Question 160
A release must prove that pure semantic search misses precise part identifiers will not recur. The application is used in this situation: Product catalog queries mix synonyms and exact serial codes. Which acceptance test most directly checks that failure mode?
- Try off-script wording and verify the mandatory sequence
- Simulate failed verification and confirm no notification
- Query both model codes and paraphrased product descriptions
- Test scans with tables, handwriting and missing labels
Explanation
Correct Answer
C:Querying both model codes and paraphrased product descriptions is aligned with the failure that pure semantic search misses precise part identifiers; its causal relevance is that hybrid ranking combines lexical precision and semantic relevance for the situation where product catalog queries mix synonyms and exact serial codes.
Incorrect Answers
A:Although the exercise might catch generative routing skips a mandatory disclosure step, it would not directly validate the mechanism that hybrid ranking combines lexical precision and semantic relevance in the target implementation.
B:Although the exercise might catch an unconstrained planner sends a notification before authorization, it would not directly validate the mechanism that hybrid ranking combines lexical precision and semantic relevance in the target implementation.
D:For vendors send variable-layout invoices and shipping documents, the alternate test is about whether document processing models extract structured fields from document content; it cannot prove the outcome needed after the reported regression where pure semantic search misses precise part identifiers.
Question 161
An agent rollout is intended for this situation: Knowledge retrieval is temporarily unavailable. The known problem is that the agent invents answers rather than saying evidence is missing. Which success measure connects to that specific outcome?
- Unauthorized grounding documents included in answers
- Duplicate mutations per concurrent event batch
- Safe failure responses under dependency loss
- Unauthorized action rate in adversarial evaluations
Explanation
Correct Answer
C:Because failure injection checks controlled degradation and abstention, the outcome worth following is safe failure responses under dependency loss; the metric is tied to the reported issue—the agent invents answers rather than saying evidence is missing—for the scenario where knowledge retrieval is temporarily unavailable.
Incorrect Answers
A:The proposed metric concerns users have different rights to the same knowledge corpus, whereas the failure in this case is that the agent invents answers rather than saying evidence is missing; the indicator needed here is safe failure responses under dependency loss.
B:Several events refer to one business order at once; its intended measurement is duplicate mutations per concurrent event batch; substituting that result here would not capture whether the agent invents answers rather than saying evidence is missing has been corrected.
D:The proposed metric concerns benchmarks cover only cooperative user wording, whereas the failure in this case is that the agent invents answers rather than saying evidence is missing; the indicator needed here is safe failure responses under dependency loss.
Question 162
A quality engineer is writing a regression case after the incident where the agent repeats an unhelpful default message. The underlying business scenario is: Users use novel requests outside the supported topic inventory. What should the test exercise?
- Install the agent to a pilot Microsoft 365 audience
- Simulate failed verification and confirm no notification
- Test ambiguous requests with incomplete required inputs
- Submit an unsupported question and verify safe routing
- Run representative notes and validate schema and exclusions
Explanation
Correct Answer
D:The agent repeats an unhelpful default message; a release test must establish that this no longer occurs; submitting an unsupported question and verifying safe routing measures the specific behavior for which fallback handles unmatched intent and can ask for clarification or escalate.
Incorrect Answers
A:In the alternate setting where employees want curated assistance without owning a separate LLM runtime, the test exercises declarative agents define instructions, knowledge and actions in the supported host; even a pass would not show that the agent repeats an unhelpful default message has been prevented.
B:Although the exercise might catch an unconstrained planner sends a notification before authorization, it would not directly validate the mechanism that fallback handles unmatched intent and can ask for clarification or escalate in the target implementation.
C:This check exercises a different system behavior: specific tool metadata helps the planner choose and parameterize actions; a pass there cannot establish that the agent repeats an unhelpful default message no longer occurs in the required workflow.
E:This check exercises a different system behavior: prompt actions encapsulate reusable inputs and expected outputs; a pass there cannot establish that the agent repeats an unhelpful default message no longer occurs in the required workflow.
Question 163
Operations has not established the cause of this defect: An agent updates the wrong pipeline field from an informal phrase. The associated setting is: Sales users say deal stage differently across regions. Which inspection should precede remediation?
- Inspect match decisions and false-positive entity joins
- Inspect entity-to-term mappings and regional synonym lists
- Compare source timestamps and principal mapping on connector calls
- Validate indexed records and source ACL mappings
Explanation
Correct Answer
B:For the case where sales users say deal stage differently across regions, the diagnostic path must explain the failure that an agent updates the wrong pipeline field from an informal phrase; inspecting entity-to-term mappings and regional synonym lists targets the mechanism by which canonical business terminology maps language to underlying Dynamics records.
Incorrect Answers
A:Two systems share company names but not primary keys; its matching diagnostic check is reasonable for Copilot attaches one customer's risk note to another account; the present investigation needs evidence for the different failure where an agent updates the wrong pipeline field from an informal phrase.
