Question 70
A retrieval product will ingest thousands of PDFs. What infrastructure input must be sized before choosing the request topology?
- Specify allowed data flows, retention obligations, and residency constraints.
- In this design, convert the request into measurable quality, latency, throughput, and risk targets.
- Document supported languages and measure task quality for each representative locale.
- Estimate document volume, query rate, indexing load, and latency budgets.
- Set out perceived latency and progressive-response behavior, including streaming.
Explanation
Correct Answer
D:Retrieval storage and preprocessing costs differ from model inference; capacity planning needs both workloads.
Incorrect Answers
A:Information governance determines which data can cross a service boundary and which controls the design must provide; the scenario's controlling requirement is retrieval infrastructure sizing, not the related regulated data boundaries concern.
B:Acceptance criteria turn adjectives into testable constraints; model selection without task metrics can optimize the wrong property; the effect concerns measurable product requirements; the success condition is retrieval infrastructure sizing.
C:Language performance varies by input distribution and task type, so a single English evaluation cannot establish fitness; when the goal is retrieval infrastructure sizing, optimizing multilingual quality targets misses that operational boundary.
E:Time to first useful output may matter more than complete-generation time for an interactive interface; the SLA should reflect that; this targets conversational latency requirements, while the system must demonstrate retrieval infrastructure sizing.
Question 71
A company wants to compare a one-call prompt against an agent solution. What should the technical design specify for the comparison?
- Assess source-supported claims and rejection of unsupported questions on a held-out set.
- Distinguish asynchronous completion deadlines from interactive response requirements.
- Fix representative tasks, success criteria, and operational budget for both approaches.
- Document which decisions, tool calls, versions, and approvals must be traceable.
Explanation
Correct Answer
C:Comparative evaluation is only meaningful when task difficulty, measured outcomes, and cost constraints are held consistent.
Incorrect Answers
A:Grounding is a product behavior that needs verification against retrieval evidence, not a vague instruction to avoid hallucinations; the effect concerns grounded answer acceptance; the success condition is architecture evaluation baseline.
B:Deferred workloads can be queued and batched for efficiency, while live requests need immediate request-response behavior; changing deferred completion SLA does not establish architecture evaluation baseline in this application.
D:Audit logs are useful only when they capture the events needed to reconstruct actions and establish accountable control; even if multi-criterion integration succeeds, architecture evaluation baseline still needs separate verification.
Question 72
A support app must continue answering simple FAQs if its order-management API is unavailable. Which nonfunctional requirement should be written?
- Describe allowed data flows, retention obligations, and residency constraints.
- Set out a shared application service contract and client-specific interaction needs.
- Document authorization roles and approval thresholds for each mutating action.
- Define degraded-mode behavior and safe dependency-failure responses.
Explanation
Correct Answer
D:Graceful degradation lets the application preserve read-only functionality while refusing actions that require unavailable authoritative systems.
Incorrect Answers
A:Information governance determines which data can cross a service boundary and which controls the design must provide; the consequence is improved regulated data boundaries, not verified dependency failure policy.
B:A stable backend API can centralize model rules and security, while each client has different presentation and latency expectations; it changes multichannel integration boundaries; it does not prove dependency failure policy in this case.
C:Tool design depends on who can commit side effects; missing approval requirements cannot be repaired by prompt wording; for dependency failure policy, this instead manages action authorization requirements; the requested decision remains unresolved.
Question 73
An enterprise app uses customer records and depends on a model that can be replaced during a release. Which TWO engineering controls are necessary for a responsible rollout? Choose TWO.
- Store a single shared key in the client application.
- Compare the new model and prompt on a versioned acceptance suite before deployment.
- Rely on informal demonstrations as the only approval test.
- Use the public API without a server-side authorization layer.
- Set explicit rules for allowed data use and storage before API integration.
Explanation
Correct Answers
B:Controlled regression evaluation exposes behavior changes while the old configuration remains available for rollback. Comparison should include representative customer cases, safety requirements, latency, and a predefined rollback threshold.
E:Residency, retention, and permitted data flows determine whether sensitive records may be sent to an inference endpoint. Data handling constraints must also cover retention, permitted regions, and authorized downstream tools.
