Cisco 810-110 Practice Test 17
Question 160
A team wants to improve future performance from human feedback. Which practical correction or implementation decision is MOST directly appropriate?
- Implement a kill switch and supervised resumption rules.
- Version feedback datasets and validate changes before release.
- Define confidence and risk thresholds for handoff with full context.
- Allow any comment to rewrite deployed model policy without approval.
Explanation
Correct Answer
B:Version feedback datasets and validate changes before release. The intended behavior can be checked because repeated evaluation shows improvements without unauthorized changes to production weights.
Incorrect Answers
A:Implement a kill switch and supervised resumption rules. This could help Human override and reversible control, but it does not resolve the Feedback from reviewers without unapproved learning case. HITL feedback can guide later refinement but does not imply instant retraining.
C:Define confidence and risk thresholds for handoff with full context. This could help Low-confidence escalation to a specialist, but it does not resolve the Feedback from reviewers without unapproved learning case. HITL feedback can guide later refinement but does not imply instant retraining.
D:Allow any comment to rewrite deployed model policy without approval. This would not correct the Feedback from reviewers without unapproved learning workflow. Automatically treating every reviewer note as new global instructions may destabilize behavior.
Question 161
An assistant must synthesize structured highlights from multiple documents with traceability. An engineer must explain the recommendation for Long-document chunk and merge. Which underlying principle is the MOST accurate?
- Map-reduce summarization trades extra steps for bounded context and source traceability.
- Model repositories list capabilities that must match task modality requirements.
- Termination controls bound output but do not replace semantic completeness checks.
- Input and output consume bounded model resources and output limits can end responses early.
Explanation
Correct Answer
A:Map-reduce summarization trades extra steps for bounded context and source traceability. Therefore the implementation decision is to summarize bounded document chunks with source IDs, then combine and verify extracted claims.
Incorrect Answers
B:Model repositories list capabilities that must match task modality requirements. This describes the separate Modality-aware model repository selection mechanism; the present case instead requires Long-document chunk and merge. Dropping source identifiers during aggregation makes claims hard to verify.
C:Termination controls bound output but do not replace semantic completeness checks. This describes the separate Stop conditions and maximum output mechanism; the present case instead requires Long-document chunk and merge. Dropping source identifiers during aggregation makes claims hard to verify.
D:Input and output consume bounded model resources and output limits can end responses early. This describes the separate Response output token reservation mechanism; the present case instead requires Long-document chunk and merge. Dropping source identifiers during aggregation makes claims hard to verify.
Question 162
A development team splits testing and implementation among specialized agent roles. Select THREE statements that correctly combine the Multi-agent handoff with scoped roles method with evidence or its governing principle.
- Store verified checkpoints, operation IDs, and source references for safe resumption.
- Every receiving agent gets verified inputs and accountable completion criteria.
- Define clear handoff artifacts, owners, limits, and review of intermediate outputs.
- Specialization can help complex workflows when interfaces are controlled.
- Separate a proposed plan from a validated executor with allowlisted operations.
- Give every agent unrestricted credentials and no assigned responsibility.
Explanation
Correct Answers
B:Every receiving agent gets verified inputs and accountable completion criteria. This is directly observable validation of the Multi-agent handoff with scoped roles result.
C:Define clear handoff artifacts, owners, limits, and review of intermediate outputs. This is the intended action for Multi-agent handoff with scoped roles.
D:Specialization can help complex workflows when interfaces are controlled. This explains why the implementation and evidence fit together.
Incorrect Answers
A:Store verified checkpoints, operation IDs, and source references for safe resumption. This serves the different Durable agent checkpoints and resume objective, not the selected Multi-agent handoff with scoped roles plan. Specialization can help complex workflows when interfaces are controlled.
E:Separate a proposed plan from a validated executor with allowlisted operations. This serves the different Planner-executor role separation objective, not the selected Multi-agent handoff with scoped roles plan. Specialization can help complex workflows when interfaces are controlled.
F:Give every agent unrestricted credentials and no assigned responsibility. This is not part of a safe or effective Multi-agent handoff with scoped roles design. Unstructured agents can contradict each other or lose required evidence.
Question 163
An enterprise ingests documents daily and needs deleted policies to stop appearing. Which practical correction or implementation decision is MOST directly appropriate?
- Tune chunking using retrieval evaluation and token budgets.
- Build retrieval with scoped context and citation verification.
- Assume updating the source website instantly updates every cached embedding.
- Implement change-triggered re-indexing and deletion propagation.
Explanation
Correct Answer
D:Implement change-triggered re-indexing and deletion propagation. The intended behavior can be checked because new versions appear and deleted versions no longer surface in queries.
Incorrect Answers
A:Tune chunking using retrieval evaluation and token budgets. This could help Chunking and overlap for document retrieval, but it does not resolve the Retrieval index freshness case. RAG indexes are derived copies that require lifecycle synchronization.