C:Sellers need current credit exposure held outside CRM; its matching diagnostic check is reasonable for Copilot relies on last quarter's manually imported account report; the present investigation needs evidence for the different failure where an agent updates the wrong pipeline field from an informal phrase.
D:For the failure that Copilot cannot find approved wiki pages by natural language, validating indexed records and source ACL mappings is relevant; this question instead reports that an agent updates the wrong pipeline field from an informal phrase, so different evidence is needed.
Question 164
Troubleshooting starts from a business complaint: Requests fail despite another approved deployment being available. The affected process is: Peak load can saturate a primary model endpoint. Which investigation is most directly tied to the reported defect?
- Compare retrieved versions against the policy registry
- Compare host limitations against the requirements matrix
- Compare metered usage with adoption and concurrency assumptions
- Correlate throttling errors with fallback activation
Explanation
Correct Answer
D:For the situation where peak load can saturate a primary model endpoint, the defect is that requests fail despite another approved deployment being available; correlating throttling errors with fallback activation helps verify the cause-and-effect path: health-aware fallback uses a compatible approved model when needed.
Incorrect Answers
A:In the process where policy documents contain draft, superseded, and currently effective revisions, the diagnostic procedure could explain the agent cites a superseded rule despite a newer policy existing; this particular incident instead turns on whether health-aware fallback uses a compatible approved model when needed and why requests fail despite another approved deployment being available.
B:The evidence this method produces applies to a workflow where a workflow needs proprietary scheduling, model choice and runtime control; it cannot by itself establish why requests fail despite another approved deployment being available in the current scenario.
C:The proposed inspection is associated with an agent scales from a trial to enterprise usage, where workload forecasts include model usage, connectors, storage and observability; it would not expose the failure mode that requests fail despite another approved deployment being available with the required specificity.
Question 165
Business leaders reject an activity-only metric for this process: Search spans internal manuals with different owners and permissions. They need to detect the issue where a model retrieves a draft manual without any accountable owner. Which performance indicator fits?
- Correct citations to accessible policy material
- Share of grounding records with complete governance metadata
- Citations referencing in-force policy versions
- Percentage of agent metrics matching reconciled source calculations
Explanation
Correct Answer
B:Share of grounding records with complete governance metadata; this outcome measure belongs to the workflow where search spans internal manuals with different owners and permissions; a change in it can help detect whether the defect that a model retrieves a draft manual without any accountable owner is resolved.
Incorrect Answers
A:The alternative tracks correct citations to accessible policy material, arising from the mechanism where supported SharePoint grounding retrieves organizational content within access controls; its movement alone says little about the failure that a model retrieves a draft manual without any accountable owner.
C:For the workflow where policy documents contain draft, superseded, and currently effective revisions, citations referencing in-force policy versions is an actionable measure; it does not establish whether a model retrieves a draft manual without any accountable owner in the situation under review.
D:The proposed metric concerns sales and finance agents must agree on revenue and margin calculations, whereas the failure in this case is that a model retrieves a draft manual without any accountable owner; the indicator needed here is share of grounding records with complete governance metadata.
Question 166
Operations has not established the cause of this defect: Copilot misclassifies a complaint as an ordinary information request. The associated setting is: Agents must distinguish an incident, complaint, inquiry and entitlement. Which inspection should precede remediation?
- Verify virtual entity enablement and row-level access
- Inspect plugin capabilities and authentication configuration
- Review confusion pairs against the customer-service glossary
- Trace latency distributions and retries for connector calls
Explanation
Correct Answer
C:In the setting where agents must distinguish an incident, complaint, inquiry and entitlement, reviewing confusion pairs against the customer-service glossary can isolate the failure in which Copilot misclassifies a complaint as an ordinary information request, because business-term definitions align user vocabulary with governed service entities.
Incorrect Answers
A:In the process where finance users need current operational data in contextual help, the diagnostic procedure could explain a static knowledge upload gives stale purchase-order states; this particular incident instead turns on whether business-term definitions align user vocabulary with governed service entities and why Copilot misclassifies a complaint as an ordinary information request.
B:For the failure that read-only knowledge alone cannot create a service request, inspecting plugin capabilities and authentication configuration is relevant; this question instead reports that Copilot misclassifies a complaint as an ordinary information request, so different evidence is needed.
D:That probe evaluates bounded timeouts and graceful fallback preserve essential user work in a workflow where external credit service responds slowly during regional outages; the reported failure concerns the path for which business-term definitions align user vocabulary with governed service entities instead.
Question 167
An agent rollout is intended for this situation: Data scientists cannot explain why a newer classifier behaves differently. The known problem is that models move to production without an audit of training inputs. Which success measure connects to that specific outcome?
