Question 70
An energy-sector assistant can propose maintenance actions, but management wants to remove the server-side tool-execution policy gate. Which failure becomes most credible if approval depends only on the model deciding not to act?
- The risk that a global cost cap allows one tenant to crowd out others
- The risk that prompt-only guidance cannot neutralize malicious external text
- The risk that upstream input minimization alone may not catch derived disclosures
- The risk that a single classifier cannot handle every emergent attack pattern
- The risk that model self-restraint cannot guarantee that a bad action is blocked
Explanation
Correct Answer
E:The risk is not a generic model inconvenience: model self-restraint cannot guarantee that a bad action is blocked. The proposed safeguard changes that boundary because the control layer can require server-side authorization and business-policy checks before each side effect, and the release would need central policy service and auditable subject identity plus disallowed action attempts stopped at the executor to show the risk remains contained.
Incorrect Answers
A:If one customer can monopolize shared model capacity through abusive retries, a global cost cap allows one tenant to crowd out others; whereas removing Tool-execution policy gate primarily exposes the risk that model self-restraint cannot guarantee that a bad action is blocked.
B:When user-submitted attachments can contain instructions to steal private data, prompt-only guidance cannot neutralize malicious external text; but this is not the immediate loss associated with omitting Tool-execution policy gate: model self-restraint cannot guarantee that a bad action is blocked.
C:When generated summaries may echo sensitive identifiers unnecessary for recipients, upstream input minimization alone may not catch derived disclosures; however, the named safeguard addresses the distinct hazard that model self-restraint cannot guarantee that a bad action is blocked.
D:The other exposure arises when failures may bypass one input or model safety layer: a single classifier cannot handle every emergent attack pattern. Omitting the requested control instead leaves the more immediate risk that model self-restraint cannot guarantee that a bad action is blocked; whereas removing Tool-execution policy gate primarily exposes the risk that model self-restraint cannot guarantee that a bad action is blocked.
Question 71
The supply-chain planning architect proposes omitting Single-purpose tool contract from the production design. What is the most direct failure mode that decision would expose?
- The danger that eager inclusion of the entire catalog wastes context and confuses selection
- The danger that opaque universal tools hide security boundaries and failure semantics
- The danger that overlapping names lead to arbitrary selection and inconsistent answers
- The danger that mixing reading and mutation in one tool overgrants routine sessions
Explanation
Correct Answer
B:If the protection disappears, opaque universal tools hide security boundaries and failure semantics. The missing architectural distinction is that the integration can replace broad verbs with narrow parameterized operations and validation. The approval board should require an alternative with documented action schemas and server-side validation and validate it against prohibited-call rejection and schema-compliant invocation rates.
Incorrect Answers
A:The competing hazard is that eager inclusion of the entire catalog wastes context and confuses selection in a workload where a complex connector ecosystem cannot fit full tool descriptions in each request; but this is not the immediate loss associated with omitting Single-purpose tool contract: opaque universal tools hide security boundaries and failure semantics.
C:When several plugins expose overlapping outdated customer lookup functions, overlapping names lead to arbitrary selection and inconsistent answers; whereas removing Single-purpose tool contract primarily exposes the risk that opaque universal tools hide security boundaries and failure semantics.
D:The other exposure arises when an assistant must read audit events but rarely modify audit settings: mixing reading and mutation in one tool overgrants routine sessions. Omitting the requested control instead leaves the more immediate risk that opaque universal tools hide security boundaries and failure semantics; but this is not the immediate loss associated with omitting Single-purpose tool contract: opaque universal tools hide security boundaries and failure semantics.
Question 72
An independent assessment of a field-service coordination Claude system reports that median latency can hide the failures users notice most. Which change best closes the exposed control gap?
- Introduce Tail-latency measurement
- Introduce Cost per resolved ticket
- Introduce Calibration and abstention quality
- Introduce Task-level acceptance rubric
Explanation
Correct Answer
A:The remediation acts at the relevant design boundary: the evaluation process can track p95 and p99 end-to-end latency by request class. That counteracts the observed hazard that median latency can hide the failures users notice most. Rollout should include correlated request timing across dependencies, with regression evidence from high-percentile latency and timeout frequency.
