Question 790
Release reviewers will not accept a generic success message. The business scenario is that planner calls the wrong tool among overlapping capabilities, and the historical failure is: The team proposes adding another similar tool. Which test is appropriately targeted?
- Submit a malicious mislabeled training record and inspect review
- Classify a mixed set of complaints and check routing consistency
- Compare controlled tests before and after tool curation
- Reproduce a sanitized failure and confirm the new regression catches it
Explanation
Correct Answer
C:Comparing controlled tests before and after tool curation is aligned with the failure that the team proposes adding another similar tool; its causal relevance is that clear capabilities and restricted sets improve selection reliability for the situation where planner calls the wrong tool among overlapping capabilities.
Incorrect Answers
A:The verification concerning business analysts can submit examples but not change production behavior checks separated data-curation and release permissions preserve integrity against unvetted labeled data poisons a retrained classifier; its result says little about the observed issue where the team proposes adding another similar tool.
B:The tested behavior is that structured defect taxonomy assigns the right remediation owner; the necessary evidence here concerns clear capabilities and restricted sets improve selection reliability and the previously observed the team proposes adding another similar tool.
D:A pass on that verification would speak to support tickets reveal an unusual multistep conversation failure; it would not show that clear capabilities and restricted sets improve selection reliability, which is needed to avoid the observed defect the team proposes adding another similar tool.
Question 791
Security and business owners want to keep this capability: Callers provide account numbers and amounts by speech. They must also prevent the following: A speech misunderstanding might trigger a high-impact action for the wrong account. Which guardrail should the design incorporate?
- Store policy thresholds outside model-editable instructions
- Require identity checks and human fallback for ambiguous audio
- Enforce source-system entitlements in the integration
- Limit deployment to approved users and surfaces
Explanation
Correct Answer
B:The business setting is callers provide account numbers and amounts by speech, with the exposure that a speech misunderstanding might trigger a high-impact action for the wrong account; the enforceable protection must correspond to the path where spoken confirmation and verification reduce high-impact recognition errors.
Incorrect Answers
A:Another process could need this guardrail because a background agent might exceed spending authority by editing its own prompt; the design being assessed supports callers provide account numbers and amounts by speech and has the explicit exposure that a speech misunderstanding might trigger a high-impact action for the wrong account.
C:The safeguard for sellers need current credit exposure held outside CRM deals with a different threat: sales enrichment might query external records beyond the seller's entitled scope; even after it is enforced, the exposure that a speech misunderstanding might trigger a high-impact action for the wrong account remains possible in the target workflow where spoken confirmation and verification reduce high-impact recognition errors.
D:That protection is tied to the mechanism that host-specific support, packaging and admin approval shape deployment; it does not prevent the distinct risk where a speech misunderstanding might trigger a high-impact action for the wrong account in the system under review.
Question 792
A risk analysis highlights a threat in this workflow: Analysts compare unstructured reports but must retain human interpretation. Exposure: A narrative insight may be mistaken for permission to approve a financial decision. Which enforceable safeguard addresses it at the relevant boundary?
- Validate sending-agent identity and delegated authorization
- Promote complex or low-confidence cases to stronger fallback
- Preserve table and row access controls for retrieval
- Require human review for high-impact low-confidence classifications
- Prohibit autonomous financial decisions and present uncertainty
Explanation
Correct Answer
E:The business setting is analysts compare unstructured reports but must retain human interpretation, with the exposure that a narrative insight may be mistaken for permission to approve a financial decision; the enforceable protection must correspond to the path where retrieval supplies supporting evidence while analysts own analytical decisions.
Incorrect Answers
A:The proposed control targets an external agent might accept delegated tasks from a spoofed sender, while the given risk scenario is a narrative insight may be mistaken for permission to approve a financial decision; these exposures arise in different operational contexts.
