Question 80
An enterprise assistant must maintain the same output format across several departments despite inconsistent employee phrasing. What should the solution architect prioritize?
- Specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
- Delimit external document text as data, not instructions, and keep tool permissions separate.
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Extract shared policy into a versioned prompt component managed by release control.
Explanation
Correct Answer
C:Higher-level instructions provide reusable behavior; app-side validation still checks the actual response contract.
Incorrect Answers
A:The proposed measure is aimed at sensitive-inference guardrails, not centralized stable instruction design. The architecture must instead put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template; a separate change is still needed for centralized stable instruction design.
B:The relevant trade-off is not primarily trust boundaries in retrieved documents. A successful fix needs centralized stable instruction design; otherwise centralized stable instruction design remains unverified even after this change.
D:The option belongs to the broader work on versioned reusable prompt policy, but it cannot by itself deliver centralized stable instruction design. A decision on this case needs the team to put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template and verify that outcome.
Question 81
A customer message contains text claiming to be a system administrator who can disable safety reviews. Which design choice best matches the stated constraint?
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
- Delimit external document text as data, not instructions, and keep tool permissions separate.
- Record desired content elements in a testable output rubric before optimizing concision.
Explanation
Correct Answer
B:Instruction hierarchy and server-side controls prevent lower-trust content from becoming an actual access grant.
Incorrect Answers
A:The proposed action prioritizes centralized stable instruction design. That can coexist with the right solution, but it cannot replace prompt-injection-resistant authorization: the application must still treat user-supplied content as untrusted and enforce authorization outside the prompt.
C:This addresses trust boundaries in retrieved documents; however, the problem asks for prompt-injection-resistant authorization. Without the safeguard to treat user-supplied content as untrusted and enforce authorization outside the prompt, the original failure mode is still possible.
D:That choice would be responsive to mandatory information preserved through revisions, but the reported defect is different. It leaves prompt-injection-resistant authorization unproved because there is no commitment to treat user-supplied content as untrusted and enforce authorization outside the prompt.
Question 82
A team copies a 200-line policy into every individual user prompt and receives inconsistent revisions. Which architecture adjustment is most appropriate?
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
- Extract shared policy into a versioned prompt component managed by release control.
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
Explanation
Correct Answer
C:A single maintained template avoids silent divergence and supports controlled regression testing. Make versioned reusable prompt policy part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:Improving centralized stable instruction design does not imply versioned reusable prompt policy. The selected design must extract shared policy into a versioned prompt component managed by release control, whereas this choice omits that direct protection or verification step.
B:This recommendation targets sensitive-inference guardrails at a different point in the workflow. It leaves versioned reusable prompt policy outside the design contract and does not require the team to extract shared policy into a versioned prompt component managed by release control.
D:This would make sense as an additional project activity for prompt-injection-resistant authorization. The priority required here is versioned reusable prompt policy, so the workflow still needs a control that will extract shared policy into a versioned prompt component managed by release control.
Question 83
A structured invoice extractor occasionally omits required keys even when the instruction says to use JSON. Which next step is most defensible?
- Combine a clear schema description with parsing, validation and retry-or-escalate handling.
- Design prompts to request missing information or return an explicit missing-data state.
- State authority, escalation rules and examples of excluded tasks in the system policy.
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
Explanation
Correct Answer
A:Prompts can encourage a format, but application validation is needed before downstream reliance. Make schema-backed output reliability part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
B:This shifts attention toward correct handling of missing required inputs when the relevant decision is about schema-backed output reliability. Unless the architecture can combine a clear schema description with parsing, validation and retry-or-escalate handling, the failure described in the stem remains unresolved.
C:For a workflow whose primary weakness was role-specific escalation rules, that proposal could be reasonable. Here the unmet obligation is schema-backed output reliability, and that requires the system to combine a clear schema description with parsing, validation and retry-or-escalate handling.
D:Although centralized stable instruction design can be valuable, it is not the acceptance condition in the question. The missing provision is schema-backed output reliability, addressed by the control to combine a clear schema description with parsing, validation and retry-or-escalate handling.