Incorrect Answers
B:When raw tokens per request omit cost of corrections and failures, the evaluation process can include review, retries and inference in outcome-based unit economics; yet it would not correct the observed weakness: median latency can hide the failures users notice most.
C:Where incomplete evidence should yield escalation rather than invented certainty, the evaluation process can evaluate correctness jointly with justified abstentions across uncertainty levels; but this remedy operates on another boundary, not the case where median latency can hide the failures users notice most.
D:When generic satisfaction scores may reward confident errors, the evaluation process can define observable criteria for a usable output rather than aesthetic fluency; whereas the field-service coordination defect follows the path in which median latency can hide the failures users notice most.
Question 73
A healthcare operations architecture gate expects the following behavior: a stronger model can justify tradeoffs across heterogeneous evidence. Which TWO actions together supply the design intervention and its operating dependency? Choose TWO.
- Implement Empirical model bake-off
- Implement Reasoning-capable model for complex synthesis
- Track routed quality at the target cost and p95 latency
- Establish benchmark cases with multi-hop evidence
- Establish long-context benchmark cases and token accounting
Explanation
Correct Answers
B:The control itself makes the following capability possible: a stronger model can justify tradeoffs across heterogeneous evidence. That matters when a lightweight classifier can omit crucial cross-document reasoning; it does not replace the required enabling condition or outcome test.
D:This step creates the enabling basis of benchmark cases with multi-hop evidence. That preparation supports the mechanism that a stronger model can justify tradeoffs across heterogeneous evidence; merely choosing the control does not prove its configuration or authority is ready.
Incorrect Answers
A:This design choice is aimed at cases where stakeholders choose a model based on brand rather than target workload; it is effective because the model configuration can evaluate candidate models on identical task data and operating constraints. The rollout described is instead governed by the case where multi-document arbitration requires several interdependent inferences and requires a different primary intervention; however, the intended intervention requires that a stronger model can justify tradeoffs across heterogeneous evidence.
C:This test would validate a different benefit: a fast adequate tier handles easy cases and escalates only uncertain ones; but that indicator does not verify audited synthesis accuracy on difficult cases for the selected architecture.
E:The alternative setup enables a different mechanism: the model configuration can test effective context utilization and recall before choosing a model configuration; whereas this intervention requires benchmark cases with multi-hop evidence as its operating precondition.
Question 74
The procurement services release team needs a numerical gate for Human-approved recommendation service. Rather than equating deployment with success, it wants to catch cases in which automating the last step could bypass required accountability. Which indicator should it track?
- Audit pass rate on a stratified set of authentic exceptions
- Audit reviewer override and acceptance rates by decision category
- Audit proportion of prohibited workflows rejected before execution
- Audit risk-adjusted impact and implementation effort per use case
Explanation
Correct Answer
B:The acceptance measure focuses on reviewer override and acceptance rates by decision category at the task or workflow boundary. Model fluency would not establish whether Claude drafts ranked recommendations while a designated reviewer authorizes each action, and it could conceal cases where automating the last step could bypass required accountability.
Incorrect Answers
A:When the proposed automation encounters exceptions that have never been tested, small-scope trials expose feasibility and exception handling before a broad commitment; a favorable result on that test would not establish reviewer override and acceptance rates by decision category for this workflow, where caseworkers must retain authority for consequential final decisions; by contrast, the procurement services release needs evidence from reviewer override and acceptance rates by decision category.
C:When a client demands a guarantee of fully accurate autonomous legal advice, identifying out-of-scope judgments prevents unsupported automation promises; a favorable result on that test would not establish reviewer override and acceptance rates by decision category for this workflow, where caseworkers must retain authority for consequential final decisions; but the acceptance gate must instead inspect reviewer override and acceptance rates by decision category to detect the targeted failure.
D:The proposed measure would suit a setting where candidate use cases differ in savings, uncertainty and personal-data exposure, since a scored decision framework surfaces high-value lower-risk starting points; by contrast, the procurement services release needs evidence from reviewer override and acceptance rates by decision category.
Question 75
A legal operations team is selecting the next improvement for its AI service; a proposal lists features but omits response time and recovery requirements. What design choice has the strongest direct justification?
