Question 20
A service desk receives bursts of tickets and requires reliable delivery to several downstream systems. Which response follows from the actual operational constraint?
- For this deployment, persist state and idempotency keys around tool actions with bounded retries and a failure queue.
- As a scoped first step, decouple ingestion from inference and side effects using queues, workers and back-pressure.
- In the release design, add user feedback capture with source identifiers and a reviewed correction-to-evaluation loop.
- During architecture review, separate intake normalization, Claude classification, schema validation and controlled record creation.
Explanation
Correct Answer
B:Asynchronous buffering isolates burst traffic and permits bounded concurrency without dropping requests; validation needs to show that a load test preserves accepted ticket count while worker concurrency stays within budget.
Incorrect Answers
A:Durable execution separates transient external failures from duplicate irreversible side effects; this applies when an operations workflow needs recoverable execution after one downstream tool times out, but here a service desk receives bursts of tickets and requires reliable delivery to several downstream systems.
C:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; it is justified if a document assistant returns answers but users cannot correct an inaccurate citation, not the case in which a service desk receives bursts of tickets and requires reliable delivery to several downstream systems.
D:Explicit stage boundaries prevent free-form model output from directly mutating the case database; its scope includes cases where a triage assistant receives messy emails and should create only validated structured case records, whereas the problem presented is that a service desk receives bursts of tickets and requires reliable delivery to several downstream systems.
Question 21
A knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions. Which change targets the constraint rather than a proxy metric?
- In the implementation sequence, version prompts, retrieval indexes, models and evaluations together with trace identifiers.
- For the operating team, persist state and idempotency keys around tool actions with bounded retries and a failure queue.
- For this deployment, decouple ingestion from inference and side effects using queues, workers and back-pressure.
- As a scoped first step, add user feedback capture with source identifiers and a reviewed correction-to-evaluation loop.
Explanation
Correct Answer
A:Change attribution needs correlated component versions and item-level evidence across the entire request path; production testing should demonstrate that a failed sample can be traced to the exact prompt, model, corpus snapshot and retrieval result.
Incorrect Answers
B:Durable execution separates transient external failures from duplicate irreversible side effects; this makes sense when an operations workflow needs recoverable execution after one downstream tool times out, while this problem involves a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions.
C:Asynchronous buffering isolates burst traffic and permits bounded concurrency without dropping requests; this technique fits where a service desk receives bursts of tickets and requires reliable delivery to several downstream systems, and not where a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions.
D:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; the trigger would be that a document assistant returns answers but users cannot correct an inaccurate citation, rather than the risk that a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions.
Question 22
A message-processing pilot writes case records correctly until a deliberately malformed category reaches its persistence API. Which measurement is the most relevant release criterion?
- The scenario test demonstrates that malformed categories are rejected before any write, and valid records include source provenance.
- The end-to-end check confirms that an unapproved draft remains invisible to patient-facing systems in an end-to-end test.
- The observed outcome is that a cross-role probe returns no snippets or citations from unauthorized tax folders.
- The regression suite verifies that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
Explanation
Correct Answer
A:Explicit stage boundaries prevent free-form model output from directly mutating the case database; the reviewer is concerned with situations where a triage assistant receives messy emails and should create only validated structured case records.
Incorrect Answers
B:The output boundary enforces clinical review independently of the model's confidence in the summary; the other workload is one where a healthcare summarizer cannot expose partial notes before a clinician approves them, whereas the release criterion addresses a message-processing pilot writes case records correctly until a deliberately malformed category reaches its persistence API.
C:Authorized evidence must be constrained upstream; a final natural-language warning cannot substitute for ACL enforcement; this applies when a tax research tool generates plausible citations pointing to documents the user cannot view, while the required safeguard concerns a message-processing pilot writes case records correctly until a deliberately malformed category reaches its persistence API.
D:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; the alternative helps if a document assistant returns answers but users cannot correct an inaccurate citation, not the risk where a message-processing pilot writes case records correctly until a deliberately malformed category reaches its persistence API.
