Question 130
A customer wants confidence in a financial answer without receiving private internal reasoning. Which output contract is most useful?
- Return the result, evidence, assumptions, and checkable calculations
- Remove all supporting calculations to keep the answer concise
- Return only an unexplained confidence percentage
- Require every hidden internal token to prove accuracy
Explanation
Correct Answer
A:Verification can rely on observable support rather than a private reasoning transcript. Clear assumptions and reproducible calculations make the answer assessable.
Incorrect Answers
B:Concision should not eliminate the material needed to check a financial conclusion. The contract can be compact while preserving the calculation basis.
C:A percentage alone does not show whether inputs or calculations are valid. It provides less verifiable support than evidence and explicit assumptions.
D:A reasoning transcript is not necessary or sufficient for correctness. The customer needs evidence that can be checked against the task and source data.
Question 131
A product catalog changes daily, while the assistant's task instructions are stable. Where should current prices primarily come from?
- The model's general knowledge about similar products
- A monthly export whose refresh schedule is slower than required
- A maintained data source retrieved for the request
- A hardcoded price list inside the permanent role prompt
Explanation
Correct Answer
C:Frequently changing facts need an updateable evidence path. Stable instructions should define how to use the facts rather than embed an ever-growing stale catalog.
Incorrect Answers
A:General knowledge cannot establish current internal prices. Similar products may differ in timing, region, discounts, or commercial terms.
B:An export can be a useful source when its freshness meets the task. Here a slower refresh cannot support the required current pricing information.
D:Embedding volatile facts in a rarely updated prompt creates a maintenance and freshness problem. The catalog should have its own controlled data lifecycle.
Question 132
A team can fit an entire archive into context, but most questions concern a few current records. Which factor should drive the choice between full-context and retrieval approaches?
- Always use the largest possible input because it fits
- Choose solely by the number of files in the archive
- Always retrieve exactly one passage regardless of question
- Measured answer quality, evidence coverage, latency, and cost
Explanation
Correct Answer
D:Capacity alone does not establish usefulness. Compare representative tasks to see whether selecting relevant evidence or supplying the archive better meets the complete requirements.
Incorrect Answers
A:Fitting is a technical limit, not a quality guarantee. Irrelevant material can add processing cost and distract from the necessary evidence.
B:File count does not capture content size, relevance, or query structure. The decision needs workload-level evidence rather than a superficial inventory measure.
C:Some questions require several complementary records. A fixed one-passage rule can under-supply evidence even when retrieval is the right general approach.
Question 133
A worker reviewing one contract receives confidential notes from unrelated clients because the coordinator forwards all history. What context change is appropriate?
- Remove all contract evidence and retain only instructions
- Ask the worker to forget unrelated notes after reading them
- Keep all history because more context is always beneficial
- Provide only task-relevant authorized context to that worker
Explanation
Correct Answer
D:Delegation does not justify sharing every available record. A scoped handoff reduces irrelevant material and preserves the authorization boundary of the assigned task.
Incorrect Answers
A:Isolation must preserve the information needed to perform the assigned review. Eliminating relevant evidence prevents useful analysis.
B:A later instruction does not undo the disclosure. The coordinator should prevent unnecessary information from entering the worker's context.
C:Additional context can create confidentiality and relevance problems. The worker's needs and permissions should govern what information is supplied.
Question 134
A classification prompt contains correct examples, but every positive example includes a greeting and every negative example omits one. The model starts classifying by greeting presence. What revision targets this shortcut?
- Remove greetings from every future user message before classification
- Reverse all example labels while retaining their text
- Vary greetings independently of labels while retaining valid examples
- Add more positive examples that also contain greetings
Explanation
Correct Answer
C:The demonstration set accidentally associates an irrelevant surface feature with the answer. Breaking that association helps focus the examples on the actual classification criterion.
Incorrect Answers
A:That changes legitimate input and may be infeasible outside the prompt. The immediate prompt defect is the unbalanced association in its demonstrations.
B:The current labels are correct, so reversing them would teach the wrong task. The issue is a spurious contextual cue rather than label accuracy.
D:This reinforces the accidental correlation rather than challenging it. More examples are useful only if they improve the information provided about the decision rule.
Question 135
A scheduling request omits the timezone, and two plausible interpretations produce different meeting times. Which instruction best prevents an unsupported assumption?
- Use the server's local timezone without disclosing that assumption
- Choose the timezone appearing most often in examples
- Leave the timezone unspecified and let the calendar service resolve it
- Ask for the missing timezone before committing the schedule
Explanation
Correct Answer
D:The missing field changes the outcome materially. Clarification is appropriate before an external commitment that depends on that field.