- Deploy Current-state service blueprint
- Deploy Acceptance criteria negotiation
- Deploy Workflow observation interviews
- Deploy Functional versus nonfunctional split
Explanation
Correct Answer
D:Under the stated constraints, the delivery team can capture distinct behavior, latency, security, availability and data constraints. That directly addresses the operational failure without relying on a generic prompt reminder. Production depends on measurable quality attributes and named owners; the result to monitor is proportion of nonfunctional requirements with acceptance tests.
Incorrect Answers
A:If automation is proposed without mapping present handoffs and failure paths, the delivery team can chart the existing input-to-outcome process and exception volumes; that does not provide the capability actually needed: the delivery team can capture distinct behavior, latency, security, availability and data constraints.
B:For a workload where stakeholders ask for a system that is accurate enough without specifying thresholds, the delivery team can turn goals into testable task-specific, risk-sensitive success criteria; the alternative cannot replace this key requirement: the delivery team can capture distinct behavior, latency, security, availability and data constraints.
C:When sponsors describe an ideal process that differs from daily operations, the delivery team can observe actual exception handling and frontline work before mapping automation; whereas the legal operations case must guard against the outcome that feature completeness does not demonstrate production acceptability.
Question 76
During solution discovery in internal audit, generated summaries may echo sensitive identifiers unnecessary for recipients. What should the architects prioritize to resolve this issue?
- Deploy Output privacy filter
- Deploy Ingress content classification
- Deploy Defense-in-depth incident response
- Deploy Rate and spend limits by tenant
Explanation
Correct Answer
A:The mechanism is specific: the control layer can enforce recipient-specific redaction and validated output fields. A credible implementation requires recipient entitlement map and redaction tests, while its acceptance test monitors sensitive data leakage rate in response audits. The alternative approaches leave the risk that upstream input minimization alone may not catch derived disclosures.
Incorrect Answers
B:When user-submitted attachments can contain instructions to steal private data, the control layer can classify untrusted inputs and prevent them from acquiring instruction authority; whereas the financial operations case must guard against the outcome that upstream input minimization alone may not catch derived disclosures.
C:For a workload where failures may bypass one input or model safety layer, the control layer can combine prevention, detection, containment and audited recovery; the missing operational behavior is specific: the control layer can enforce recipient-specific redaction and validated output fields.
D:If one customer can monopolize shared model capacity through abusive retries, the control layer can enforce bounded usage budgets and anomaly controls per tenant; the current decision instead depends on a specific capability: the control layer can enforce recipient-specific redaction and validated output fields.
Question 77
A software platform technical leadership group is designing a Claude-supported workflow because several plugins expose overlapping outdated customer lookup functions. Which architectural decision most directly meets that requirement?
- Adopt Single-purpose tool contract
- Adopt Tool ownership and lifecycle review
- Adopt Progressive tool discovery
- Adopt Task-scoped tool allowlist
- Adopt Read-write capability separation
Explanation
Correct Answer
B:The proposed control fits because the integration can retire or consolidate redundant capabilities with a responsible owner. Its production prerequisites include inventory with owner, version and retirement process; check duplicate-capability count and chosen-tool reliability on representative tasks. Without this boundary, overlapping names lead to arbitrary selection and inconsistent answers.
Incorrect Answers
A:For a workload where a generic execute-any-query tool lets a model submit unreviewed commands, the integration can replace broad verbs with narrow parameterized operations and validation; whereas the software platform case must guard against the outcome that overlapping names lead to arbitrary selection and inconsistent answers.
C:For a workload where a complex connector ecosystem cannot fit full tool descriptions in each request, the integration can advertise compact discovery metadata and load detailed definitions on demand; the stated requirement remains unmet: the integration can retire or consolidate redundant capabilities with a responsible owner.
D:For a workload where a research agent is offered dozens of irrelevant mutation tools, the integration can expose only tools needed for the active task and user permission scope; whereas the software platform case must guard against the outcome that overlapping names lead to arbitrary selection and inconsistent answers.
E:For a workload where an assistant must read audit events but rarely modify audit settings, the integration can expose read-only access independently from privileged changes; this rollout still needs the following guarantee: the integration can retire or consolidate redundant capabilities with a responsible owner.
Question 78
A media-licensing service proposes removing Tail-latency measurement because its mean response duration looks acceptable. Which customer-visible delay risk could be overlooked?
