Question 290
During solution discovery in media licensing, generated citations may be fluent yet unsupported by source passages. Which design response addresses the stated constraint most directly?
- Tool-chain compromise assessment
- Hallucination impact assessment
- Automation bias countermeasure
- Model behavior drift surveillance
Explanation
Correct Answer
B:The mechanism is specific: the control layer can assess error severity and enforce evidence or abstention at decision points. A credible implementation requires gold answers and claim-verification criteria, while its acceptance test monitors unsupported claims weighted by business severity. The alternative approaches leave the risk that fluency is not a reliability guarantee for factual claims.
Incorrect Answers
A:When multiple connectors relay credentials and untrusted results, the control layer can identify propagation paths for compromised tokens and manipulated outputs; whereas the media licensing case must guard against the outcome that fluency is not a reliability guarantee for factual claims.
C:If reviewers rubber-stamp model recommendations under time pressure, the control layer can design independent checks and reason-coded overrides for high-stakes cases; the alternative cannot replace this key requirement: the control layer can assess error severity and enforce evidence or abstention at decision points.
D:When new prompt versions or changing input mix alter policy compliance, the control layer can establish post-release monitoring and re-evaluation triggers; whereas the media licensing case must guard against the outcome that fluency is not a reliability guarantee for factual claims.
Question 291
A research administration service needs an implementation in which the evaluation process can add attacker-controlled content and escalation attempts to evaluation. Select THREE complementary elements: the design intervention, its prerequisite, and the evidence needed to test it. Choose THREE.
- Implement Held-out evaluation split
- Track successful unauthorized instruction rate under injected content
- Establish threat model and approved red-team payloads
- Establish reviewer rubric and adjudication procedure
- Track version-correct answers after policy refresh
- Implement Adversarial injection fixtures
Explanation
Correct Answers
B:This is an acceptance signal rather than another design control: successful unauthorized instruction rate under injected content. Evaluating that signal on representative tasks tests whether the evaluation process can add attacker-controlled content and escalation attempts to evaluation and identifies cases where normal benign examples do not exercise trust boundaries.
C:Before rollout, the team needs threat model and approved red-team payloads. This is a deployment dependency for the intervention whose effect is that the evaluation process can add attacker-controlled content and escalation attempts to evaluation, not an alternative metric or unrelated design feature.
F:Implementing the intervention changes the operational mechanism so that the evaluation process can add attacker-controlled content and escalation attempts to evaluation. The failure to prevent is that normal benign examples do not exercise trust boundaries; configuration and validation remain separate release obligations.
Incorrect Answers
A:This design choice is aimed at cases where prompt engineers tune repeatedly on the same example corpus; it is effective because the evaluation process can reserve untouched cases for final comparison after iteration. The rollout described is instead governed by the case where a tool agent ingests external documents that can contain hostile instructions and requires a different primary intervention; however, the intended intervention requires that the evaluation process can add attacker-controlled content and escalation attempts to evaluation.
D:The alternative setup enables a different mechanism: the evaluation process can collect independent reviewer scores without revealing model or variant; whereas this intervention requires threat model and approved red-team payloads as its operating precondition.
E:This test would validate a different benefit: the evaluation process can test on later source changes to measure freshness and drift response; but that indicator does not verify successful unauthorized instruction rate under injected content for the selected architecture.
Question 292
A research administration audit challenges the need for Explicit intermediate state store. Which risk would remain unmanaged if its function were not provided elsewhere?
- An exposure in which undifferentiated feedback cannot guide targeted changes
- An exposure in which relying on conversation text alone can duplicate side effects after failure
- An exposure in which a network timeout can trigger repeated external state changes
- An exposure in which response-only logs cannot explain the path to an outcome
- An exposure in which unvalidated inputs propagate ambiguity into generated actions
Explanation
Correct Answer
B:Removing the safeguard exposes a failure in which relying on conversation text alone can duplicate side effects after failure. The architecture normally counters this because durable state records completed stages and enables safe continuation. A compensating design would require stage IDs, persistence design and replay policy and evidence from recovered workflows without repeated write operations.