B:Many calls require only a small fixed-label output; its guardrail addresses the possibility that a small model might make a high-complexity judgment with insufficient confidence; the present architecture must constrain the different threat that a narrative insight may be mistaken for permission to approve a financial decision.
C:The safeguard for users ask about governed CRM records rather than prose manuals deals with a different threat: Dataverse retrieval might reveal a row not authorized to the caller; even after it is enforced, the exposure that a narrative insight may be mistaken for permission to approve a financial decision remains possible in the target workflow where retrieval supplies supporting evidence while analysts own analytical decisions.
D:Where the classifier confuses complaints with refund requests, the guard can address a low-confidence label might be used for an irreversible eligibility decision; the exposure that a narrative insight may be mistaken for permission to approve a financial decision remains relevant to the present assessment.
Question 793
Business leaders reject an activity-only metric for this process: Operations teams review multiple proposed next actions daily. They need to detect the issue where recommendations disappear in conversation history. Which performance indicator fits?
- User understanding of verified versus suggested content
- Correctly completed legacy UI tasks without unsafe side effects
- Complex exception tasks resolved with supported reasoning traces
- Ticket updates completed with attributable credentials
- Tool operations executed with valid delegated identity
- Agent recommendations with recorded outcome and owner
Explanation
Correct Answer
F:The observed shortcoming is that recommendations disappear in conversation history; measuring agent recommendations with recorded outcome and owner gives business owners a directly relevant indicator for the process where operations teams review multiple proposed next actions daily.
Incorrect Answers
A:Users need to distinguish verified fields from AI suggestions; its intended measurement is user understanding of verified versus suggested content; substituting that result here would not capture whether recommendations disappear in conversation history has been corrected.
B:A change in correctly completed legacy UI tasks without unsafe side effects is informative for a legacy portal has no integration API for a clerical workflow; it does not establish whether the current process achieves agent recommendations with recorded outcome and owner.
C:For the workflow where procurement exceptions require comparing several policies and records, complex exception tasks resolved with supported reasoning traces is an actionable measure; it does not establish whether recommendations disappear in conversation history in the situation under review.
D:Ticket updates completed with attributable credentials; that indicator describes a distinct workflow where an agent needs to update a governed ticket status; it does not quantify agent recommendations with recorded outcome and owner, which is needed for the stated case.
E:Tool operations executed with valid delegated identity; that indicator describes a distinct workflow where several agents need the same inventory query operation; it does not quantify agent recommendations with recorded outcome and owner, which is needed for the stated case.
Question 794
In a business situation where knowledge retrieval is temporarily unavailable, a previous build showed this defect: The agent invents answers rather than saying evidence is missing. Which reproducible test belongs in the release gate?
- Block the search endpoint during a policy question
- Use an annotated query with a known authoritative paragraph
- Try injection, invalid inputs and out-of-scope actions
- Query identical terms as HR and non-HR users
Explanation
Correct Answer
A:For the required workflow where knowledge retrieval is temporarily unavailable, the test needs to reproduce or prevent the failure where the agent invents answers rather than saying evidence is missing; blocking the search endpoint during a policy question exercises the path where failure injection checks controlled degradation and abstention.
Incorrect Answers
B:In a solution where an assistant responds fluently but cites unrelated support articles, using an annotated query with a known authoritative paragraph could catch the defect that surface quality scores appear healthy despite wrong documents; the current release gate must challenge the failure where the agent invents answers rather than saying evidence is missing.
C:A pass on that verification would speak to benchmarks cover only cooperative user wording; it would not show that failure injection checks controlled degradation and abstention, which is needed to avoid the observed defect the agent invents answers rather than saying evidence is missing.
D:In the alternate setting where users have different rights to the same knowledge corpus, the test exercises retrieval applies effective source entitlements to candidate documents; even a pass would not show that the agent invents answers rather than saying evidence is missing has been prevented.
