Topic 11 Practice Test 2 covers AI Governance, Compliance, Lineage, Auditability, Fairness, and Responsible AI in Domain 3 (Tasks 3.3-3.4). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A scientific assistant should escalate low-confidence answers and operations wants to measure how often uncertainty occurs. Which approach best satisfies the requirement? Choose ONE.
- Policy-as-code deployment gate
- Confidence and uncertainty telemetry
- Prompt A/B fairness test
- Human review for high-impact exception
Correct Answer: B
Correct Answer
Answer B is correct because Confidence and uncertainty telemetry: captures calibrated confidence or uncertainty indicators so downstream logic and operators can distinguish strong evidence from uncertain output. T11P2Q01 needs quantify uncertainty for routing and oversight; it directly fits because it will record a validated confidence or uncertainty signal and apply an escalation threshold.
Incorrect Answers
Answer A is incorrect because Policy-as-code deployment gate: evaluates machine-readable governance requirements in the delivery path and blocks releases that violate mandatory controls. Best used when organizational policy must be enforced consistently before production deployment; T11P2Q01 instead needs quantify uncertainty for routing and oversight, so it misses the deciding constraint.
Answer C is incorrect because Prompt A/B fairness test: runs controlled prompt variants against the same evaluation set so teams can measure whether wording changes improve fairness without guessing. Best used when a prompt change is proposed specifically to reduce disparate behavior; T11P2Q01 instead needs quantify uncertainty for routing and oversight, so it misses the deciding constraint.
Answer D is incorrect because Human review for high-impact exception: routes cases beyond automated policy confidence or authority to an accountable human reviewer before a consequential action is taken. Best used when the AI system must not autonomously decide certain high-risk exceptions; T11P2Q01 instead needs quantify uncertainty for routing and oversight, so it misses the deciding constraint.
Question 2
An agent produced an unexpected result after calling several tools, and the engineering review needs to see which orchestration steps and actions occurred. What should the developer implement? Choose ONE.
- Policy version correlation
- Bedrock agent trace
- Central model inventory
- Bedrock LLM-as-a-judge evaluation
Correct Answer: B
Correct Answer
Answer B is correct because Bedrock agent trace: captures agent orchestration details such as tool selection and intermediate execution steps for supported troubleshooting and transparency workflows. T11P2Q02 needs inspect the agent’s actual execution path rather than infer it from the final answer; it directly fits because it will enable and review Amazon Bedrock agent tracing for the invocation.
Incorrect Answers
Answer A is incorrect because Policy version correlation: records the exact policy or guardrail version alongside each decision so later audits can reproduce which rules were active. Best used when governance controls change over time and a generic resource name is insufficient for reconstruction; T11P2Q02 instead needs inspect the agent’s actual execution path rather than infer it from the final answer, so it misses the deciding constraint.
Answer C is incorrect because Central model inventory: maintains a governed inventory of deployed or approved AI assets with owners, versions, risk classifications, and lifecycle state. Best used when oversight teams need to know what AI systems actually exist across the organization; T11P2Q02 instead needs inspect the agent’s actual execution path rather than infer it from the final answer, so it misses the deciding constraint.
Answer D is incorrect because Bedrock LLM-as-a-judge evaluation: uses a separate evaluator model to score generator responses against selected evaluation criteria and provide scoring explanations. Best used when large-scale qualitative assessment needs automated comparative scoring with a defined rubric; T11P2Q02 instead needs inspect the agent’s actual execution path rather than infer it from the final answer, so it misses the deciding constraint.
Question 3
A benefits assistant recommends one plan and policy requires a short explanation of the eligibility factors and evidence considered. Which design most directly addresses the constraint? Choose ONE.
- User-facing rationale summary
- Predeployment responsible-AI approval gate
- Human evaluation for subjective harm review
- Risk-tiered approval workflow
Correct Answer: A
Correct Answer
Answer A is correct because User-facing rationale summary: provides a concise explanation of the factors or evidence behind an AI result without exposing protected implementation secrets or raw internal reasoning. T11P2Q03 needs explain the result in terms users can inspect and challenge; it directly fits because it will return a concise evidence-based rationale summary alongside the recommendation.
Incorrect Answers
Answer B is incorrect because Predeployment responsible-AI approval gate: requires documented review of model limitations, fairness evidence, safety controls, and ownership before a production release is authorized. Best used when responsible-AI evidence must be evaluated before users are exposed to a new model or prompt configuration; T11P2Q03 instead needs explain the result in terms users can inspect and challenge, so it misses the deciding constraint.
