Question 110
The acceptance board asks how to verify a corrected design. The setting is: Candidate processes are attractive but permissions and data quality are weak. Last time, the first-wave pilot requests broad access to unreliable records. Which scenario should the team execute?
- Run a representative extraction holdout at production concurrency
- Add a confidential draft and verify it is excluded for unauthorized users
- Compare prioritized roadmap with a risk-adjusted benefit matrix
- Propose a concession above limit and confirm approval enforcement
Explanation
Correct Answer
C:The first-wave pilot requests broad access to unreliable records; a release test must establish that this no longer occurs; comparing prioritized roadmap with a risk-adjusted benefit matrix measures the specific behavior for which readiness assessment places safe high-value workloads ahead of risky ones.
Incorrect Answers
A:In a solution where high-volume narrow extraction needs predictable latency and cost, running a representative extraction holdout at production concurrency could catch the defect that a general large model exceeds per-document unit economics; the current release gate must challenge the failure where the first-wave pilot requests broad access to unreliable records.
B:For search spans internal manuals with different owners and permissions, the alternate test is about whether metadata enables filtering, review, retention and traceable answers; it cannot prove the outcome needed after the reported regression where the first-wave pilot requests broad access to unreliable records.
D:Proposing a concession above limit and confirming approval enforcement verifies an adjacent objective where service, sales and finance jointly decide high-value concessions; it leaves untested the causal path in which readiness assessment places safe high-value workloads ahead of risky ones for this case.
Question 111
A design adds autonomous behavior to a process where an internal model sometimes returns confidential text verbatim. This creates a documented risk: A model might expose sensitive user data in uncontrolled prompt output. Which enforced control is the best fit?
- Prevent sensitive data from entering prompts unless required
- Reject missing, invalid or out-of-scope authorization tokens
- Require reviewed datasets and independent promotion sign-off
- Scope the identity to only required resource permissions
Explanation
Correct Answer
A:In the scenario where an internal model sometimes returns confidential text verbatim, the explicit risk is that a model might expose sensitive user data in uncontrolled prompt output; enforcement must follow the mechanism by which safety controls and authorization checks limit misuse and disclosure.
Incorrect Answers
B:The alternate boundary mitigates an API might trust model-generated identity claims instead of valid authorization, which is relevant to prompt wording persuades the agent to submit a privileged operation; it leaves the current workflow vulnerable to a model might expose sensitive user data in uncontrolled prompt output.
C:Business analysts can submit examples but not change production behavior; its guardrail addresses the possibility that unreviewed training examples might poison a promoted model; the present architecture must constrain the different threat that a model might expose sensitive user data in uncontrolled prompt output.
D:A backend agent must call an approved Azure service without embedded secrets; its guardrail addresses the possibility that long-lived credentials might grant an agent unnecessary Azure resource privileges; the present architecture must constrain the different threat that a model might expose sensitive user data in uncontrolled prompt output.
Question 112
A proposed deployment supports this business case: Agents suggest payment adjustments and inventory holds. Reviewers also warn that an updated connector might gain a direct privileged posting route. Which restriction mitigates that concrete risk?
- Require data-use approval and encrypted governed storage
- Do not approve procurement while financial hold status is unknown
- Block release if privileged writes bypass approvals
- Rotate compromised keys and enforce least-privilege callers
Explanation
Correct Answer
C:Because an updated connector might gain a direct privileged posting route, a guardrail must constrain the actual system behavior: authorization regression gates verify posting authority remains constrained; protecting a response alone cannot enforce that boundary.
Incorrect Answers
A:This restriction is associated with proprietary records are copied to an open test workspace and addresses the exposure that restricted fine-tuning documents might be readable by another project; it does not specifically protect the target workflow against an updated connector might gain a direct privileged posting route.