B:Build retrieval with scoped context and citation verification. This could help Retrieval-augmented question answering, but it does not resolve the Retrieval index freshness case. RAG indexes are derived copies that require lifecycle synchronization.
C:Assume updating the source website instantly updates every cached embedding. This would not correct the Retrieval index freshness workflow. A stale vector index can keep serving withdrawn or superseded policy passages.
Question 164
A retailer wants to compare product returns across regions and categories. If the relevant Segmentation of exploratory datasets safeguard is omitted, which outcome is the MOST directly associated failure?
- Presenting speculative ideas as proven results can mislead stakeholders.
- Untracked edits can make visually convincing plots impossible to audit.
- Treating an outlier as a proven fraud event without context can mislead action.
- Comparing raw group totals without population sizes can misstate relative rates.
Explanation
Correct Answer
D:Comparing raw group totals without population sizes can misstate relative rates. Aggregation choices can reveal or hide patterns through grouping and denominators. The corrective approach is to compute group-specific statistics and verify comparability.
Incorrect Answers
A:Presenting speculative ideas as proven results can mislead stakeholders. This failure pattern belongs chiefly to Ideation as hypotheses rather than facts; it is not the failure most directly tested in the Segmentation of exploratory datasets situation. Comparing raw group totals without population sizes can misstate relative rates.
B:Untracked edits can make visually convincing plots impossible to audit. This failure pattern belongs chiefly to Reproducible EDA with provenance; it is not the failure most directly tested in the Segmentation of exploratory datasets situation. Comparing raw group totals without population sizes can misstate relative rates.
C:Treating an outlier as a proven fraud event without context can mislead action. This failure pattern belongs chiefly to Exploratory data anomaly identification; it is not the failure most directly tested in the Segmentation of exploratory datasets situation. Comparing raw group totals without population sizes can misstate relative rates.
Question 165
An AI workflow accepts user data and later writes model-produced field values into a CRM. Which approach BEST meets the requirement while respecting the system constraints?
- Validate incoming text and generated structured outputs against domain rules before side effects.
- Build a test set of unanswerable, contradictory, and malicious prompts with expected behavior.
- Disable the CRM's input validation because the prompt already requests valid JSON.
- Treat response instructions as formatting guidance and independently verify factual grounding.
Explanation
Correct Answer
A:Validate incoming text and generated structured outputs against domain rules before side effects. Prompt wording is not a security control; deterministic validation must gate execution. Verification should demonstrate that invalid or out-of-policy data is rejected before the write.
Incorrect Answers
B:Build a test set of unanswerable, contradictory, and malicious prompts with expected behavior. This primarily addresses Adversarial evaluation of hallucinations, not the specific Input and output validation guardrails decision. Here, prompt wording is not a security control; deterministic validation must gate execution.
C:Disable the CRM's input validation because the prompt already requests valid JSON. This is an unsafe or ineffective shortcut for Input and output validation guardrails. Executing arbitrary generated fields can corrupt records or trigger unauthorized actions.
D:Treat response instructions as formatting guidance and independently verify factual grounding. This primarily addresses Separation of prompt guidance from evidence verification, not the specific Input and output validation guardrails decision. Here, prompt wording is not a security control; deterministic validation must gate execution.
Question 166
A searchable knowledge base needs to match paraphrases like password reset and account recovery. Which approach BEST meets the requirement while respecting the system constraints?
- Sort document IDs alphabetically and call it vector semantic retrieval.
- Apply per-user authorization and metadata filters before exposing retrieved context.
- Retrieve authorized relevant passages and ground generation in them with source references.
- Use compatible vector embeddings for queries and documents with a similarity index.
Explanation
Correct Answer
D:Use compatible vector embeddings for queries and documents with a similarity index. Similarity search compares embedding geometry; it does not guarantee factual truth. Verification should demonstrate that paraphrased queries retrieve relevant indexed passages in held-out tests.
Incorrect Answers
A:Sort document IDs alphabetically and call it vector semantic retrieval. This is an unsafe or ineffective shortcut for Embedding similarity retrieval. Ranking solely by unrelated raw identifiers misses semantic relevance.
B:Apply per-user authorization and metadata filters before exposing retrieved context. This primarily addresses Retrieval ACL and metadata filtering, not the specific Embedding similarity retrieval decision. Here, similarity search compares embedding geometry; it does not guarantee factual truth.
C:Retrieve authorized relevant passages and ground generation in them with source references. This primarily addresses Retrieval-augmented question answering, not the specific Embedding similarity retrieval decision. Here, similarity search compares embedding geometry; it does not guarantee factual truth.
Question 167
A developer wants a draft Python parser from a natural-language specification. After the recommended Code-generation model with execution tests change is implemented, which observation MOST directly verifies that it works?
- Paraphrased queries retrieve relevant indexed passages in held-out tests.
