Question 400 A low-privilege employee asks a multi-agent supervisor to retrieve confidential payroll and the document agent forwards unrestricted snippets to the data agent. Which design change directly satisfies the stated constraints? Route the structured analytics intent to a permissioned Genie data interface using the approved semantic definitions. Apply caller identity and row/document entitlements before every delegated tool call, including inter-agent handoffs. Classify the incident as a privacy/access-control failure, block unauthorized retrieval and validate output against disclosure rules. Route policy prose to the document agent and current counts to governed Genie data tooling, then reconcile the combined answer with citations. ExplanationCorrect Answer B:The action targets the stated cause: Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. Check the repair by: replay with an unauthorized identity and verify neither delegated tool trace contains payroll records. Other Answers A:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. C:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. D:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path.
Question 401 A low-privilege employee asks a multi-agent supervisor to retrieve confidential payroll and the document agent forwards unrestricted snippets to the data agent. Which causal explanation best accounts for the failure? Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. Different facts live in different authoritative systems; a multi-agent router can delegate to each without conflating evidence types. Business aggregate questions require trusted table/query semantics, not document similarity that may return stale marketing summaries. A final response filter prevents any disclosure even after an unrestricted payroll snippet is shared with another agent. ExplanationCorrect Answer A:Apply caller identity and row/document entitlements before every delegated tool call, including inter-agent handoffs. Other Answers B:That mechanism describes another problem: A user asks a mixed question about warranty terms and current claim counts, but one general assistant fabricates both from a single paragraph. It does not explain this case, where delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. C:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. D:The claimed final filter does not undo a prior cross-agent transfer of confidential payroll material; caller permissions must be enforced before retrieval and delegation.
Question 402 A low-privilege employee asks a multi-agent supervisor to retrieve confidential payroll and the document agent forwards unrestricted snippets to the data agent. Which concrete acceptance test best demonstrates the proposed change works? Test a mixed question and verify the policy citation and SQL-backed count both appear with correct provenance. Query with a public identity and confirm no restricted source ID enters the prompt or returned text. Replay with an unauthorized identity and verify neither delegated tool trace contains payroll records. Compare the answer to a validated SQL aggregate and inspect permissions and the executed Genie query. ExplanationCorrect Answer C:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. Other Answers A:Replay with an unauthorized identity and verify neither delegated tool trace contains payroll records. B:Replay with an unauthorized identity and verify neither delegated tool trace contains payroll records. D:Replay with an unauthorized identity and verify neither delegated tool trace contains payroll records.
Question 403 A low-privilege employee asks a multi-agent supervisor to retrieve confidential payroll and the document agent forwards unrestricted snippets to the data agent. Which operational practice should be made a durable requirement? Keep intent-routing tests for structured measures and enforce semantic-model governance. Propagate least-privilege identity and audit access decisions across all agent boundaries. Enforce pre-retrieval access checks and monitor suspected sensitive-data leakage in traces. Document per-agent evidence boundaries and validate cross-agent answer aggregation. ExplanationCorrect Answer B:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. Other Answers A:That delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. C:Delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path. D:This control addresses a different exposure: A user asks a mixed question about warranty terms and current claim counts, but one general assistant fabricates both from a single paragraph. It would not prevent this failure, because delegating to a specialist does not grant new privileges; cross-agent context is another potential disclosure path.
Question 404 A supervisor agent always chooses the document agent for “how many” questions, even when the user mentions a governed metric view by name. What should the implementation team change first to resolve the described defect? Add intent tests using schema hints and metric identifiers, routing metric computations to Genie and policy interpretation to documents. Log tool result identifiers and prompt context spans in the trace, then use recorded traces to review the failing branches. Route the structured analytics intent to a permissioned Genie data interface using the approved semantic definitions. Route policy prose to the document agent and current counts to governed Genie data tooling, then reconcile the combined answer with citations. ExplanationCorrect Answer A:The action targets the stated cause: Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. Check the repair by: measure router confusion on ambiguous analytic/document prompts and inspect target-agent selections before rollout. Other Answers B:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. C:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. D:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval.
Question 405 A supervisor agent always chooses the document agent for “how many” questions, even when the user mentions a governed metric view by name. Why does the current approach fail under these conditions? Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. Business aggregate questions require trusted table/query semantics, not document similarity that may return stale marketing summaries. Different facts live in different authoritative systems; a multi-agent router can delegate to each without conflating evidence types. Trace context provides causality between retrieval, tool execution and final generation; isolated logs cannot reconstruct the path. ExplanationCorrect Answer A:Add intent tests using schema hints and metric identifiers, routing metric computations to Genie and policy interpretation to documents. Other Answers B:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. C:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. D:That mechanism describes another problem: A multi-tool agent is returning unpredictable answers because intermediate tool results are not linked to the generated text in its MLflow experiment. It does not explain this case, where surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval.