Question 795
During integration testing, an agent guesses a policy instead of escalating. The associated business context is: The system cannot always find evidence for regulated guidance. Which architectural remedy is the strongest match?
- Require approval before irreversible agent actions
- Define autonomy tiers by consequence and reversibility
- Route high-stakes reasoning to a stronger validated model
- Implement an abstain-or-escalate response policy for uncertain outputs
Explanation
Correct Answer
D:For the need that the system cannot always find evidence for regulated guidance, the relevant mechanism is that confidence and evidence checks route insufficiently supported cases to humans; that architectural dependency addresses the observed failure where an agent guesses a policy instead of escalating.
Incorrect Answers
A:Confirmation and human approval separate advice from execution authority; its main purpose is the agent may refund large amounts or deactivate accounts, not correction of the reported defect where an agent guesses a policy instead of escalating in the target process.
B:This design would make sense if teams propose that an agent perform both low-risk lookups and financial postings, particularly where the same agent can read an invoice and commit a payment unchecked; the requested architecture instead needs the dependency that confidence and evidence checks route insufficiently supported cases to humans because an agent guesses a policy instead of escalating.
C:In the case where some cases need cross-document reasoning and detailed justification, the alternate design depends on complexity classification selects a capable model for difficult tasks; it does not correct the target failure, an agent guesses a policy instead of escalating.
Question 796
Before introducing a fix, the team must explain why old scores cannot be reproduced because prompts and tests changed. The affected setting is: An updated agent performs differently months after release. What should be inspected?
- Compare deployed agent configuration to its approved manifest
- Review the signed acceptance matrix and results
- Compare release report hashes with preserved test artifacts
- Compare embedding versions and chunking metadata
Explanation
Correct Answer
C:The investigation directly tests the causal connection: versioned evaluation sets, model IDs and thresholds document release claims; it collects evidence for a workflow where an updated agent performs differently months after release but the outcome is instead that old scores cannot be reproduced because prompts and tests changed.
Incorrect Answers
A:Model and tool changes are promoted independently; its matching diagnostic check is reasonable for a release uses stale prompts against a new tool signature; the present investigation needs evidence for the different failure where old scores cannot be reproduced because prompts and tests changed.
B:Candidate improves one metric while violating another; examining that other path does not determine whether versioned evaluation sets, model IDs and thresholds document release claims, which is what distinguishes the issue in the stated use case.
D:A search index changes embedding model and chunking approach; its matching diagnostic check is reasonable for existing vectors are queried with incompatible assumptions; the present investigation needs evidence for the different failure where old scores cannot be reproduced because prompts and tests changed.
Question 797
Before introducing a fix, the team must explain why untrusted clients submit requests and exfiltrate outputs. The affected setting is: A sensitive model endpoint is reachable using a broadly shared key. What should be inspected?
- Trace source exception through target state and error queue
- Review backend authorization logs against actor identities
- Review endpoint access logs and allowed network paths
- Inspect cache keys and authorization on reused entries
Explanation
Correct Answer
C:Where untrusted clients submit requests and exfiltrate outputs, the useful evidence comes from reviewing endpoint access logs and allowed network paths; its focus matches the architectural mechanism that identity-based access and supported network isolation reduce exposure in the process where a sensitive model endpoint is reachable using a broadly shared key.
Incorrect Answers
A:The evidence this method produces applies to a workflow where a blocked invoice triggers a hold on related procurement actions; it cannot by itself establish why untrusted clients submit requests and exfiltrate outputs in the current scenario.
B:The proposed probe concerns prompt wording persuades the agent to submit a privileged operation, where server-side access checks enforce real actor entitlements; this incident instead requires evidence about identity-based access and supported network isolation reduce exposure and the failure that untrusted clients submit requests and exfiltrate outputs.
D:In the separate scenario where shared result caching makes answers fast across roles, the same investigation could explain one user's cached confidential answer appears for another user; it does not isolate the reported cause in a system where a sensitive model endpoint is reachable using a broadly shared key.