Question 84
A legal advice bot invents citations if the uploaded corpus contains no relevant law. Which approach should be recommended?
- State authority, escalation rules and examples of excluded tasks in the system policy.
- Explicitly instruct evidence-based answering and abstention, then verify source references in code.
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Extract shared policy into a versioned prompt component managed by release control.
Explanation
Correct Answer
B:Grounding policy and validation make unsupported outputs visible rather than rewarding confident completion. Make enforced evidence and abstention behavior part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:That choice would be responsive to role-specific escalation rules, but the reported defect is different. It leaves enforced evidence and abstention behavior unproved because there is no commitment to explicitly instruct evidence-based answering and abstention, then verify source references in code.
C:Consider its actual operating scope: prompt-injection-resistant authorization. The requested assurance is enforced evidence and abstention behavior, which calls for an implementation that can explicitly instruct evidence-based answering and abstention, then verify source references in code, not merely for a related improvement.
D:The alternative could reduce risk associated with centralized stable instruction design, yet the described behavior requires enforced evidence and abstention behavior. That outcome depends on the capability to explicitly instruct evidence-based answering and abstention, then verify source references in code.
E:The proposed measure is aimed at versioned reusable prompt policy, not enforced evidence and abstention behavior. The architecture must instead explicitly instruct evidence-based answering and abstention, then verify source references in code; a separate change is still needed for enforced evidence and abstention behavior.
Question 85
A recruitment assistant produces demographic assumptions from resumes, contrary to organizational policy. Which change most directly satisfies this requirement?
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- State authority, escalation rules and examples of excluded tasks in the system policy.
- Specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
- Extract shared policy into a versioned prompt component managed by release control.
Explanation
Correct Answer
C:Prompt instructions are one layer of fairness control; testing and review must catch policy violations. Make sensitive-inference guardrails part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:This would make sense as an additional project activity for centralized stable instruction design. The priority required here is sensitive-inference guardrails, so the workflow still needs a control that will specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
B:By itself, this improves role-specific escalation rules, but the case is not asking the architect to optimize that dimension. The requirement is sensitive-inference guardrails, supported by an approach to specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
D:The option might improve versioned reusable prompt policy; that remains insufficient for sensitive-inference guardrails. The case is resolved by sensitive-inference guardrails, which depends on a design that will specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
Question 86
A prompt change increases summary brevity but removes essential audit evidence. Which solution best resolves this operational gap?
- Record desired content elements in a testable output rubric before optimizing concision.
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
- Delimit external document text as data, not instructions, and keep tool permissions separate.
- Combine a clear schema description with parsing, validation and retry-or-escalate handling.
Explanation
Correct Answer
A:Prompt templates should encode nonnegotiable information so style tweaks do not regress correctness. Make mandatory information preserved through revisions part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
B:It is aimed at a different outcome, prompt-injection-resistant authorization. No evidence is given that it provides mandatory information preserved through revisions; the needed mechanism is to record desired content elements in a testable output rubric before optimizing concision at the appropriate boundary.
C:It emphasizes trust boundaries in retrieved documents without closing the stated gap in mandatory information preserved through revisions. By contrast, the remedy to record desired content elements in a testable output rubric before optimizing concision acts on the condition that makes this scenario fail.
D:That is primarily a technique for schema-backed output reliability. It lacks the explicit measure for mandatory information preserved through revisions that is necessary here, namely a system that will record desired content elements in a testable output rubric before optimizing concision.
Question 87
The assistant receives customer instructions embedded in a PDF that conflict with the firm's approved workflow. What design decision would best fit these facts?
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
- Delimit external document text as data, not instructions, and keep tool permissions separate.
- Specify prohibited inference boundaries and evaluate representative edge cases with hard downstream checks.
Explanation
Correct Answer
C:Treating retrieved content as untrusted lowers the risk of instruction laundering via RAG. Make trust boundaries in retrieved documents part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:The alternative could reduce risk associated with centralized stable instruction design, yet the described behavior requires trust boundaries in retrieved documents. That outcome depends on the capability to delimit external document text as data, not instructions, and keep tool permissions separate.