B:The safeguard for a blocked invoice triggers a hold on related procurement actions deals with a different threat: a procurement approval might proceed before a finance hold is resolved; even after it is enforced, the exposure that an updated connector might gain a direct privileged posting route remains possible in the target workflow where authorization regression gates verify posting authority remains constrained.
D:A sensitive model endpoint is reachable using a broadly shared key; its guardrail addresses the possibility that a leaked endpoint credential might permit inference from an untrusted client; the present architecture must constrain the different threat that an updated connector might gain a direct privileged posting route.
Question 113
A release must prove that a client can request arbitrary portions of training examples will not recur. The application is used in this situation: An internal model sometimes returns confidential text verbatim. Which acceptance test most directly checks that failure mode?
- Make identical queries as permitted and denied users
- Test extraction and secret-disclosure prompts with synthetic fixtures
- Try a forbidden refund despite a permissive model instruction
- Submit a malicious mislabeled training record and inspect review
Explanation
Correct Answer
B:For the required workflow where an internal model sometimes returns confidential text verbatim, the test needs to reproduce or prevent the failure where a client can request arbitrary portions of training examples; testing extraction and secret-disclosure prompts with synthetic fixtures exercises the path where safety controls and authorization checks limit misuse and disclosure.
Incorrect Answers
A:In a solution where shared result caching makes answers fast across roles, making identical queries as permitted and denied users could catch the defect that one user's cached confidential answer appears for another user; the current release gate must challenge the failure where a client can request arbitrary portions of training examples.
C:Although the exercise might catch API trusts the model's text assertion of permission, it would not directly validate the mechanism that safety controls and authorization checks limit misuse and disclosure in the target implementation.
D:For business analysts can submit examples but not change production behavior, the alternate test is about whether separated data-curation and release permissions preserve integrity; it cannot prove the outcome needed after the reported regression where a client can request arbitrary portions of training examples.
Question 114
Release reviewers will not accept a generic success message. The business scenario is that a new model reduces price but also degrades exact extractions, and the historical failure is: Finance sees savings without the downstream rework cost. Which test is appropriately targeted?
- Re-run a sampled historical release on its frozen corpus
- Submit a document without classification and verify it stays excluded
- Re-run the same multi-intent corpus on both configurations
- Re-evaluate extraction accuracy at forecast traffic levels
Explanation
Correct Answer
D:The check is targeted rather than generic: joint quality and consumption measures capture real efficiency; its pass/fail evidence applies to the workflow where a new model reduces price but also degrades exact extractions and to the prior defect that finance sees savings without the downstream rework cost.
Incorrect Answers
A:The suggested test serves an updated agent performs differently months after release, in which old scores cannot be reproduced because prompts and tests changed; the present scenario instead concerns the reported problem where finance sees savings without the downstream rework cost.
B:For new documents enter a production source automatically, the alternate test is about whether ingestion gates check classification, freshness and ownership; it cannot prove the outcome needed after the reported regression where finance sees savings without the downstream rework cost.
C:The verification concerning a new topic description changes which agent is invoked checks trace comparisons reveal shifts in plan selection and step order against regression appears only for compound user requests; its result says little about the observed issue where finance sees savings without the downstream rework cost.
Question 115
An investment review asks what measurable result would justify continuing an AI solution where two applications try to overwrite the same customer credit field. The original issue is: Parallel agent updates create conflicting authoritative values. Which indicator should be tracked?
- Share of grounding records with complete governance metadata
- Median elapsed time from design decision to approval
- Cross-app case-state agreement rate
- Conflicting cross-application writes per thousand transactions
Explanation
Correct Answer
D:For the use case where two applications try to overwrite the same customer credit field, the indicator conflicting cross-application writes per thousand transactions is worth baselining against the earlier issue where parallel agent updates create conflicting authoritative values.
Incorrect Answers
A:The proposed metric concerns search spans internal manuals with different owners and permissions, whereas the failure in this case is that parallel agent updates create conflicting authoritative values; the indicator needed here is conflicting cross-application writes per thousand transactions.