Incorrect Answers
A:The other exposure arises when reviewers only click thumbs-up or thumbs-down without reason codes: undifferentiated feedback cannot guide targeted changes. Omitting the requested control instead leaves the more immediate risk that relying on conversation text alone can duplicate side effects after failure; whereas removing Explicit intermediate state store primarily exposes the risk that relying on conversation text alone can duplicate side effects after failure.
C:The competing hazard is that a network timeout can trigger repeated external state changes in a workload where a retrying agent sometimes creates the same support task twice; however, the named safeguard addresses the distinct hazard that relying on conversation text alone can duplicate side effects after failure.
D:The competing hazard is that response-only logs cannot explain the path to an outcome in a workload where compliance teams cannot reconstruct which tool outputs informed a recommendation; whereas removing Explicit intermediate state store primarily exposes the risk that relying on conversation text alone can duplicate side effects after failure.
E:If incoming tickets have missing identifiers and mixed attachment types, unvalidated inputs propagate ambiguity into generated actions; but this is not the immediate loss associated with omitting Explicit intermediate state store: relying on conversation text alone can duplicate side effects after failure.
Question 293
A manufacturing quality release candidate fails review after assessors observe that trusting an inbound URL lets attackers forge operational events. What intervention should take priority?
- Enforce Tenant-bound retrieval filtering
- Enforce OAuth consent and delegated scopes
- Enforce Short-lived scoped service credentials
- Enforce Authenticated webhook processing
- Enforce Per-user delegated authorization
Explanation
Correct Answer
D:This intervention addresses the failure at its source because the integration can verify message authenticity, timestamp and replay protection before accepting commands. To keep the issue from recurring, establish signing keys and bounded timestamp tolerance and verify forged and replayed webhook rejection rate; otherwise trusting an inbound URL lets attackers forge operational events can persist.
Incorrect Answers
A:Although the alternative reduces the risk that post-generation redaction cannot undo a cross-tenant retrieval leak because the integration can apply tenant and document ACL filters before results reach the model; but this remedy operates on another boundary, not the case where trusting an inbound URL lets attackers forge operational events.
B:This addresses the failure where administrator-wide credentials bypass user-controlled consent semantics, because the integration can obtain user consent for minimum scopes and handle revocation; whereas the manufacturing quality defect follows the path in which trusting an inbound URL lets attackers forge operational events.
C:Although the alternative reduces the risk that long-lived broad secrets amplify compromise impact because the integration can issue narrowly scoped renewable credentials for the job identity; yet it would not correct the observed weakness: trusting an inbound URL lets attackers forge operational events.
E:When a shared service token can reveal records across users, the integration can enforce each call using the requesting user's authorized scope; whereas the manufacturing quality defect follows the path in which trusting an inbound URL lets attackers forge operational events.
Question 294
The defining problem for an internal audit pilot is that code generation sometimes bypasses mandatory lint and secret scans. Which proposed approach addresses it most precisely?
- Deploy Workspace separation for sensitive tasks
- Deploy Hook-based policy checks
- Deploy Shared MCP connector governance
- Deploy Reproducible development environment
Explanation
Correct Answer
B:The architectural distinction is that the engineering workflow can run deterministic checks at controlled workflow boundaries. The decision should be implemented with authorized hooks and maintained check commands and evaluated against blocked noncompliant changes before merge; otherwise, natural-language reminders cannot substitute for enforceable test execution.
Incorrect Answers
A:For a workload where confidential projects share agent context and local tool access inadvertently, the engineering workflow can isolate project roots, secrets and tool permissions by approved workspace; the stated scenario depends on a different mechanism: the engineering workflow can run deterministic checks at controlled workflow boundaries.
C:For a workload where development teams connect tools with overlapping credentials and contracts, the engineering workflow can register reviewed MCP servers with clear ownership and allowed scopes; whereas the enterprise knowledge management case must guard against the outcome that natural-language reminders cannot substitute for enforceable test execution.
D:If agent-assisted fixes work on one laptop but fail in CI, the engineering workflow can pin dependencies, tooling and environment assumptions across contributors; the stated scenario depends on a different mechanism: the engineering workflow can run deterministic checks at controlled workflow boundaries.
Question 295
The manufacturing quality staging review finds that hand-edited prompts drift across teams and environments. Which design intervention is the most appropriate corrective action?
