Question 300
A staging test has to detect the specific defect that an update changes a field mapping and grounded answers become stale. In the workflow where AI assistance reads operational records through Dataverse, which exercise verifies the corrective behavior?
- Interrupt processing after the first app commits
- Execute both approved and rejected concession scenarios
- Modify a safe test purchase order and verify the mapped output
- Try accessing a restricted training corpus from an unrelated project
Explanation
Correct Answer
C:An update changes a field mapping and grounded answers become stale; a release test must establish that this no longer occurs; modifying a safe test purchase order and verifying the mapped output measures the specific behavior for which integration tests protect virtual-entity schema and source semantics.
Incorrect Answers
A:For a blocked invoice triggers a hold on related procurement actions, the alternate test is about whether failure-injection tests verify retries and compensation between systems; it cannot prove the outcome needed after the reported regression where an update changes a field mapping and grounded answers become stale.
B:This check exercises a different system behavior: transaction-path tests confirm recommendation, review and posting records; a pass there cannot establish that an update changes a field mapping and grounded answers become stale no longer occurs in the required workflow.
D:For proprietary records are copied to an open test workspace, the alternate test is about whether segregated datasets and access controls protect sensitive data; it cannot prove the outcome needed after the reported regression where an update changes a field mapping and grounded answers become stale.
Question 301
Support engineers report this failure: Previously reported errors recur in the next release. The business context is: Fixes are marked complete when a prompt is deployed. What evidence should be collected to isolate the cause?
- Compare affected cohorts and failure types before and after release
- Compare requested export destinations against policy allowlists
- Compare resolved, abandoned and escalated session cohorts
- Trace token acquisition and authorization results by environment
Explanation
Correct Answer
A:Where previously reported errors recur in the next release, the useful evidence comes from comparing affected cohorts and failure types before and after release; its focus matches the architectural mechanism that outcome verification checks whether the change fixed the user problem in the process where fixes are marked complete when a prompt is deployed.
Incorrect Answers
B:Model-level refusal alone may not prevent a tool call; examining that other path does not determine whether outcome verification checks whether the change fixed the user problem, which is what distinguishes the issue in the stated use case.
C:The evidence this method produces applies to a workflow where support leads need resolution and escalation insights over time; it cannot by itself establish why previously reported errors recur in the next release in the current scenario.
D:The evidence this method produces applies to a workflow where an imported agent can authenticate in dev but not production; it cannot by itself establish why previously reported errors recur in the next release in the current scenario.
Question 302
The fault affects an AI-powered workflow in which staff need a contextual workspace assembled around one customer case. Specifically, generic chat loses visibility into related records and actions. Which diagnostic investigation best distinguishes the relevant failure path?
- Inspect outbound payloads and role-based access decisions
- Trace agent plan for extra tool invocations
- Trace page data bindings and invoked record APIs
- Compare prompt versions used by dependent apps
Explanation
Correct Answer
C:For the situation where staff need a contextual workspace assembled around one customer case, the defect is that generic chat loses visibility into related records and actions; tracing page data bindings and invoked record APIs helps verify the cause-and-effect path: generated interface components can be grounded in governed app data and defined actions.
Incorrect Answers
A:In the process where a canvas app passes customer records to a model action, the diagnostic procedure could explain hidden fields from the UI appear in the prompt payload; this particular incident instead turns on whether generated interface components can be grounded in governed app data and defined actions and why generic chat loses visibility into related records and actions.
B:The evidence this method produces applies to a workflow where an employee requests a standard equipment return initiation; it cannot by itself establish why generic chat loses visibility into related records and actions in the current scenario.
D:For the failure that each app drifts to a different unreviewed AI prompt, comparing prompt versions used by dependent apps is relevant; this question instead reports that generic chat loses visibility into related records and actions, so different evidence is needed.