- Model accuracy under agreed input-quality stress levels
- Deployed models with complete registry evidence
- True performance drift detected before critical SLO breach
- Time needed to recover a known-good model/data pair
Explanation
Correct Answer
B:For the business use where data scientists cannot explain why a newer classifier behaves differently, the result to quantify is deployed models with complete registry evidence; it connects measurement to the mechanism where a model registry ties versions to lineage, evaluations and approval, not merely to agent traffic.
Incorrect Answers
A:Model accuracy under agreed input-quality stress levels; that indicator describes a distinct workflow where production documents differ from clean training samples; it does not quantify deployed models with complete registry evidence, which is needed for the stated case.
C:The alternative tracks true performance drift detected before critical SLO breach, arising from the mechanism where drift monitoring and outcomes justify retraining when performance shifts; its movement alone says little about the failure that models move to production without an audit of training inputs.
D:The alternative tracks time needed to recover a known-good model/data pair, arising from the mechanism where versioned data transformations and model artifacts enable consistent rollback; its movement alone says little about the failure that models move to production without an audit of training inputs.
Question 168
The architecture board has evidence that separate assistants report contradictory totals for the same period. The affected situation is: Sales and finance agents must agree on revenue and margin calculations. What change most directly establishes the missing behavior?
- Create a curated semantic layer for business metrics
- Use event-driven integration for finance-to-supply-chain exceptions
- Expose governed operational entities through Dataverse
- Establish a cross-functional AI Center of Excellence playbook
Explanation
Correct Answer
A:The defect is that separate assistants report contradictory totals for the same period; the expected business case is that sales and finance agents must agree on revenue and margin calculations; the design fits because shared metric definitions and lineage prevent model-dependent arithmetic.
Incorrect Answers
B:The alternative addresses delayed reconciliation hides material fulfillment blockers by a design in which business events and durable orchestration propagate changes asynchronously; it does not supply the missing relationship for sales and finance agents must agree on revenue and margin calculations, where separate assistants report contradictory totals for the same period.
C:Different solutions apply incompatible status and customer fields; the alternate design can address that situation because Dataverse tables and governed relationships provide a reusable business-data surface, whereas this question needs a solution to separate assistants report contradictory totals for the same period.
D:One division releases agents without risk review; the alternate design can address that situation because a CoE publishes patterns, reusable controls, standards and escalation paths, whereas this question needs a solution to separate assistants report contradictory totals for the same period.
Question 169
In a business situation where a retraining release changes the tokenizer and label order, a previous build showed this defect: Downstream apps interpret correct predictions as the wrong category. Which reproducible test belongs in the release gate?
- Test unseen invoices with ambiguous tax fields
- Rebuild a test index and switch traffic under observation
- Re-run a sampled historical release on its frozen corpus
- Invoke both versions with boundary category samples
Explanation
Correct Answer
D:The check is targeted rather than generic: bundled preprocessing, model and label maps maintain contract consistency; its pass/fail evidence applies to the workflow where a retraining release changes the tokenizer and label order and to the prior defect that downstream apps interpret correct predictions as the wrong category.
Incorrect Answers
A:The tested behavior is that stratified field metrics protect high-impact values; the necessary evidence here concerns bundled preprocessing, model and label maps maintain contract consistency and the previously observed downstream apps interpret correct predictions as the wrong category.
B:The tested behavior is that explicit index versions support compatible retrieval and rollback; the necessary evidence here concerns bundled preprocessing, model and label maps maintain contract consistency and the previously observed downstream apps interpret correct predictions as the wrong category.
C:This check exercises a different system behavior: versioned evaluation sets, model IDs and thresholds document release claims; a pass there cannot establish that downstream apps interpret correct predictions as the wrong category no longer occurs in the required workflow.
Question 170
A pilot scorecard needs evidence of business improvement rather than only invocation counts. The use case is: Some cases need cross-document reasoning and detailed justification. Known defect: Small-model outputs omit critical contractual exceptions. Which KPI should be selected?
- High-complexity case pass rate by model route
- Unsupported-host requirements met by the custom design
- New-policy answer accuracy after ingestion
- Accuracy-adjusted cost per completed business task
Explanation
Correct Answer
A:Because complexity classification selects a capable model for difficult tasks, the outcome worth following is high-complexity case pass rate by model route; the metric is tied to the reported issue—small-model outputs omit critical contractual exceptions—for the scenario where some cases need cross-document reasoning and detailed justification.
Incorrect Answers
B:Unsupported-host requirements met by the custom design; that indicator describes a distinct workflow where a workflow needs proprietary scheduling, model choice and runtime control; it does not quantify high-complexity case pass rate by model route, which is needed for the stated case.
C:The premise of this indicator is retrieval separates volatile knowledge from model parameters; the architecture being evaluated instead serves some cases need cross-document reasoning and detailed justification, and its success is measured by high-complexity case pass rate by model route.
D:For the workflow where real task mix differs from prelaunch performance tests, accuracy-adjusted cost per completed business task is an actionable measure; it does not establish whether small-model outputs omit critical contractual exceptions in the situation under review.