- Prioritize Parameterized prompt template
- Prioritize Prompt version pinning
- Prioritize Role and instruction separation
- Prioritize Explicit schema and refusal conditions
Explanation
Correct Answer
A:The reported failure calls for the control under which the model configuration can keep stable instructions versioned and inject only validated task variables. A prompt-only reminder or unrelated metric would not address the risk that hand-edited prompts drift across teams and environments. Evidence of resolution requires template registry and variable validation and an improvement in output consistency across template versions and inputs.
Incorrect Answers
B:When an untracked hot edit invalidates comparison with prior test results, the model configuration can pin and trace prompt versions through release and rollback; but this remedy operates on another boundary, not the case where hand-edited prompts drift across teams and environments.
C:This addresses the failure where promoting document content into authority enables prompt injection, because the model configuration can place durable task rules in trusted instruction channels and treat retrieved text as data; whereas the manufacturing quality defect follows the path in which hand-edited prompts drift across teams and environments.
D:Where the caller needs a machine-readable response even when evidence is insufficient, the model configuration can define output fields and abstention behavior for unanswerable inputs; yet it would not correct the observed weakness: hand-edited prompts drift across teams and environments.
Question 296
The manufacturing quality go-live board accepts the premise that the delivery team can present candidate designs on a common Pareto comparison. Which preparation is indispensable before it can rely on Cost-latency-quality tradeoff chart operationally?
- Confirm dated assumption log and validation plan
- Confirm common risk taxonomy and mitigation evidence
- Confirm controlled performance benchmark and cost attribution
- Confirm cost categories and migration work breakdown
Explanation
Correct Answer
C:The proposed design needs controlled performance benchmark and cost attribution as its foundation. This matters because the delivery team can present candidate designs on a common Pareto comparison; a plausible demonstration without the prerequisite can still leave a case where presenting only token price hides outcome quality and downstream rework. The relevant check is stakeholder comprehension of total cost and quality tradeoffs.
Incorrect Answers
A:For the separate challenge where a design assumes source quality and permissions that have not been confirmed, the delivery team can track assumptions, validation owner and consequences of failure; but the production dependency for Cost-latency-quality tradeoff chart is controlled performance benchmark and cost attribution.
B:The setup is useful to enable the alternative's behavior where the delivery team can explain residual risk, enforceable mitigations and explicit escalation boundaries and a sponsor wants autonomy while the risk team wants control of irreversible actions; for this release, the team must first verify controlled performance benchmark and cost attribution.
D:Where a model switch looks cheap but requires extensive template and integration rework, the delivery team can include maintenance, migration, vendor dependencies and personnel effort in comparison; but the production dependency for Cost-latency-quality tradeoff chart is controlled performance benchmark and cost attribution.
Question 297
A contract administration team is advancing Hallucination impact assessment beyond its proof of concept. Engineers expect that the control layer can assess error severity and enforce evidence or abstention at decision points. What evidence of setup readiness should be required first?
- Confirm gold answers and claim-verification criteria
- Confirm versioning, drift indicators and verified review samples
- Confirm uncertainty rubric and explanation requirements
- Confirm training, reasonable workload and escalation authority
Explanation
Correct Answer
A:Readiness depends on gold answers and claim-verification criteria, not just a design approval. That condition supports the mechanism in which the control layer can assess error severity and enforce evidence or abstention at decision points. Without it, fluency is not a reliability guarantee for factual claims; monitoring unsupported claims weighted by business severity helps verify the deployed result.
Incorrect Answers
B:This prerequisite supports a method used when new prompt versions or changing input mix alter policy compliance, because the control layer can establish post-release monitoring and re-evaluation triggers; however, the named intervention needs gold answers and claim-verification criteria before its protection can be enforced.
C:The setup is useful to enable the alternative's behavior where the control layer can require source grounding and explicit uncertainty when proof is unavailable and an assistant states confident conclusions without revealing evidence limitations; but the production dependency for Hallucination impact assessment is gold answers and claim-verification criteria.
D:The setup is useful to enable the alternative's behavior where the control layer can design independent checks and reason-coded overrides for high-stakes cases and reviewers rubber-stamp model recommendations under time pressure; for this release, the team must first verify gold answers and claim-verification criteria.
Question 298
A research administration delivery team is designing a Claude-supported workflow because two outputs have subtle differences in usefulness and tone. Which architectural decision most directly meets that requirement?