Incorrect Answers
A:Client-visible credentials increase exposure and do not validate whether the model meets quality requirements. A key embedded in a client can be extracted and reused outside the intended enterprise policy boundary.
C:Demonstrations sample too few cases to reveal systematic regression or exposure issues. A repeatable benchmark reveals edge cases that a polished demonstration may never exercise.
D:Direct unauthenticated access does not establish data rights or lifecycle release quality. A public endpoint does not remove the application's obligation to authenticate users and authorize access to records.
Question 74
A prompt revision improves demo results but causes production regressions. What release practice was missing?
- Supply rollout controls and a fast feature rollback or kill switch.
- Execute integration tests with simulated throttling and dependency failures.
- Evaluate versioned prompts on representative regression cases before deployment.
- Contrast API contract tests and representative behavior before and after the change.
Explanation
Correct Answer
C:A prompt is executable application behavior; test gates and rollback history should cover it like code. A tested prompt release should record its revision and provide a rollback target when evaluation metrics or user outcomes deteriorate.
Incorrect Answers
A:Operational containment limits impact when a new model, prompt, or connector misbehaves after deployment; it changes controllable production rollback; it does not prove prompt release regression testing in this case.
B:Unit tests of isolated logic do not prove the complete request, retry, and tool-response path behaves correctly; for prompt release regression testing, this instead manages integration failure testing; the requested decision remains unresolved.
D:Internal restructuring should not silently alter externally promised response schemas, error behavior, or supported user workflows; because refactoring contract preservation is its concern, prompt release regression testing remains outside its effect here.
Question 75
A production model identifier will be retired next quarter. What should the maintenance plan include?
- Allocate measurable acceptance criteria and responsibility for ongoing quality.
- Test an available successor and schedule a controlled migration before retirement.
- Write down and rehearse restoration of configuration, durable state, and secret references.
- Include runtime observability and incident response for model and tool outcomes.
Explanation
Correct Answer
B:Deprecation and retirement are lifecycle events; delaying validation can cause avoidable service disruption or behavior changes.
Incorrect Answers
A:AI capabilities require defined ownership for evaluations, observed failures, and user-facing policy decisions after release; the consequence is improved lifecycle ownership, not verified planned model retirement response.
C:Recovery requires more than redeploying code; operations must reconstruct approved settings and resume tasks without duplicating side effects; this implements operational recovery readiness rather than the demanded planned model retirement response.
D:Deployment is not the end of the system lifecycle; ongoing traces, metrics, and triage are needed as inputs and dependencies change; the scenario's controlling requirement is planned model retirement response, not the related postdeployment operations concern.
Question 76
A team ships an agent without tracking its tool interactions after go-live. What operational phase has been neglected?
- Give measurable acceptance criteria and responsibility for ongoing quality.
- Evaluate an available successor and schedule a controlled migration before retirement.
- Add runtime observability and incident response for model and tool outcomes.
- Maintain revisions of and test integration contracts, and handle incompatible payloads explicitly.
- Track data and knowledge-base changes as inputs to regression evaluation.
Explanation
Correct Answer
C:Deployment is not the end of the system lifecycle; ongoing traces, metrics, and triage are needed as inputs and dependencies change.
Incorrect Answers
A:AI capabilities require defined ownership for evaluations, observed failures, and user-facing policy decisions after release; the effect concerns lifecycle ownership; the success condition is postdeployment operations.
B:Deprecation and retirement are lifecycle events; delaying validation can cause avoidable service disruption or behavior changes; because planned model retirement response is its concern, postdeployment operations remains outside its effect here.
D:Schema evolution affects the application's tool boundary even if model code is untouched; contract discipline catches breaks early; the scenario's controlling requirement is postdeployment operations, not the related dependent API schema evolution concern.
E:A stable prompt does not guarantee stable results when source documents, schemas, and retrieval rankings change; it follows data-driven regression drift constraints but leaves postdeployment operations unspecified.
Question 77
A new policy requires redacting certain identifiers before any model request. Where does that work belong in the development lifecycle?