- The answer identifies visible relevant features and avoids unsupported hidden details.
- The generated program passes representative unit tests and security review.
- The model produces coherent textual summaries that reference the supplied messages.
Explanation
Correct Answer
C:The generated program passes representative unit tests and security review. This directly tests whether the Code-generation model with execution tests goal was met. Code generation proposes source text but does not prove the code executes correctly.
Incorrect Answers
A:Paraphrased queries retrieve relevant indexed passages in held-out tests. This is evidence about Embedding similarity retrieval, not direct proof of the required Code-generation model with execution tests behavior. The relevant check is that the generated program passes representative unit tests and security review.
B:The answer identifies visible relevant features and avoids unsupported hidden details. This is evidence about Multimodal visual question answering, not direct proof of the required Code-generation model with execution tests behavior. The relevant check is that the generated program passes representative unit tests and security review.
D:The model produces coherent textual summaries that reference the supplied messages. This is evidence about Autoregressive language generation versus image diffusion, not direct proof of the required Code-generation model with execution tests behavior. The relevant check is that the generated program passes representative unit tests and security review.
Question 168
A developer requests a generated transformation that must produce precise test outputs. An engineer must explain the recommendation for External calculator and code validation. Which underlying principle is the MOST accurate?
- Attribution reduces unsupported claims but citations must be checked for correctness.
- Few-shot prompting uses examples to convey patterns without modifying model weights.
- Prompt wording is not a security control; deterministic validation must gate execution.
- Generative fluent text is not proof of exact arithmetic or code execution.
Explanation
Correct Answer
D:Generative fluent text is not proof of exact arithmetic or code execution. Therefore the implementation decision is to use a trusted calculator or test harness to verify numerical and executable results.
Incorrect Answers
A:Relying on unaudited mental arithmetic can produce confident numeric errors. Attribution reduces unsupported claims but citations must be checked for correctness.
B:Relying on unaudited mental arithmetic can produce confident numeric errors. Few-shot prompting uses examples to convey patterns without modifying model weights.
C:Relying on unaudited mental arithmetic can produce confident numeric errors. Prompt wording is not a security control; deterministic validation must gate execution.
Question 169
A claims agent can be in gather evidence, review, approve, or complete states. If the relevant Finite-state agent orchestration safeguard is omitted, which outcome is the MOST directly associated failure?
- Free-form malformed arguments can corrupt external systems.
- Untracked state transitions can skip required reviews or double-execute steps.
- Treating every published resource URI as an executable write operation misuses the protocol.
- Giving free-form planning unrestricted tool credentials creates avoidable side-effect risk.
Explanation
Correct Answer
B:Untracked state transitions can skip required reviews or double-execute steps. State machines make agent progress and failure handling inspectable. The corrective approach is to use a durable state graph with transition validation.
Incorrect Answers
A:Free-form malformed arguments can corrupt external systems. This failure pattern belongs chiefly to Structured tool contracts and validation; it is not the failure most directly tested in the Finite-state agent orchestration situation. Untracked state transitions can skip required reviews or double-execute steps.
C:Treating every published resource URI as an executable write operation misuses the protocol. This failure pattern belongs chiefly to MCP tool execution primitive; it is not the failure most directly tested in the Finite-state agent orchestration situation. Untracked state transitions can skip required reviews or double-execute steps.
D:Giving free-form planning unrestricted tool credentials creates avoidable side-effect risk. This failure pattern belongs chiefly to Planner-executor role separation; it is not the failure most directly tested in the Finite-state agent orchestration situation. Untracked state transitions can skip required reviews or double-execute steps.
Question 170
An adversary uses search-result text to induce an unrelated outbound network action. Which practical correction or implementation decision is MOST directly appropriate?
- Treat an address found in untrusted text as explicit user approval to email secrets.
- Isolate retrieved content and constrain side-effect tools to user-authorized tasks.
- Maintain trusted instruction hierarchy and harden tool authorization.
- Enforce least-privilege tools and user authorization for all outbound sharing.
Explanation
Correct Answer
D:Enforce least-privilege tools and user authorization for all outbound sharing. The intended behavior can be checked because no protected content is sent to an unapproved recipient in attack simulations.
Incorrect Answers
A:Treat an address found in untrusted text as explicit user approval to email secrets. This would not correct the Injection-driven data exfiltration workflow. Trusting a document-defined destination can leak confidential data.
B:Isolate retrieved content and constrain side-effect tools to user-authorized tasks. This could help Indirect injection in retrieved webpages, but it does not resolve the Injection-driven data exfiltration case. Prompt injection can target tool permissions, not only the content of a model's answer.
C:Maintain trusted instruction hierarchy and harden tool authorization. This could help Direct user instruction override attempt, but it does not resolve the Injection-driven data exfiltration case. Prompt injection can target tool permissions, not only the content of a model's answer.