B:The proposed metric concerns business, platform, security and legal teams make conflicting AI decisions, whereas the failure in this case is that parallel agent updates create conflicting authoritative values; the indicator needed here is conflicting cross-application writes per thousand transactions.
C:Cross-app case-state agreement rate. That indicator describes a distinct workflow where sales and service agents need the same customer and case context; it does not quantify conflicting cross-application writes per thousand transactions, which is needed for the stated case.
Question 116
One enterprise rollout has 2 separately testable requirements. (1) Business use: Real task mix differs from prelaunch performance tests. Known defect: Router saves tokens but sends difficult tasks to weak models. (2) Business use: Many calls require only a small fixed-label output. Known defect: Universal use of a premium model inflates cost. Which 2 architectural choices directly satisfy these stated cases? Choose TWO.
- Account for capacity, consumption and monitoring in TCO
- Normalize identifiers and remove duplicated grounding records
- Route simple classification requests to a validated inexpensive model
- Use quality-and-cost telemetry to tune routing thresholds
Explanation
Correct Answers
C:For the independent requirement that many calls require only a small fixed-label output, task-aware routing chooses economical models after quality gates; it addresses the defect universal use of a premium model inflates cost, while real task mix differs from prelaunch performance tests needs its own architecture.
D:For the independent requirement that real task mix differs from prelaunch performance tests, feedback-driven thresholds rebalance accuracy, latency and spend; it addresses the defect router saves tokens but sends difficult tasks to weak models, while many calls require only a small fixed-label output needs its own architecture.
Incorrect Answers
A:This design is associated with the different failure where token and action costs exceed the pilot estimate in a workflow where an agent scales from a trial to enterprise usage; it does not directly solve the listed conditions: router saves tokens but sends difficult tasks to weak models; universal use of a premium model inflates cost.
B:This design is associated with the different failure where responses merge two unrelated customers with similar names in a workflow where multiple systems use inconsistent customer identifiers and duplicate cases; it does not directly solve the listed conditions: router saves tokens but sends difficult tasks to weak models; universal use of a premium model inflates cost.
Question 117
An investment review asks what measurable result would justify continuing an AI solution where answers depend on policies revised every week. The original issue is: A trained model remembers obsolete guidance after source changes. Which indicator should be tracked?
- Improvement in domain error rate versus approved baseline
- Cost per accepted classification at required accuracy
- New-policy answer accuracy after ingestion
- Authorized structured-record answer accuracy
- Incremental maintenance effort of the extension
Explanation
Correct Answer
C:This scorecard needs an outcome related to answers depend on policies revised every week; new-policy answer accuracy after ingestion. That indicator can be checked against the earlier symptom that a trained model remembers obsolete guidance after source changes.
Incorrect Answers
A:That scorecard concerns the activity in which a unique domain vocabulary produces persistent measured errors and its result of improvement in domain error rate versus approved baseline; the reported shortcoming here is a trained model remembers obsolete guidance after source changes.
B:A change in cost per accepted classification at required accuracy is informative for many calls require only a small fixed-label output; it does not establish whether the current process achieves new-policy answer accuracy after ingestion.
D:The proposed metric concerns users ask about governed CRM records rather than prose manuals, whereas the failure in this case is that a trained model remembers obsolete guidance after source changes; the indicator needed here is new-policy answer accuracy after ingestion.
E:That scorecard concerns the activity in which current users need one governed custom action inside their workflow and its result of incremental maintenance effort of the extension; the reported shortcoming here is a trained model remembers obsolete guidance after source changes.
Question 118
The acceptance board asks how to verify a corrected design. The setting is: Answers about product policy require traceable evidence. Last time, helpful-looking responses contain unverifiable claims. Which scenario should the team execute?