B:The proposed measure is aimed at prompt-injection-resistant authorization, not trust boundaries in retrieved documents. The architecture must instead delimit external document text as data, not instructions, and keep tool permissions separate; a separate change is still needed for trust boundaries in retrieved documents.
D:The control would be justified if sensitive-inference guardrails were the limiting factor. In this scenario the decisive target is trust boundaries in retrieved documents; success requires an architecture that will delimit external document text as data, not instructions, and keep tool permissions separate.
Question 88
An internal agent needs to refuse a task outside its defined role but employees use ambiguous language. Which intervention most directly addresses the failure?
- Treat user-supplied content as untrusted and enforce authorization outside the prompt.
- State authority, escalation rules and examples of excluded tasks in the system policy.
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Design prompts to request missing information or return an explicit missing-data state.
Explanation
Correct Answer
B:Explicit boundary guidance reduces scope creep; external permissions remain the final enforcement layer. Make role-specific escalation rules part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:The option might improve prompt-injection-resistant authorization; that remains insufficient for role-specific escalation rules. The case is resolved by role-specific escalation rules, which depends on a design that will state authority, escalation rules and examples of excluded tasks in the system policy.
C:The design concern is real but different: centralized stable instruction design. It does not impose the specific requirement for role-specific escalation rules; a suitable solution must state authority, escalation rules and examples of excluded tasks in the system policy and demonstrate the result.
D:This addresses correct handling of missing required inputs; however, the problem asks for role-specific escalation rules. Without the safeguard to state authority, escalation rules and examples of excluded tasks in the system policy, the original failure mode is still possible.
Question 89
A model frequently interprets missing input fields as permission to guess values. What should the solution architect prioritize?
- Extract shared policy into a versioned prompt component managed by release control.
- Put stable role, policy boundaries and output requirements in a controlled system instruction and versioned template.
- Design prompts to request missing information or return an explicit missing-data state.
- State authority, escalation rules and examples of excluded tasks in the system policy.
Explanation
Correct Answer
C:A documented uncertainty branch is safer than manufactured fields in transactional applications. Make correct handling of missing required inputs part of a versioned application contract, with host-side enforcement for properties a prompt cannot guarantee.
Incorrect Answers
A:That is primarily a technique for versioned reusable prompt policy. It lacks the explicit measure for correct handling of missing required inputs that is necessary here, namely a system that will design prompts to request missing information or return an explicit missing-data state.
B:Improving centralized stable instruction design does not imply correct handling of missing required inputs. The selected design must design prompts to request missing information or return an explicit missing-data state, whereas this choice omits that direct protection or verification step.
D:This is a valid architectural concern when optimizing role-specific escalation rules. Here it leaves correct handling of missing required inputs unaddressed, whereas the intended change must design prompts to request missing information or return an explicit missing-data state before the case can be considered resolved.
Question 90
A structured labeling task uses unusual internal categories that the model confuses with common English meanings. Which design choice best matches the stated constraint?
- Separate development examples from held-out evaluation instances and prevent data leakage.
- Use balanced, varied demonstrations and check position sensitivity.
- Add a small, diverse set of labeled examples covering borderline distinctions.
- Keep a concise zero-shot instruction and verify outcomes on a held-out set.
Explanation
Correct Answer
C:Few-shot patterns clarify proprietary decision boundaries better than repeating vague category names. Check few-shot category boundary calibration using independent cases and verified intermediate evidence, because demonstrations can otherwise hide brittle assumptions.
Incorrect Answers
A:The control would be justified if uncontaminated few-shot evaluation were the limiting factor. In this scenario the decisive target is few-shot category boundary calibration; success requires an architecture that will add a small, diverse set of labeled examples covering borderline distinctions.
B:The option belongs to the broader work on robust example ordering, but it cannot by itself deliver few-shot category boundary calibration. A decision on this case needs the team to add a small, diverse set of labeled examples covering borderline distinctions and verify that outcome.
D:For a workflow whose primary weakness was efficient zero-shot operation, that proposal could be reasonable. Here the unmet obligation is few-shot category boundary calibration, and that requires the system to add a small, diverse set of labeled examples covering borderline distinctions.