Answer C is incorrect because Human evaluation for subjective harm review: uses qualified reviewers when nuanced, contextual, or high-impact judgments are not adequately captured by automated metrics. Best used when governance needs defensible review of ambiguous outputs where machine scoring is insufficient; T11P2Q03 instead needs explain the result in terms users can inspect and challenge, so it misses the deciding constraint.
Answer D is incorrect because Risk-tiered approval workflow: applies stronger review requirements to higher-impact AI uses instead of treating every use case identically. Best used when governance effort should reflect the consequence of the proposed AI decision; T11P2Q03 instead needs explain the result in terms users can inspect and challenge, so it misses the deciding constraint.
Question 4
A medical-information assistant retrieves conflicting source passages and cannot determine which guidance applies to the user’s jurisdiction. What is the most appropriate implementation? Choose ONE.
- Longitudinal fairness baseline
- Automated misuse detection
- Uncertainty disclosure with escalation
- SageMaker Model Card intended-use record
Correct Answer: C
Correct Answer
Answer C is correct because Uncertainty disclosure with escalation: tells the user when evidence is insufficient and routes the case to a safer fallback instead of presenting speculation as fact. T11P2Q04 needs avoid pretending certainty when supporting evidence conflicts; it directly fits because it will disclose the uncertainty and escalate or decline according to policy.
Incorrect Answers
Answer A is incorrect because Longitudinal fairness baseline: compares current fairness measurements with an approved baseline so governance can detect drift rather than evaluate each snapshot in isolation. Best used when the organization needs to know whether disparity is worsening over time; T11P2Q04 instead needs avoid pretending certainty when supporting evidence conflicts, so it misses the deciding constraint.
Answer B is incorrect because Automated misuse detection: continuously analyzes usage or outputs for patterns that indicate prohibited behavior, abuse, or policy violations. Best used when governance requires ongoing detection after deployment rather than a one-time prelaunch review; T11P2Q04 instead needs avoid pretending certainty when supporting evidence conflicts, so it misses the deciding constraint.
Answer D is incorrect because SageMaker Model Card intended-use record: documents a model’s purpose, intended and disallowed uses, risk rating, evaluation evidence, and other governance details in a structured lifecycle artifact. Best used when reviewers need one governed record explaining where a model may and may not be used; T11P2Q04 instead needs avoid pretending certainty when supporting evidence conflicts, so it misses the deciding constraint.
Question 5
A compliance analyst must validate each generated interpretation against the exact contract language used by the assistant. Which solution best fits this production requirement? Choose ONE.
- Representative fairness dataset
- Versioned SageMaker Model Card update
- Bias drift monitoring
- Evidence snippet presentation
Correct Answer: D
Correct Answer
Answer D is correct because Evidence snippet presentation: shows the relevant source excerpts that support an answer so reviewers can compare the generated claim with underlying material. T11P2Q05 needs let the reviewer inspect the supporting text directly; it directly fits because it will present the relevant retrieved evidence snippets with their source identifiers.
Incorrect Answers
Answer A is incorrect because Representative fairness dataset: includes realistic cases from relevant user groups so fairness evaluation measures the populations and scenarios the system will actually encounter. Best used when a balanced synthetic set that omits important user groups can give false assurance; T11P2Q05 instead needs let the reviewer inspect the supporting text directly, so it misses the deciding constraint.
Answer B is incorrect because Versioned SageMaker Model Card update: creates a new model-card version when governed details change, preserving prior versions as an immutable audit history. Best used when governance must retain what reviewers previously approved rather than overwrite it; T11P2Q05 instead needs let the reviewer inspect the supporting text directly, so it misses the deciding constraint.
Answer C is incorrect because Bias drift monitoring: tracks fairness-related measurements over time so changing traffic, data, or model behavior can reveal emerging disparities. Best used when a system passed fairness evaluation initially but its population mix and prompts change over months; T11P2Q05 instead needs let the reviewer inspect the supporting text directly, so it misses the deciding constraint.
Question 6
A loan-document assistant has 95% aggregate extraction accuracy, but governance needs to know whether error rates differ materially across language groups. Which approach best satisfies the requirement? Choose ONE.
- Bedrock Guardrail policy enforcement
- Automated governance alert and remediation
- SageMaker lineage graph
- Fairness metric by protected cohort
Correct Answer: D
Correct Answer
Answer D is correct because Fairness metric by protected cohort: measures outcome or quality differences across relevant groups instead of averaging all users together. T11P2Q06 needs measure disparities that an overall average can conceal; it directly fits because it will calculate the approved fairness or quality metric separately for each relevant cohort.