- The danger that median latency can hide the failures users notice most
- The danger that accuracy on answered cases alone can reward refusing everything
- The danger that a successful business task may still leak restricted records
- The danger that generic satisfaction scores may reward confident errors
Explanation
Correct Answer
A:This is the direct residual hazard: median latency can hide the failures users notice most. The omitted intervention matters because the evaluation process can track p95 and p99 end-to-end latency by request class. A reviewer should demand correlated request timing across dependencies and measure high-percentile latency and timeout frequency before treating the risk as controlled.
Incorrect Answers
B:The competing hazard is that accuracy on answered cases alone can reward refusing everything in a workload where incomplete evidence should yield escalation rather than invented certainty; whereas removing Tail-latency measurement primarily exposes the risk that median latency can hide the failures users notice most.
C:The other exposure arises when agent integration must respect access restrictions for multiple tenants: a successful business task may still leak restricted records; but this is not the immediate loss associated with omitting Tail-latency measurement: median latency can hide the failures users notice most.
D:The other exposure arises when executives see fluent answers but cannot judge operational correctness: generic satisfaction scores may reward confident errors; however, the named safeguard addresses the distinct hazard that median latency can hide the failures users notice most.
Question 79
A financial operations product engineering team is designing a Claude-supported workflow because coding agents request shell access beyond the repository task. Which architectural decision most directly meets that requirement?
- Implement Repository-level project instructions
- Implement Reproducible development environment
- Implement Least-privilege tool permissions
- Implement Workspace separation for sensitive tasks
- Implement Hook-based policy checks
Explanation
Correct Answer
C:The architectural distinction is that the engineering workflow can configure approved actions and deny risky operations by default. The decision should be implemented with team permission policy and escalation route and evaluated against blocked unsafe command attempts in controlled tests; otherwise, unrestricted execution can alter secrets or infrastructure accidentally.
Incorrect Answers
A:For a workload where developers get inconsistent coding behavior across branches and members, the engineering workflow can version shared Claude Code guidance alongside the project and review changes; that mechanism leaves another need unresolved: the engineering workflow can configure approved actions and deny risky operations by default.
B:When agent-assisted fixes work on one laptop but fail in CI, the engineering workflow can pin dependencies, tooling and environment assumptions across contributors; whereas the financial operations case must guard against the outcome that unrestricted execution can alter secrets or infrastructure accidentally.
D:When confidential projects share agent context and local tool access inadvertently, the engineering workflow can isolate project roots, secrets and tool permissions by approved workspace; that mechanism leaves another need unresolved: the engineering workflow can configure approved actions and deny risky operations by default.
E:When code generation sometimes bypasses mandatory lint and secret scans, the engineering workflow can run deterministic checks at controlled workflow boundaries; whereas the financial operations case must guard against the outcome that unrestricted execution can alter secrets or infrastructure accidentally.
Question 80
Following a regression in legal operations, the team establishes that automating the last step could bypass required accountability. Which proposed redesign most directly prevents recurrence?
- Introduce Human-approved recommendation service
- Introduce Baseline workflow benchmarking
- Introduce Outcome-first use-case framing
- Introduce Phased value hypothesis testing
Explanation
Correct Answer
A:The reported failure calls for the control under which Claude drafts ranked recommendations while a designated reviewer authorizes each action. A prompt-only reminder or unrelated metric would not address the risk that automating the last step could bypass required accountability. Evidence of resolution requires a documented decision authority and approval queue and an improvement in reviewer override and acceptance rates by decision category.
Incorrect Answers
B:Where executives claim the assistant will save time without measuring current work, measuring existing effort establishes a credible incremental-value estimate; whereas the legal operations defect follows the path in which automating the last step could bypass required accountability.
C:Although the alternative reduces the risk that feature demonstrations can obscure the absence of an accountable success condition because a measurable business outcome narrows the solution boundary before model selection; yet it would not correct the observed weakness: automating the last step could bypass required accountability.
D:Although the alternative reduces the risk that a portfolio of speculative features can crowd out working high-value uses because staged hypotheses tie each funding tranche to measured evidence; but this remedy operates on another boundary, not the case where automating the last step could bypass required accountability.