- Introduce runtime observability and incident response for model and tool outcomes.
- Add privacy transformations to design, implementation, and automated verification.
- Exercise an available successor and schedule a controlled migration before retirement.
- Record and rehearse restoration of configuration, durable state, and secret references.
Explanation
Correct Answer
B:Compliance is cross-cutting; a late manual checklist cannot reliably protect every request path. Put the transformation in shared data-handling code, and exercise automated cases for file uploads, direct text, and asynchronous requests.
Incorrect Answers
A:Deployment is not the end of the system lifecycle; ongoing traces, metrics, and triage are needed as inputs and dependencies change; the consequence is improved postdeployment operations, not verified privacy controls across SDLC.
C:Deprecation and retirement are lifecycle events; delaying validation can cause avoidable service disruption or behavior changes; even if planned model retirement response succeeds, privacy controls across SDLC still needs separate verification.
D:Recovery requires more than redeploying code; operations must reconstruct approved settings and resume tasks without duplicating side effects; a remedy for operational recovery readiness is not a substitute for privacy controls across SDLC.
Question 78
An agent passes unit tests but fails when a real connector rate-limits it. What testing stage should address this?
- Observe data and knowledge-base changes as inputs to regression evaluation.
- Assess versioned prompts on representative regression cases before deployment.
- Run integration tests with simulated throttling and dependency failures.
- Track versions of and test integration contracts, and handle incompatible payloads explicitly.
Explanation
Correct Answer
C:Unit tests of isolated logic do not prove the complete request, retry, and tool-response path behaves correctly.
Incorrect Answers
A:A stable prompt does not guarantee stable results when source documents, schemas, and retrieval rankings change; for integration failure testing, a control centered on data-driven regression drift is not the appropriate substitute.
B:A prompt is executable application behavior; test gates and rollback history should cover it like code; it follows prompt release regression testing constraints but leaves integration failure testing unspecified.
D:Schema evolution affects the application's tool boundary even if model code is untouched; contract discipline catches breaks early; the consequence is improved dependent API schema evolution, not verified integration failure testing.
Question 79
A proposed AI release has no defined way to disable risky tools during an incident. What should release engineering add?
- Include privacy transformations to design, implementation, and automated verification.
- Provide rollout controls and a fast feature rollback or kill switch.
- Evaluate API contract tests and representative behavior before and after the change.
- Label versions of and test integration contracts, and handle incompatible payloads explicitly.
Explanation
Correct Answer
B:Operational containment limits impact when a new model, prompt, or connector misbehaves after deployment. A kill switch should revoke the affected capability promptly while preserving an audit trail for incident investigation.
Incorrect Answers
A:Compliance is cross-cutting; a late manual checklist cannot reliably protect every request path; for this request, controllable production rollback is the missing piece, not privacy controls across SDLC.
C:Internal restructuring should not silently alter externally promised response schemas, error behavior, or supported user workflows; it follows refactoring contract preservation constraints but leaves controllable production rollback unspecified.
D:Schema evolution affects the application's tool boundary even if model code is untouched; contract discipline catches breaks early; the effect concerns dependent API schema evolution; the success condition is controllable production rollback.
Question 80
A support agent uses evolving retrieval content, but developers only test the original prompt. What lifecycle risk remains?
- Review versioned prompts on representative regression cases before deployment.
- Monitor data and knowledge-base changes as inputs to regression evaluation.
- Offer rollout controls and a fast feature rollback or kill switch.
- Set measurable acceptance criteria and responsibility for ongoing quality.
Explanation
Correct Answer
B:A stable prompt does not guarantee stable results when source documents, schemas, and retrieval rankings change.
Incorrect Answers
A:A prompt is executable application behavior; test gates and rollback history should cover it like code; for data-driven regression drift, a control centered on prompt release regression testing is not the appropriate substitute.
C:Operational containment limits impact when a new model, prompt, or connector misbehaves after deployment; this implements controllable production rollback rather than the demanded data-driven regression drift.
D:AI capabilities require defined ownership for evaluations, observed failures, and user-facing policy decisions after release; testing data-driven regression drift would still fail after improving lifecycle ownership.