations afterward.
Incorrect Answers
B:Without an atomic reservation, multiple workers can each see the same available balance and collectively commit more than remains.
C:A model estimate is not the enforced maximum charge in this scenario. Admission based on that estimate cannot establish the promised hard expenditure bound.
D:Post-completion accounting records expenditure but cannot prevent already-admitted concurrent operations from exceeding the balance before their charges arrive.
Question 294
A question concerns one policy clause, but retrieval supplies hundreds of unrelated pages. What context change is most useful?
- Select the relevant passages while preserving necessary surrounding definitions.
- Remove the user’s actual question to make room.
- Always fill the entire available window.
- Keep only the longest document.
Explanation
Correct Answer
A:Select the passages needed to answer the clause-level question and retain source references, reducing distracting material without losing supporting evidence.
Incorrect Answers
B:The model still needs the task that determines relevance.
C:Capacity is an upper bound, not a requirement to add irrelevant material.
D:Length does not identify the passage needed for the answer.
Question 295
A retrieved paragraph says “these exceptions apply” but its heading and referenced rule were omitted. What should retrieval include?
- Only a neighboring passage selected by length rather than structural relevance.
- Enough neighboring structure to interpret the paragraph correctly.
- A guessed rule generated from general knowledge.
- Only the paragraph repeated several times.
Explanation
Correct Answer
B:Chunk boundaries should preserve necessary definitions and references rather than isolate grammatically complete but semantically incomplete text.
Incorrect Answers
A:The missing heading and referenced rule supply meaning; an arbitrary long passage may leave the pronoun and exception scope unresolved.
C:A guess is not evidence of the source’s intended rule.
D:Repetition does not restore the missing referent.
Question 296
A context-pruning step removes verbose tool logs. Which details should survive when later actions depend on them?
- All repeated debug lines but none of the result fields.
- Only the final prose summary, omitting all durable resource identifiers.
- Stable resource identifiers, outcomes, and unresolved action state.
- A new guessed identifier for every resource.
Explanation
Correct Answer
C:Preserve resource identifiers, outcomes, and unresolved decisions that future steps require; verbose execution logs can be removed without erasing operational state.
Incorrect Answers
A:This retains noise while deleting decision-critical information.
B:Later operations may need exact identifiers to target previously created resources; descriptive prose alone can make those targets ambiguous.
D:Invented identifiers break the link to actual external objects.
Question 297
A durable note claims a customer authorized a change. What provenance should be retained?
- The authorizing source, scope, and relevant validity conditions.
- The model’s preferred explanation of the customer’s intent.
- Only the word approved.
- An unrelated task’s approval record.
Explanation
Correct Answer
A:Retain who authorized which action, when, and under what scope or validity conditions. A vague remembered claim must not become unlimited permission for later actions.
Incorrect Answers
B:Interpretation is not a substitute for the underlying authorization evidence.
C:That omits who approved what and whether the approval remains applicable.
D:Permission for another operation does not establish authority here.
Question 298
A research subtask requires reading noisy documents unrelated to the main conversation’s final reasoning. What can context isolation achieve?
- Automatically grant the worker broader data permissions.
- Guarantee the research findings are correct.
- Make the worker’s token usage disappear.
- Keep exploratory detail separate and return only relevant supported findings.
Explanation
Correct Answer
D:Isolation reduces pollution of the main context while allowing useful evidence to be transferred deliberately.
Incorrect Answers
A:Context boundaries are not authorization grants.
B:Separate context does not eliminate factual or reasoning errors.
C:Delegated processing still consumes resources even when its transcript is not copied back.
Question 299
A retrieval query is rewritten to improve search recall. What must the rewrite preserve?
- Every unrelated topic mentioned in earlier sessions.
- Only the most common keyword.
- A more convenient question with an easier answer.
- The user’s entities, restrictions, and intended information need.
Explanation
Correct Answer
D:Keep entity, time, scope, and exclusion constraints intact; a broader rewrite can retrieve relevant-looking material for a different question.
Incorrect Answers
A:Unrelated history adds noise rather than faithful query expansion.
B:Dropping specific entities or exclusions can make results irrelevant.
C:Retrieval must serve the actual request rather than silently substitute another one.
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.