Question 303
One enterprise rollout has 3 separately testable requirements. (1) Business use: Production documents differ from clean training samples. Known defect: Scanned forms suddenly produce inaccurate predictions. (2) Business use: Teams promote a custom model because it feels better. Known defect: Candidate improves one metric while violating another. (3) Business use: A document model performs well overall but fails on tax identifiers. Known defect: A high aggregate score masks critical field errors. Which 3 diagnostic procedures directly satisfy these stated cases? Choose THREE.
- Compare errors by scan quality and source drift
- Compare denial logs and executed actions for attack samples
- Review the signed acceptance matrix and results
- Compare precision and recall by field and layout type
Explanation
Correct Answers
A:Comparing errors by scan quality and source drift can reveal why scanned forms suddenly produce inaccurate predictions when production documents differ from clean training samples; that finding is needed alongside the independent diagnosis for teams promote a custom model because it feels better; a document model performs well overall but fails on tax identifiers.
C:Reviewing the signed acceptance matrix and results can reveal why candidate improves one metric while violating another when teams promote a custom model because it feels better; that finding is needed alongside the independent diagnosis for production documents differ from clean training samples; a document model performs well overall but fails on tax identifiers.
D:Comparing precision and recall by field and layout type can reveal why a high aggregate score masks critical field errors when a document model performs well overall but fails on tax identifiers; that finding is needed alongside the independent diagnosis for production documents differ from clean training samples; teams promote a custom model because it feels better.
Incorrect Answers
B:This diagnosis is associated with the different failure where a malicious instruction triggers a privileged tool in a workflow where benchmarks cover only cooperative user wording; it does not directly solve the listed conditions: scanned forms suddenly produce inaccurate predictions; candidate improves one metric while violating another; a high aggregate score masks critical field errors.
Question 304
A risk analysis highlights a threat in this workflow: Many teams copy slightly different instructions for the same task. Exposure: A prompt editor could change production decisions without tracked approval. Which enforceable safeguard addresses it at the relevant boundary?
- Restrict production prompt edits to approved maintainers
- Require exception ownership and policy-version review
- Fund security, evaluation, hosting and incident response
- Keep a known-good model and agent configuration for rollback
Explanation
Correct Answer
A:Many teams copy slightly different instructions for the same task; its risk scenario is that a prompt editor could change production decisions without tracked approval; the safeguard fits because governed templates provide versioned contracts and test fixtures and the restriction addresses that specific exposure.
Incorrect Answers
B:That protection is tied to the mechanism that a CoE publishes patterns, reusable controls, standards and escalation paths; it does not prevent the distinct risk where a prompt editor could change production decisions without tracked approval in the system under review.
C:The proposed control targets a custom build might reach production without the needed operational controls, while the given risk scenario is a prompt editor could change production decisions without tracked approval; these exposures arise in different operational contexts.
D:That protection is tied to the mechanism that controlled release gates restrict blast radius and support rollback; it does not prevent the distinct risk where a prompt editor could change production decisions without tracked approval in the system under review.
Question 305
A staging test has to detect the specific defect that a simple scripted topic cannot handle legitimate variants. In the workflow where procurement exceptions require comparing several policies and records, which exercise verifies the corrective behavior?
- Simulate failed verification and confirm no notification
- Test edge cases with conflicting documents and incomplete facts
- Run representative notes and validate schema and exclusions
- Create a test request and verify backend audit
- Test mixed intents with missing action parameters
- Send representative messages through translation and human review
Explanation
Correct Answer
B:For the required workflow where procurement exceptions require comparing several policies and records, the test needs to reproduce or prevent the failure where a simple scripted topic cannot handle legitimate variants; testing edge cases with conflicting documents and incomplete facts exercises the path where reasoning can plan through evidence and tools within enforced boundaries.
Incorrect Answers
A:Although the exercise might catch an unconstrained planner sends a notification before authorization, it would not directly validate the mechanism that reasoning can plan through evidence and tools within enforced boundaries in the target implementation.