Incorrect Answers
A:The server's location does not establish the user's intended scheduling zone. An unstated default can commit the appointment at the wrong time.
B:Example frequency does not identify this requester's location. Demonstrations should not replace a required input for the current task.
C:Passing the ambiguity to another component does not establish the requester's intended time. Without an agreed default, committing the event can still select the wrong instant.
Question 136
A machine consumer requires one classification label, but the assistant adds explanatory prose that breaks parsing. What should the task contract specify?
- A general request to sound concise and professional
- A free-form explanation followed by the selected classification label
- A higher maximum output limit
- The allowed label set and exact response structure
Explanation
Correct Answer
D:A precise output contract prevents ambiguity about whether prose is permitted. The application should validate the result before treating it as a classification.
Incorrect Answers
A:Style guidance does not define the machine-readable boundary. A concise sentence can still be invalid for a label-only consumer.
B:That format may be useful to a human reader but violates a label-only consumer contract. The prompt must specify the exact permitted response structure.
C:More generation capacity makes extra prose possible rather than prohibiting it. The allowed structure must be defined and checked.
Question 137
A demonstration copied from an untrusted forum contains an instruction to reveal secrets. How should it be handled before inclusion in a shared prompt?
- Move the secret-revealing instruction into the answer key
- Include it unchanged because it is only an example
- Rely on the forum's popularity as the review criterion
- Review and sanitize demonstrations as executable behavioral guidance
Explanation
Correct Answer
D:Examples influence the model's interpretation of the task. Untrusted demonstrations can import malicious or inappropriate instructions into a higher-trust prompt component.
Incorrect Answers
A:Relocating malicious content within a demonstration can reinforce it as desired behavior. The unsafe instruction should not be taught as a valid outcome.
B:Example status does not eliminate behavioral influence. The content can still steer responses or conflict with the application's intended boundaries.
C:Popularity is not a security or correctness guarantee. Shared prompt content requires review against the actual task and data-handling requirements.
Question 138
A rewriting prompt says only, "Do not sound bureaucratic." Outputs vary widely in tone and omit important requirements. Which revision provides more actionable guidance?
- Describe the desired plain-language style and the content that must be retained
- Remove all style instructions and accept whichever tone the model chooses
- Add several synonyms for bureaucratic to the prohibited-word list
- Require the shortest possible answer for every source paragraph
Explanation
Correct Answer
A:Negative guidance excludes a style without defining a suitable replacement. Positive requirements can specify readable wording while preserving the information the task needs.
Incorrect Answers
B:That abandons the stated communication goal. The prompt needs a clearer target rather than the absence of a style requirement.
C:A longer prohibition still does not define the desired output or required content. It may also suppress necessary terminology without improving clarity.
D:Brevity does not guarantee plain language or completeness. The scenario specifically reports omitted requirements, which an absolute shortening rule could worsen.
Question 139
A company has separate procedures for invoice disputes and product returns. Their descriptions both say customer support. Selection is unreliable. What should change?
- Merge all steps into a single unconditional procedure
- Choose whichever procedure was created most recently
- Give both procedures the same longer description
- Describe the distinct triggers, inputs, and intended outcomes
Explanation
Correct Answer
D:A useful capability description communicates when to apply it. Discriminating conditions reduce confusion between procedures that share a broad business category.
Incorrect Answers
A:Combining incompatible workflows can cause irrelevant actions. Shared category membership does not make their triggers and decision paths identical.
B:Creation time does not establish relevance to the current request. Selection should follow the task's actual inputs and desired outcome.
C:Additional identical text does not distinguish applicability. It preserves the ambiguity that caused incorrect selection.
Question 140
A reusable procedure explains how to reconcile invoices but embeds a supplier bank account that changes occasionally. What separation is best?
- Keep the procedure reusable and resolve bank details from approved live data
- Copy the account into more examples to reinforce it
- Treat the embedded account as permanent because the procedure is approved
- Reuse the bank details embedded in the last successful execution
Explanation
Correct Answer
A:Procedural knowledge and mutable operational facts have different lifecycles. The workflow should retrieve the current authorized account rather than trust an old embedded value.
Incorrect Answers
B:Repetition does not improve freshness or authority. It makes the stale detail more influential while preserving the underlying update problem.
C:Approval of a procedure does not make every included fact timeless. Mutable bank details need their own verification and update path.
D:A previous successful payment does not establish current account details. The procedure should resolve authoritative live values for the new transaction.