Question 798
During integration testing, a large context prompt quietly grows every release. The associated business context is: Model bills increase although request volume is stable. Which architectural remedy is the strongest match?
- Analyze percentile latency instead of averages alone
- Preserve model-compatible preprocessing and postprocessing contracts
- Segment token spending by use case and model route
- Tune model routing based on segmented evaluation outcomes
Explanation
Correct Answer
C:The architecture must reconcile two facts: model bills increase although request volume is stable, yet a large context prompt quietly grows every release; the selected pattern does so through the mechanism that token telemetry isolates prompt, retrieval and generation cost changes.
Incorrect Answers
A:When a small portion of calls time out during peak demand, tail percentiles surface extreme experiences affecting users; in the stated workflow, however, a large context prompt quietly grows every release; the relevant architectural behavior is that token telemetry isolates prompt, retrieval and generation cost changes.
B:The technical behavior of this alternative is that bundled preprocessing, model and label maps maintain contract consistency; the user case being assessed is model bills increase although request volume is stable with a different failure: a large context prompt quietly grows every release.
D:Lower-cost routing degrades a business-critical subgroup; the alternate design can address that situation because stratified evaluation guides thresholds for difficult requests, whereas this question needs a solution to a large context prompt quietly grows every release.
Question 799
The acceptance board asks how to verify a corrected design. The setting is: Production components should not be edited casually. Last time, a maker changes production agent logic outside the release process. Which scenario should the team execute?
- Submit urgent and ambiguous cases to the candidate agent
- Shift limited traffic to a candidate and monitor regressions
- Simulate rollback during a nonproduction close cycle
- Attempt an unapproved production modification in test governance
Explanation
Correct Answer
D:For the required workflow where production components should not be edited casually, the test needs to reproduce or prevent the failure where a maker changes production agent logic outside the release process; attempting an unapproved production modification in test governance exercises the path where managed solution deployment supports controlled downstream customization boundaries.
Incorrect Answers
A:Submitting urgent and ambiguous cases to the candidate agent verifies an adjacent objective where a support agent changes how it classifies priority intents; it leaves untested the causal path in which managed solution deployment supports controlled downstream customization boundaries for this case.
B:A pass on that verification would speak to business applications depend on the agent endpoint continuously; it would not show that managed solution deployment supports controlled downstream customization boundaries, which is needed to avoid the observed defect a maker changes production agent logic outside the release process.
C:The suggested test serves invoice posting agents affect critical close processes, in which a prompt change during close generates unexplained exceptions; the present scenario instead concerns the reported problem where a maker changes production agent logic outside the release process.
Question 800
A policy review covers an agent for which average accuracy hides failures on complex finance questions. Its most relevant exposure is that a model router change might silently degrade the highest-risk task segment. What safeguard should be required before release?
- Roll back a routing change when a critical segment regresses
- Use bounded retries and approved fallback capacity
- Require a tested human escalation path for unmatched cases
- Require reviewed datasets and independent promotion sign-off
Explanation
Correct Answer
A:In the context where average accuracy hides failures on complex finance questions, the exposure is that a model router change might silently degrade the highest-risk task segment; containing it requires a guardrail in the operation where stratified evaluation guides thresholds for difficult requests.
Incorrect Answers
B:A regional launch multiplies concurrent agent requests; its guardrail addresses the possibility that retry storms might amplify a rate-limit event into a service outage; the present architecture must constrain the different threat that a model router change might silently degrade the highest-risk task segment.
C:The alternate guard addresses a new support agent might route urgent cases to a general queue for a support agent changes how it classifies priority intents; the present threat, a model router change might silently degrade the highest-risk task segment, needs a different enforcement boundary.
D:The proposed control targets unreviewed training examples might poison a promoted model, while the given risk scenario is a model router change might silently degrade the highest-risk task segment; these exposures arise in different operational contexts.