Question 406 A supervisor agent always chooses the document agent for “how many” questions, even when the user mentions a governed metric view by name. Which observation would establish that the requirement is now met? Select a failed response in MLflow and navigate to the exact source/tool span responsible for its unsupported content. Compare the answer to a validated SQL aggregate and inspect permissions and the executed Genie query. Measure router confusion on ambiguous analytic/document prompts and inspect target-agent selections before rollout. Test a mixed question and verify the policy citation and SQL-backed count both appear with correct provenance. ExplanationCorrect Answer C:Surface cue words alone are insufficient to distinguish analytical aggregation from narrative retrieval. Other Answers A:Measure router confusion on ambiguous analytic/document prompts and inspect target-agent selections before rollout. B:Measure router confusion on ambiguous analytic/document prompts and inspect target-agent selections before rollout. D:Measure router confusion on ambiguous analytic/document prompts and inspect target-agent selections before rollout.
Question 407 A pyfunc wrapper receives a request object with question and user_id, but its predict method sends the whole object as prompt text to an LLM. Which engineering action most directly addresses this specific requirement? Convert provider responses into a stable JSON-compatible typed output in post-processing, rejecting missing or invalid fields. Parse and validate question separately, resolve authorized context from user_id, then transform the model output to the serving contract. Package the retriever dependencies and configuration with the logged model and test the loaded artifact in a clean environment. Parse and validate model outputs against approved class labels before writing the curated reporting table. ExplanationCorrect Answer B:The action targets the stated cause: A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. Check the repair by: send a synthetic unauthorized user and verify no private context is included, while the returned object matches the signature. Other Answers A:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. C:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. D:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs.
Question 408 A pyfunc wrapper receives a request object with question and user_id, but its predict method sends the whole object as prompt text to an LLM. What architectural constraint is most directly responsible? Model flavor packaging needs reproducible code and dependencies; local notebook imports are not automatically available to remote serving. Inference returns generated content; SQL invocation alone does not guarantee every label meets a downstream business contract. SDK response objects are not necessarily serving wire-format values; pyfunc output must satisfy the declared interface. A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. ExplanationCorrect Answer D:Parse and validate question separately, resolve authorized context from user_id, then transform the model output to the serving contract. Other Answers A:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. B:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. C:That mechanism describes another problem: A PythonModel returns a raw SDK object that cannot be JSON serialized by the endpoint, even though model inference succeeds. It does not explain this case, where a pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs.
Question 409 A pyfunc wrapper receives a request object with question and user_id, but its predict method sends the whole object as prompt text to an LLM. What validation evidence would most specifically confirm the resolution? Load the registered artifact in an isolated runtime and execute the same retrieval-dependent test request successfully. Invoke the serving interface and verify real JSON serialization and schema conformance on success and error paths. Run a labeled batch with malformed responses and verify invalid rows are quarantined rather than promoted. Send a synthetic unauthorized user and verify no private context is included, while the returned object matches the signature. ExplanationCorrect Answer D:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. Other Answers A:Send a synthetic unauthorized user and verify no private context is included, while the returned object matches the signature. B:Send a synthetic unauthorized user and verify no private context is included, while the returned object matches the signature. C:Send a synthetic unauthorized user and verify no private context is included, while the returned object matches the signature.
Question 410 A pyfunc wrapper receives a request object with question and user_id, but its predict method sends the whole object as prompt text to an LLM. What prevention measure is most targeted to this underlying cause? Keep input parsing, entitlement checks and output serialization as independently tested pyfunc steps. Add quality constraints, retry ceilings and exception tables to batch inference workflows. Maintain signature fixtures and separate post-processing tests for provider SDK updates. Freeze dependency versions and run clean-environment smoke tests before registry promotion. ExplanationCorrect Answer A:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. Other Answers B:A pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. C:This control addresses a different exposure: A PythonModel returns a raw SDK object that cannot be JSON serialized by the endpoint, even though model inference succeeds. It would not prevent this failure, because a pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs. D:That a pyfunc model's predict interface is an adapter; it must explicitly map request fields into safe model inputs and typed outputs.