Question 23
A citation repair is approved by a reviewer, but the next evaluation still uses an obsolete source version. Which production-like test finding is most convincing?
- A controlled trial finds that a timeout replay completes once without creating duplicate orders or losing audit events.
- The measured evidence confirms that a cross-role probe returns no snippets or citations from unauthorized tax folders.
- The scenario test demonstrates that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
- The verified result is that an unapproved draft remains invisible to patient-facing systems in an end-to-end test.
Explanation
Correct Answer
C:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; this checks the situation in which a document assistant returns answers but users cannot correct an inaccurate citation.
Incorrect Answers
A:Durable execution separates transient external failures from duplicate irreversible side effects; it is relevant if an operations workflow needs recoverable execution after one downstream tool times out, instead of when a citation repair is approved by a reviewer, but the next evaluation still uses an obsolete source version.
B:Authorized evidence must be constrained upstream; a final natural-language warning cannot substitute for ACL enforcement; it serves conditions where a tax research tool generates plausible citations pointing to documents the user cannot view, not the case in which a citation repair is approved by a reviewer, but the next evaluation still uses an obsolete source version.
D:The output boundary enforces clinical review independently of the model's confidence in the summary; this makes sense when a healthcare summarizer cannot expose partial notes before a clinician approves them, but this decision concerns a citation repair is approved by a reviewer, but the next evaluation still uses an obsolete source version.
Question 24
A payment workflow times out just after a remote commit, and the next attempt might repeat the transaction. Which finding should the release gate require?
- The representative case test finds that malformed categories are rejected before any write, and valid records include source provenance.
- The negative-case evaluation confirms that a timeout replay completes once without creating duplicate orders or losing audit events.
- A controlled trial finds that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
- The measured evidence confirms that an unapproved draft remains invisible to patient-facing systems in an end-to-end test.
Explanation
Correct Answer
B:Durable execution separates transient external failures from duplicate irreversible side effects; it is the decisive test when an operations workflow needs recoverable execution after one downstream tool times out.
Incorrect Answers
A:Explicit stage boundaries prevent free-form model output from directly mutating the case database; that would be useful if a triage assistant receives messy emails and should create only validated structured case records, while this problem involves a payment workflow times out just after a remote commit, and the next attempt might repeat the transaction.
C:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; that is defensible if a document assistant returns answers but users cannot correct an inaccurate citation, rather than the risk that a payment workflow times out just after a remote commit, and the next attempt might repeat the transaction.
D:The output boundary enforces clinical review independently of the model's confidence in the summary; the other workload is one where a healthcare summarizer cannot expose partial notes before a clinician approves them, rather than the condition that a payment workflow times out just after a remote commit, and the next attempt might repeat the transaction.
Question 25
A clinician sees draft summaries in a downstream portal before anyone has recorded approval. Which verified outcome most directly addresses the reported risk?
- The verified result is that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
- The recorded test establishes that an unapproved draft remains invisible to patient-facing systems in an end-to-end test.
- The representative case test finds that a failed sample can be traced to the exact prompt, model, corpus snapshot and retrieval result.
- The negative-case evaluation confirms that a cross-role probe returns no snippets or citations from unauthorized tax folders.
Explanation
Correct Answer
B:The output boundary enforces clinical review independently of the model's confidence in the summary; this is a functional check for the case where a healthcare summarizer cannot expose partial notes before a clinician approves them.
Incorrect Answers
A:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; the approach helps when a document assistant returns answers but users cannot correct an inaccurate citation, whereas the release criterion addresses a clinician sees draft summaries in a downstream portal before anyone has recorded approval.
C:Change attribution needs correlated component versions and item-level evidence across the entire request path; it is relevant if a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions, not the risk where a clinician sees draft summaries in a downstream portal before anyone has recorded approval.