C:Running representative notes and validating schema and exclusions verifies an adjacent objective where an agent must turn service notes into a fixed JSON summary; it leaves untested the causal path in which reasoning can plan through evidence and tools within enforced boundaries for this case.
D:The tested behavior is that a plugin provides authorized callable tools to an agent; the necessary evidence here concerns reasoning can plan through evidence and tools within enforced boundaries and the previously observed a simple scripted topic cannot handle legitimate variants.
E:Testing mixed intents with missing action parameters verifies an adjacent objective where users request a status check and an authorized follow-up action; it leaves untested the causal path in which reasoning can plan through evidence and tools within enforced boundaries for this case.
F:In the alternate setting where support documents arrive in several customer languages, the test exercises translation provides standardized text input while preserving original evidence; even a pass would not show that a simple scripted topic cannot handle legitimate variants has been prevented.
Question 306
A policy review covers an agent for which two applications try to overwrite the same customer credit field. Its most relevant exposure is that two apps might overwrite each other's supposedly authoritative customer values. What safeguard should be required before release?
- Quarantine unclassified or ownerless content pending review
- Require finance approval of assumptions and sensitivity ranges
- Restrict write APIs to the authoritative owning service
- Avoid publishing benefits without attribution and uncertainty bounds
- Operate independent identity, safety and observability controls
Explanation
Correct Answer
C:When two apps might overwrite each other's supposedly authoritative customer values, the use case two applications try to overwrite the same customer credit field requires a guardrail aligned with the behavior that field ownership and integration contracts prevent competing writes.
Incorrect Answers
A:The alternate guard addresses ownerless confidential drafts could become searchable before classification for search spans internal manuals with different owners and permissions; the present threat, two apps might overwrite each other's supposedly authoritative customer values, needs a different enforcement boundary.
B:The safeguard for a pilot reports gross labor savings but omits operating expenses deals with a different threat: unverified ROI assumptions might be published as approved financial forecasts; even after it is enforced, the exposure that two apps might overwrite each other's supposedly authoritative customer values remains possible in the target workflow where field ownership and integration contracts prevent competing writes.
D:For the use case where management attributes a sales lift to Copilot without a comparison, the control addresses the threat that seasonal demand could be misrepresented as attributable AI benefit; the specified agent must be protected against two apps might overwrite each other's supposedly authoritative customer values.
E:The proposed control targets an owned custom runtime might lack the host's identity and safety governance, while the given risk scenario is two apps might overwrite each other's supposedly authoritative customer values; these exposures arise in different operational contexts.
Question 307
Support engineers report this failure: Model-level refusal alone may not prevent a tool call. The business context is: A prompt injection requests data export to an outside address. What evidence should be collected to isolate the cause?
- Review critical-path durations for slow requests
- Inspect model and tool spans under one operation identifier
- Inspect field-missing and type-mismatch error clusters
- Compare requested export destinations against policy allowlists
Explanation
Correct Answer
D:Model-level refusal alone may not prevent a tool call; to attribute the defect rather than guess at a fix, comparing requested export destinations against policy allowlists examines the dependency where execution-time destination and permission checks block exfiltration.
Incorrect Answers
A:Reviewing critical-path durations for slow requests explores a separate connection, where per-span latency isolates model work, network waits and tool execution; the diagnosed requirement is a prompt injection requests data export to an outside address, and the actual defect is model-level refusal alone may not prevent a tool call.
B:A failing agent shows only a generic HTTP error; examining that other path does not determine whether execution-time destination and permission checks block exfiltration, which is what distinguishes the issue in the stated use case.
C:In the process where a Copilot prompt feeds records into downstream automation, the diagnostic procedure could explain generated output occasionally omits required keys; this particular incident instead turns on whether execution-time destination and permission checks block exfiltration and why model-level refusal alone may not prevent a tool call.