- Ask counterfactual questions and verify claims map to retrieved evidence
- Query an external article using users with different permissions
- Test an out-of-corpus question and expect a no-evidence response
- Delay a specialist and confirm timeout and retry behavior
Explanation
Correct Answer
C:Testing an out-of-corpus question and expect a no-evidence response challenges the path where cited retrieval separates factual support from free-form generation for the business setting answers about product policy require traceable evidence; the specific defect to exclude is that helpful-looking responses contain unverifiable claims.
Incorrect Answers
A:The verification concerning analysts compare unstructured reports but must retain human interpretation checks retrieval supplies supporting evidence while analysts own analytical decisions against narratives present speculation as if it were validated insight; its result says little about the observed issue where helpful-looking responses contain unverifiable claims.
B:This check exercises a different system behavior: connectors expose external enterprise content to supported Microsoft search/Copilot experiences; a pass there cannot establish that helpful-looking responses contain unverifiable claims no longer occurs in the required workflow.
D:The verification concerning downstream research agents run longer than the interactive request checks task IDs, status transitions, deadlines and durable handoff state support coordination against orchestrator returns a partial result as though the workflow finished; its result says little about the observed issue where helpful-looking responses contain unverifiable claims.
Question 119
A staging test has to detect the specific defect that engineers tune prompts without checking whether evidence was retrieved. In the workflow where an answer contradicts the handbook despite high model fluency, which exercise verifies the corrective behavior?
- Score the same question with controlled gold context
- Make identical queries as permitted and denied users
- Use an annotated query with a known authoritative paragraph
- Replay a stable request set on both prompt versions
Explanation
Correct Answer
A:Scoring the same question with controlled gold context challenges the path where separate relevance and groundedness tests isolate the failing layer for the business setting an answer contradicts the handbook despite high model fluency; the specific defect to exclude is that engineers tune prompts without checking whether evidence was retrieved.
Incorrect Answers
B:In a solution where shared result caching makes answers fast across roles, making identical queries as permitted and denied users could catch the defect that one user's cached confidential answer appears for another user; the current release gate must challenge the failure where engineers tune prompts without checking whether evidence was retrieved.
C:For an assistant responds fluently but cites unrelated support articles, the alternate test is about whether retrieval telemetry distinguishes evidence search from answer generation; it cannot prove the outcome needed after the reported regression where engineers tune prompts without checking whether evidence was retrieved.
D:Replaying a stable request set on both prompt versions verifies an adjacent objective where model bills increase although request volume is stable; it leaves untested the causal path in which separate relevance and groundedness tests isolate the failing layer for this case.
Question 120
A risk analysis highlights a threat in this workflow: Two topics claim nearly identical scheduling intents. Exposure: Overlapping scheduling intents might invoke a change before confirmation. Which enforceable safeguard addresses it at the relevant boundary?
- Disable write tools and label generated material for review
- Require confirmation when intent remains ambiguous
- Preserve conversation context and avoid infinite re-prompting
- Apply tenant policies and scoped content permissions
Explanation
Correct Answer
B:When overlapping scheduling intents might invoke a change before confirmation, the use case two topics claim nearly identical scheduling intents requires a guardrail aligned with the behavior that distinct descriptions and inputs improve topic selection.
Incorrect Answers
A:This restriction is associated with marketing specialists need tailored descriptions from existing briefings and addresses the exposure that a text-only drafting assistant might be given unnecessary record-writing powers; it does not specifically protect the target workflow against overlapping scheduling intents might invoke a change before confirmation.
C:That protection is tied to the mechanism that confidence, retry ceilings and escalation routing enable resolution; it does not prevent the distinct risk where overlapping scheduling intents might invoke a change before confirmation in the system under review.
D:The safeguard for employees want curated assistance without owning a separate LLM runtime deals with a different threat: a declarative agent might be published beyond permitted tenant groups; even after it is enforced, the exposure that overlapping scheduling intents might invoke a change before confirmation remains possible in the target workflow where distinct descriptions and inputs improve topic selection.