Incorrect Answers
Answer A is incorrect because Bedrock Guardrail policy enforcement: applies configured content, denied-topic, sensitive-information, prompt-attack, and related safeguards consistently during supported model interactions. Best used when responsible-AI policy requires a reusable runtime control around model inputs and outputs; T11P2Q06 instead needs measure disparities that an overall average can conceal, so it misses the deciding constraint.
Answer B is incorrect because Automated governance alert and remediation: connects detected violations to notifications, containment, rollback, or human review so governance findings produce timely action. Best used when a policy breach must trigger an operational response rather than only create a dashboard entry; T11P2Q06 instead needs measure disparities that an overall average can conceal, so it misses the deciding constraint.
Answer C is incorrect because SageMaker lineage graph: tracks relationships among datasets, processing or training activities, models, and deployments so model provenance can be reconstructed. Best used when an auditor must trace a deployed model back through the artifacts and jobs that produced it; T11P2Q06 instead needs measure disparities that an overall average can conceal, so it misses the deciding constraint.
Question 7
Gender-level and region-level scores look acceptable, but reviewers suspect a problem affecting one gender within a specific region. What should the developer implement? Choose ONE.
- IAM-enforced guardrail requirement
- Fairness slice intersection analysis
- AWS Glue Data Catalog source registration
- Token-level sensitive-data redaction
Correct Answer: B
Correct Answer
Answer B is correct because Fairness slice intersection analysis: evaluates intersections of attributes when a broad group metric could hide a problem affecting a smaller subgroup. T11P2Q07 needs find bias hidden inside coarse aggregate groupings; it directly fits because it will evaluate the fairness metric on the relevant intersecting slices.
Incorrect Answers
Answer A is incorrect because IAM-enforced guardrail requirement: uses the Bedrock guardrail condition key in IAM so specified inference calls are denied if the required guardrail identifier is omitted. Best used when developers must not be able to bypass the organization’s mandatory guardrail accidentally; T11P2Q07 instead needs find bias hidden inside coarse aggregate groupings, so it misses the deciding constraint.
Answer C is incorrect because AWS Glue Data Catalog source registration: registers governed datasets and metadata in a searchable catalog so data sources have stable identities and descriptions. Best used when teams need an enterprise record of which governed source fed a GenAI workflow; T11P2Q07 instead needs find bias hidden inside coarse aggregate groupings, so it misses the deciding constraint.
Answer D is incorrect because Token-level sensitive-data redaction: removes or masks sensitive spans before they are persisted or exposed, reducing the amount of regulated content in governance logs and outputs. Best used when compliance requires detailed logs without retaining raw secrets or personal identifiers; T11P2Q07 instead needs find bias hidden inside coarse aggregate groupings, so it misses the deciding constraint.
Question 8
A team rewrites system instructions to reduce stereotyping and needs evidence that the new prompt performs better across demographic slices. Which design most directly addresses the constraint? Choose ONE.
- Prompt A/B fairness test
- Model limitations in model card
- Source-attribution metadata tags
- AI response logging with correlation IDs
Correct Answer: A
Correct Answer
Answer A is correct because Prompt A/B fairness test: runs controlled prompt variants against the same evaluation set so teams can measure whether wording changes improve fairness without guessing. T11P2Q08 needs compare the fairness impact of the prompt variants under controlled inputs; it directly fits because it will run an A/B evaluation using the same representative fairness dataset.
Incorrect Answers
Answer B is incorrect because Model limitations in model card: documents known limitations, excluded uses, assumptions, risks, and evaluation findings so consumers understand where the model should not be trusted. Best used when responsible deployment requires communicating boundaries, not only benchmark strengths; T11P2Q08 instead needs compare the fairness impact of the prompt variants under controlled inputs, so it misses the deciding constraint.
Answer C is incorrect because Source-attribution metadata tags: carries source identifiers or provenance metadata with content so downstream retrieval and generation can report where evidence originated. Best used when generated content must be traceable to the specific documents that supported it; T11P2Q08 instead needs compare the fairness impact of the prompt variants under controlled inputs, so it misses the deciding constraint.
Answer D is incorrect because AI response logging with correlation IDs: records model outputs together with request, model, policy, and trace identifiers so reviewers can reconstruct a production interaction. Best used when an incident team must connect a disputed answer to the exact invocation and control configuration; T11P2Q08 instead needs compare the fairness impact of the prompt variants under controlled inputs, so it misses the deciding constraint.