D:Authorized evidence must be constrained upstream; a final natural-language warning cannot substitute for ACL enforcement; it serves conditions where a tax research tool generates plausible citations pointing to documents the user cannot view, not this situation where a clinician sees draft summaries in a downstream portal before anyone has recorded approval.
Question 26
Two employees query an indexed archive, but one can still view citations into the other's restricted folder. What independently verifiable observation is needed?
- The readiness assessment shows that a cross-role probe returns no snippets or citations from unauthorized tax folders.
- The independent test establishes that a load test preserves accepted ticket count while worker concurrency stays within budget.
- The verified result is that malformed categories are rejected before any write, and valid records include source provenance.
- The recorded test establishes that a timeout replay completes once without creating duplicate orders or losing audit events.
- The representative case test finds that an unapproved draft remains invisible to patient-facing systems in an end-to-end test.
Explanation
Correct Answer
A:Authorized evidence must be constrained upstream; a final natural-language warning cannot substitute for ACL enforcement; the required safeguard is exercised when a tax research tool generates plausible citations pointing to documents the user cannot view.
Incorrect Answers
B:Asynchronous buffering isolates burst traffic and permits bounded concurrency without dropping requests; this control is needed when a service desk receives bursts of tickets and requires reliable delivery to several downstream systems, whereas the problem presented is that two employees query an indexed archive, but one can still view citations into the other's restricted folder.
C:Explicit stage boundaries prevent free-form model output from directly mutating the case database; that would be useful if a triage assistant receives messy emails and should create only validated structured case records, but this decision concerns two employees query an indexed archive, but one can still view citations into the other's restricted folder.
D:Durable execution separates transient external failures from duplicate irreversible side effects; it serves a setting where an operations workflow needs recoverable execution after one downstream tool times out, but the actual risk is that two employees query an indexed archive, but one can still view citations into the other's restricted folder.
E:The output boundary enforces clinical review independently of the model's confidence in the summary; that is defensible if a healthcare summarizer cannot expose partial notes before a clinician approves them, but the requested boundary concerns two employees query an indexed archive, but one can still view citations into the other's restricted folder.
Question 27
A backlog surge exceeds model worker capacity, while the upstream application continues accepting requests. What outcome would demonstrate this requirement in operation?
- The staged rollout demonstrates that a load test preserves accepted ticket count while worker concurrency stays within budget.
- The audit sample demonstrates that a failed sample can be traced to the exact prompt, model, corpus snapshot and retrieval result.
- The readiness assessment shows that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
- The independent test establishes that a timeout replay completes once without creating duplicate orders or losing audit events.
Explanation
Correct Answer
A:Asynchronous buffering isolates burst traffic and permits bounded concurrency without dropping requests; it is the decisive test when a service desk receives bursts of tickets and requires reliable delivery to several downstream systems.
Incorrect Answers
B:Change attribution needs correlated component versions and item-level evidence across the entire request path; it targets cases where a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions, rather than the risk that a backlog surge exceeds model worker capacity, while the upstream application continues accepting requests.
C:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; the approach helps when a document assistant returns answers but users cannot correct an inaccurate citation, rather than the condition that a backlog surge exceeds model worker capacity, while the upstream application continues accepting requests.
D:Durable execution separates transient external failures from duplicate irreversible side effects; that belongs to situations where an operations workflow needs recoverable execution after one downstream tool times out, while the immediate question involves a backlog surge exceeds model worker capacity, while the upstream application continues accepting requests.
Question 28
A regression appears after an index rebuild, and engineers cannot identify which artifact version handled a failed request. Which observed test outcome would justify sign-off?
- The pilot evaluation finds that reviewer-approved corrections become versioned regression cases linked to retrieval sources.
- The comparison run confirms that a load test preserves accepted ticket count while worker concurrency stays within budget.
- The staged rollout demonstrates that a failed sample can be traced to the exact prompt, model, corpus snapshot and retrieval result.
- The audit sample demonstrates that a cross-role probe returns no snippets or citations from unauthorized tax folders.