- Metamorphic consistency tests
- Held-out evaluation split
- Adversarial injection fixtures
- Human blinded rating panel
- Stratified production-like test set
Explanation
Correct Answer
D:The intervention acts on the system behavior directly: the evaluation process can collect independent reviewer scores without revealing model or variant. Before rollout, verify approved enabling configuration and accountable owner. The relevant success test is inter-rater agreement and blinded quality preference; otherwise knowing model identity can bias judgments of equivalent outputs.
Incorrect Answers
A:When answers should not change when irrelevant formatting or names change, the evaluation process can apply controlled input transformations and compare invariant outcomes; it does not ensure the required behavior: the evaluation process can collect independent reviewer scores without revealing model or variant.
B:When prompt engineers tune repeatedly on the same example corpus, the evaluation process can reserve untouched cases for final comparison after iteration; whereas the research administration case must guard against the outcome that knowing model identity can bias judgments of equivalent outputs.
C:For a workload where a tool agent ingests external documents that can contain hostile instructions, the evaluation process can add attacker-controlled content and escalation attempts to evaluation; the alternative leaves the central requirement unproven: the evaluation process can collect independent reviewer scores without revealing model or variant.
E:When long-tail failure classes rarely appear in a random convenience sample, the evaluation process can stratify by task type, risk, language and source complexity; whereas the research administration case must guard against the outcome that knowing model identity can bias judgments of equivalent outputs.
Question 299
A procurement services architecture group is designing a Claude-supported workflow because downstream software requires machine-readable dispositions with fixed enums. Which architectural decision most directly meets that requirement?
- Adopt Idempotent action boundary
- Adopt Audit-correlated processing stages
- Adopt Feedback taxonomy with ground-truth capture
- Adopt Typed output validation and repair route
Explanation
Correct Answer
D:The architectural distinction is that the architecture can validate generated payloads against a contract and route failures to repair or review. The decision should be implemented with an agreed output schema and invalid-output handler and evaluated against schema-conformance rate and manual fallback count; otherwise, accepting plausible-looking prose can break consumers or misroute cases.
Incorrect Answers
A:If a retrying agent sometimes creates the same support task twice, idempotency keys and deduplication protect side-effecting writes across retries; whereas the procurement services case must guard against the outcome that accepting plausible-looking prose can break consumers or misroute cases.
B:When compliance teams cannot reconstruct which tool outputs informed a recommendation, stage-level correlation links input versions, retrieval results, model calls and final decisions; the stated scenario depends on a different mechanism: the architecture can validate generated payloads against a contract and route failures to repair or review.
C:When reviewers only click thumbs-up or thumbs-down without reason codes, structured error tags make future fixes attributable to retrieval, prompt or policy causes; that mechanism leaves another need unresolved: the architecture can validate generated payloads against a contract and route failures to repair or review.
Question 300
A supply-chain planning service passed its happy-path demo, and a reviewer asks to discard Authenticated webhook processing. Which gap would the remaining design fail to address?
- The possibility that post-generation redaction cannot undo a cross-tenant retrieval leak
- The possibility that administrator-wide credentials bypass user-controlled consent semantics
- The possibility that a shared service token can reveal records across users
- The possibility that trusting an inbound URL lets attackers forge operational events
Explanation
Correct Answer
D:The relevant failure mode is that trusting an inbound URL lets attackers forge operational events. The existing design can prevent it when the integration can verify message authenticity, timestamp and replay protection before accepting commands. The team needs signing keys and bounded timestamp tolerance and must monitor forged and replayed webhook rejection rate to support any claim of equivalent protection.
Incorrect Answers
A:When one vector index holds embeddings for several client organizations, post-generation redaction cannot undo a cross-tenant retrieval leak; however, the named safeguard addresses the distinct hazard that trusting an inbound URL lets attackers forge operational events.
B:If employees connect personal enterprise resources through an integration, administrator-wide credentials bypass user-controlled consent semantics; whereas removing Authenticated webhook processing primarily exposes the risk that trusting an inbound URL lets attackers forge operational events.
C:When a shared integration must reflect each employee's actual record rights, a shared service token can reveal records across users; but this is not the immediate loss associated with omitting Authenticated webhook processing: trusting an inbound URL lets attackers forge operational events.