Question 9
Two assistant versions generate thousands of responses, and the team needs repeatable automated scoring for tone, helpfulness, and policy alignment before human spot checks. What is the most appropriate implementation? Choose ONE.
- Bedrock LLM-as-a-judge evaluation
- AWS CloudTrail API audit history
- Output policy filter
- Automated Lambda compliance check
Correct Answer: A
Correct Answer
Answer A is correct because Bedrock LLM-as-a-judge evaluation: uses a separate evaluator model to score generator responses against selected evaluation criteria and provide scoring explanations. T11P2Q09 needs scale rubric-based comparative evaluation beyond manual review alone; it directly fits because it will run an Amazon Bedrock model evaluation job using an LLM as the judge.
Incorrect Answers
Answer B is incorrect because AWS CloudTrail API audit history: records supported AWS API activity with caller, time, source, request context, and affected resources for security and compliance review. Best used when investigators need to know which identity changed a governed AI resource; T11P2Q09 instead needs scale rubric-based comparative evaluation beyond manual review alone, so it misses the deciding constraint.
Answer C is incorrect because Output policy filter: evaluates generated content against explicit organizational rules before it reaches the user or downstream system. Best used when the organization must enforce a policy beyond the base model’s default behavior; T11P2Q09 instead needs scale rubric-based comparative evaluation beyond manual review alone, so it misses the deciding constraint.
Answer D is incorrect because Automated Lambda compliance check: runs custom deterministic policy logic when an organization needs a control not natively represented by the model service. Best used when a deployment rule can be expressed programmatically and must be checked before AI output proceeds; T11P2Q09 instead needs scale rubric-based comparative evaluation beyond manual review alone, so it misses the deciding constraint.
Question 10
An evaluation set contains culturally sensitive edge cases for which automatic toxicity scores disagree with expert policy interpretation. Which solution best fits this production requirement? Choose ONE.
- CloudWatch decision logging
- Human evaluation for subjective harm review
- Governance exception register
- Versioned guardrail promotion
Correct Answer: B
Correct Answer
Answer B is correct because Human evaluation for subjective harm review: uses qualified reviewers when nuanced, contextual, or high-impact judgments are not adequately captured by automated metrics. T11P2Q10 needs apply informed human judgment to ambiguous responsible-AI cases; it directly fits because it will use a structured human review process with a defined rubric and reviewer guidance.
Incorrect Answers
Answer A is incorrect because CloudWatch decision logging: collects application decision records and operational evidence that can be queried for review, alerts, and compliance reporting. Best used when the application must retain why a policy path accepted, blocked, or escalated an AI request; T11P2Q10 instead needs apply informed human judgment to ambiguous responsible-AI cases, so it misses the deciding constraint.
Answer C is incorrect because Governance exception register: records approved deviations from standard policy with scope, justification, owner, expiration, and compensating controls. Best used when temporary exceptions must remain visible and automatically come back for review; T11P2Q10 instead needs apply informed human judgment to ambiguous responsible-AI cases, so it misses the deciding constraint.
Answer D is incorrect because Versioned guardrail promotion: creates a stable guardrail version from a tested draft so production applications can reference an approved configuration snapshot. Best used when policy changes must be tested before they affect live traffic; T11P2Q10 instead needs apply informed human judgment to ambiguous responsible-AI cases, so it misses the deciding constraint.
Question 11
A customer-support model is retrained monthly and each release must show that fairness has not regressed relative to the approved version. Which approach best satisfies the requirement? Choose ONE.
- Longitudinal fairness baseline
- Tamper-resistant audit archive
- Human review for high-impact exception
- Evidence-backed source citations
Correct Answer: A
Correct Answer
Answer A is correct because Longitudinal fairness baseline: compares current fairness measurements with an approved baseline so governance can detect drift rather than evaluate each snapshot in isolation. T11P2Q11 needs detect fairness regression between releases; it directly fits because it will store approved baseline metrics and compare every candidate release against them.
Incorrect Answers
Answer B is incorrect because Tamper-resistant audit archive: stores finalized audit evidence under retention controls that prevent ordinary modification or premature deletion. Best used when regulators require retained AI decision records to resist alteration during the mandated period; T11P2Q11 instead needs detect fairness regression between releases, so it misses the deciding constraint.