Explanation
Correct Answer
C:Change attribution needs correlated component versions and item-level evidence across the entire request path; the condition under test is that a knowledge assistant runs weekly releases but has no way to identify whether retrieval or prompting caused regressions.
Incorrect Answers
A:Traceable corrections make retrieval and model behavior improvable rather than accumulating unclassified complaints; the trigger would be that a document assistant returns answers but users cannot correct an inaccurate citation, while the required safeguard concerns a regression appears after an index rebuild, and engineers cannot identify which artifact version handled a failed request.
B:Asynchronous buffering isolates burst traffic and permits bounded concurrency without dropping requests; the design helps when a service desk receives bursts of tickets and requires reliable delivery to several downstream systems, not the risk where a regression appears after an index rebuild, and engineers cannot identify which artifact version handled a failed request.
D:Authorized evidence must be constrained upstream; a final natural-language warning cannot substitute for ACL enforcement; this control is needed when a tax research tool generates plausible citations pointing to documents the user cannot view, whereas this scenario involves a regression appears after an index rebuild, and engineers cannot identify which artifact version handled a failed request.
Question 29
A fixed sequence of review steps is already approved by auditors and must not vary. Which proposed design action directly addresses the risk?
- In the initial solution, use an evaluator-optimizer loop with bounded iterations and test-driven feedback.
- In the proposed design, keep rules in deterministic application code and ask Claude only for ambiguous text extraction.
- Under the reviewed approach, use an augmented single Claude call with retrieved policy context and structured output.
- For this production path, use a deterministic orchestrated workflow with Claude inside bounded steps.
Explanation
Correct Answer
D:A known sequence benefits from explicit transitions and review gates rather than discretionary planning; a clear sign-off signal is that every test run follows the approved transition graph regardless of model phrasing.
Incorrect Answers
A:Iteration is justified because objectively checked outputs can drive localized improvement; the alternative helps if a coding assistant must repeat lint, test and targeted edits until checks pass or budget expires, whereas the problem presented is that a fixed sequence of review steps is already approved by auditors and must not vary.
B:Representing known rules as agent reasoning increases cost and variability without adding capability; it is justified if an approval application has a complicated branching form but every branch is fully specified by rules, but this decision concerns a fixed sequence of review steps is already approved by auditors and must not vary.
C:Adding autonomous execution or multi-stage orchestration has no benefit for a self-contained classification; its scope includes cases where a policy classifier only needs a one-pass answer with a document excerpt, but the actual risk is that a fixed sequence of review steps is already approved by auditors and must not vary.
Question 30
A market analyst must decide dynamically which unfamiliar websites and datasets to investigate. Which option supplies the needed architectural boundary?
- At this decision point, use an augmented single Claude call with retrieved policy context and structured output.
- As a candidate control, use a bounded agent loop with permitted tools, budgets and stopping criteria.
- For this deployment, use an evaluator-optimizer loop with bounded iterations and test-driven feedback.
- In the proposed design, use parallel worker calls followed by an explicitly defined aggregation stage.
Explanation
Correct Answer
B:Exploration requires model-directed next actions because the relevant evidence path is unknown upfront; a pilot is successful when trace reviews show useful tool selection while limits stop irrelevant or repeated searches.
Incorrect Answers
A:Adding autonomous execution or multi-stage orchestration has no benefit for a self-contained classification; this makes sense when a policy classifier only needs a one-pass answer with a document excerpt, rather than the risk that a market analyst must decide dynamically which unfamiliar websites and datasets to investigate.
C:Iteration is justified because objectively checked outputs can drive localized improvement; the trigger would be that a coding assistant must repeat lint, test and targeted edits until checks pass or budget expires, while the immediate question involves a market analyst must decide dynamically which unfamiliar websites and datasets to investigate.
D:Independent subtasks improve throughput without needing agents to coordinate recursively; the design helps when an annual-report analysis can divide independent companies into parallel reviews before synthesis, but the present constraint is that a market analyst must decide dynamically which unfamiliar websites and datasets to investigate.