Question 308
The fault affects an AI-powered workflow in which new documents enter a production source automatically. Specifically, unreviewed confidential drafts appear in agent answers. Which diagnostic investigation best distinguishes the relevant failure path?
- Compare deployed agent configuration to its approved manifest
- Inspect source identifiers and ingestion timestamps in index records
- Inspect vector-query filters and tenant IDs of returned hits
- Audit quarantined document reasons and promotion events
Explanation
Correct Answer
D:The reported fault is that unreviewed confidential drafts appear in agent answers in a system where new documents enter a production source automatically; evidence from auditing quarantined document reasons and promotion events can show whether ingestion gates check classification, freshness and ownership.
Incorrect Answers
A:In the process where model and tool changes are promoted independently, the diagnostic procedure could explain a release uses stale prompts against a new tool signature; this particular incident instead turns on whether ingestion gates check classification, freshness and ownership and why unreviewed confidential drafts appear in agent answers.
B:Agents begin citing an obsolete article after an ingestion job; its matching diagnostic check is reasonable for no log identifies which source version was indexed; the present investigation needs evidence for the different failure where unreviewed confidential drafts appear in agent answers.
C:Assistant retrieves another tenant's incident detail; examining that other path does not determine whether ingestion gates check classification, freshness and ownership, which is what distinguishes the issue in the stated use case.
Question 309
The architect wants evidence before altering the deployment. In a workflow where a rewritten prompt improves a single demonstration, slightly different wording breaks correct entity selection. What should the team examine first?
- Compare backend access decisions for different users
- Compare errors by scan quality and source drift
- Compare ticket assignment confusion matrices across versions
- Analyze failures by paraphrase group and boundary type
Explanation
Correct Answer
D:Where slightly different wording breaks correct entity selection, the useful evidence comes from analyzing failures by paraphrase group and boundary type; its focus matches the architectural mechanism that regression sets cover lexical variation and task boundaries in the process where a rewritten prompt improves a single demonstration.
Incorrect Answers
A:That probe evaluates delegated authorization evaluates each caller's effective permissions in a workflow where agent looks up records on behalf of employees with different rights; the reported failure concerns the path for which regression sets cover lexical variation and task boundaries instead.
B:The evidence this method produces applies to a workflow where production documents differ from clean training samples; it cannot by itself establish why slightly different wording breaks correct entity selection in the current scenario.
C:The proposed inspection is associated with a support agent changes how it classifies priority intents, where controlled service evaluations validate routing before rollout; it would not expose the failure mode that slightly different wording breaks correct entity selection with the required specificity.
Question 310
Release reviewers will not accept a generic success message. The business scenario is that data scientists cannot explain why a newer classifier behaves differently, and the historical failure is: Models move to production without an audit of training inputs. Which test is appropriately targeted?
- Reconstruct one released model's source and threshold decisions
- Evaluate candidate and baseline on identical frozen examples
- Execute an authorized and a denied deletion in staging
- Degrade sample images and test extraction stability
Explanation
Correct Answer
A:Reconstructing one released model's source and threshold decisions is aligned with the failure that models move to production without an audit of training inputs; its causal relevance is that a model registry ties versions to lineage, evaluations and approval for the situation where data scientists cannot explain why a newer classifier behaves differently.
Incorrect Answers
B:The tested behavior is that predefined gates weigh accuracy, latency and business risk; the necessary evidence here concerns a model registry ties versions to lineage, evaluations and approval and the previously observed models move to production without an audit of training inputs.
C:A pass on that verification would speak to a tool deletes a record following an ambiguous conversation; it would not show that a model registry ties versions to lineage, evaluations and approval, which is needed to avoid the observed defect models move to production without an audit of training inputs.
D:For production documents differ from clean training samples, the alternate test is about whether stress tests measure sensitivity to expected input variation; it cannot prove the outcome needed after the reported regression where models move to production without an audit of training inputs.