Answer C is incorrect because Human review for high-impact exception: routes cases beyond automated policy confidence or authority to an accountable human reviewer before a consequential action is taken. Best used when the AI system must not autonomously decide certain high-risk exceptions; T11P2Q11 instead needs detect fairness regression between releases, so it misses the deciding constraint.
Answer D is incorrect because Evidence-backed source citations: presents the evidence or source attribution supporting a generated answer so users can verify important claims. Best used when transparency requires users to see where a factual answer came from; T11P2Q11 instead needs detect fairness regression between releases, so it misses the deciding constraint.
Question 12
A multilingual assistant is evaluated only on English prompts even though half of production traffic is Spanish and Portuguese. What should the developer implement? Choose ONE.
- Confidence and uncertainty telemetry
- Governance owner and approval metadata
- Policy version correlation
- Representative fairness dataset
Correct Answer: D
Correct Answer
Answer D is correct because Representative fairness dataset: includes realistic cases from relevant user groups so fairness evaluation measures the populations and scenarios the system will actually encounter. T11P2Q12 needs make fairness testing reflect real deployment populations; it directly fits because it will build the evaluation dataset with representative language and user-group coverage.
Incorrect Answers
Answer A is incorrect because Confidence and uncertainty telemetry: captures calibrated confidence or uncertainty indicators so downstream logic and operators can distinguish strong evidence from uncertain output. Best used when the system must expose uncertainty rather than present every response with equal certainty; T11P2Q12 instead needs make fairness testing reflect real deployment populations, so it misses the deciding constraint.
Answer B is incorrect because Governance owner and approval metadata: assigns accountable owners, reviewers, and approval state to AI assets so policy exceptions and lifecycle decisions have clear responsibility. Best used when an organization needs a named decision-maker for each production GenAI system; T11P2Q12 instead needs make fairness testing reflect real deployment populations, so it misses the deciding constraint.
Answer C is incorrect because Policy version correlation: records the exact policy or guardrail version alongside each decision so later audits can reproduce which rules were active. Best used when governance controls change over time and a generic resource name is insufficient for reconstruction; T11P2Q12 instead needs make fairness testing reflect real deployment populations, so it misses the deciding constraint.
Question 13
A financial assistant must block prohibited investment-advice topics and sensitive-data leakage across several supported Bedrock models. Which design most directly addresses the constraint? Choose ONE.
- Predeployment responsible-AI approval gate
- Policy-as-code deployment gate
- Bedrock Guardrail policy enforcement
- Bedrock agent trace
Correct Answer: C
Correct Answer
Answer C is correct because Bedrock Guardrail policy enforcement: applies configured content, denied-topic, sensitive-information, prompt-attack, and related safeguards consistently during supported model interactions. T11P2Q13 needs enforce the same policy control at inference time across model choices; it directly fits because it will associate the approved Amazon Bedrock Guardrail with the application invocations.
Incorrect Answers
Answer A is incorrect because Predeployment responsible-AI approval gate: requires documented review of model limitations, fairness evidence, safety controls, and ownership before a production release is authorized. Best used when responsible-AI evidence must be evaluated before users are exposed to a new model or prompt configuration; T11P2Q13 instead needs enforce the same policy control at inference time across model choices, so it misses the deciding constraint.
Answer B is incorrect because Policy-as-code deployment gate: evaluates machine-readable governance requirements in the delivery path and blocks releases that violate mandatory controls. Best used when organizational policy must be enforced consistently before production deployment; T11P2Q13 instead needs enforce the same policy control at inference time across model choices, so it misses the deciding constraint.
Answer D is incorrect because Bedrock agent trace: captures agent orchestration details such as tool selection and intermediate execution steps for supported troubleshooting and transparency workflows. Best used when developers or reviewers need trace evidence of how an agent executed its workflow; T11P2Q13 instead needs enforce the same policy control at inference time across model choices, so it misses the deciding constraint.
Question 14
A platform team requires every Converse and InvokeModel call for a regulated app to include the approved guardrail, even if application code is changed. What is the most appropriate implementation? Choose ONE.
- User-facing rationale summary
- SageMaker Model Card intended-use record
- IAM-enforced guardrail requirement
- Central model inventory
Correct Answer: C
Correct Answer
Answer C is correct because IAM-enforced guardrail requirement: uses the Bedrock guardrail condition key in IAM so specified inference calls are denied if the required guardrail identifier is omitted. T11P2Q14 needs make guardrail use an authorization requirement rather than a coding convention; it directly fits because it will enforce the approved guardrail identifier with an IAM policy condition.
Incorrect Answers
Answer A is incorrect because User-facing rationale summary: provides a concise explanation of the factors or evidence behind an AI result without exposing protected implementation secrets or raw internal reasoning. Best used when users need understandable transparency about a consequential recommendation; T11P2Q14 instead needs make guardrail use an authorization requirement rather than a coding convention, so it misses the deciding constraint.
Answer B is incorrect because SageMaker Model Card intended-use record: documents a model’s purpose, intended and disallowed uses, risk rating, evaluation evidence, and other governance details in a structured lifecycle artifact. Best used when reviewers need one governed record explaining where a model may and may not be used; T11P2Q14 instead needs make guardrail use an authorization requirement rather than a coding convention, so it misses the deciding constraint.
Answer D is incorrect because Central model inventory: maintains a governed inventory of deployed or approved AI assets with owners, versions, risk classifications, and lifecycle state. Best used when oversight teams need to know what AI systems actually exist across the organization; T11P2Q14 instead needs make guardrail use an authorization requirement rather than a coding convention, so it misses the deciding constraint.
Question 15
A model performs poorly on handwritten documents and should not be used for that input type, even though its overall accuracy is strong. Which solution best fits this production requirement? Choose ONE.
- Versioned SageMaker Model Card update
- Uncertainty disclosure with escalation
- Risk-tiered approval workflow
- Model limitations in model card
Correct Answer: D
Correct Answer
Answer D is correct because Model limitations in model card: documents known limitations, excluded uses, assumptions, risks, and evaluation findings so consumers understand where the model should not be trusted. T11P2Q15 needs make a known limitation visible to downstream teams and auditors; it directly fits because it will record the limitation and prohibited use in the governed model card.
Incorrect Answers
Answer A is incorrect because Versioned SageMaker Model Card update: creates a new model-card version when governed details change, preserving prior versions as an immutable audit history. Best used when governance must retain what reviewers previously approved rather than overwrite it; T11P2Q15 instead needs make a known limitation visible to downstream teams and auditors, so it misses the deciding constraint.
Answer B is incorrect because Uncertainty disclosure with escalation: tells the user when evidence is insufficient and routes the case to a safer fallback instead of presenting speculation as fact. Best used when responsible behavior requires acknowledging when the system cannot support a reliable answer; T11P2Q15 instead needs make a known limitation visible to downstream teams and auditors, so it misses the deciding constraint.
Answer C is incorrect because Risk-tiered approval workflow: applies stronger review requirements to higher-impact AI uses instead of treating every use case identically. Best used when governance effort should reflect the consequence of the proposed AI decision; T11P2Q15 instead needs make a known limitation visible to downstream teams and auditors, so it misses the deciding constraint.
Question 16
A regulated workflow may answer only when a case has a current customer-consent flag in an internal system. Which approach best satisfies the requirement? Choose ONE.
- Automated misuse detection
- Automated Lambda compliance check
- SageMaker lineage graph
- Evidence snippet presentation
Correct Answer: B
Correct Answer
Answer B is correct because Automated Lambda compliance check: runs custom deterministic policy logic when an organization needs a control not natively represented by the model service. T11P2Q16 needs enforce a business-specific compliance prerequisite outside the FM itself; it directly fits because it will invoke a Lambda compliance check and block or escalate when consent is absent.
Incorrect Answers
Answer A is incorrect because Automated misuse detection: continuously analyzes usage or outputs for patterns that indicate prohibited behavior, abuse, or policy violations. Best used when governance requires ongoing detection after deployment rather than a one-time prelaunch review; T11P2Q16 instead needs enforce a business-specific compliance prerequisite outside the FM itself, so it misses the deciding constraint.
Answer C is incorrect because SageMaker lineage graph: tracks relationships among datasets, processing or training activities, models, and deployments so model provenance can be reconstructed. Best used when an auditor must trace a deployed model back through the artifacts and jobs that produced it; T11P2Q16 instead needs enforce a business-specific compliance prerequisite outside the FM itself, so it misses the deciding constraint.
Answer D is incorrect because Evidence snippet presentation: shows the relevant source excerpts that support an answer so reviewers can compare the generated claim with underlying material. Best used when citations alone are too opaque for a high-stakes review workflow; T11P2Q16 instead needs enforce a business-specific compliance prerequisite outside the FM itself, so it misses the deciding constraint.
Question 17
Safety engineers are iterating on denied topics in a draft but production must continue using the previously approved control until testing finishes. What should the developer implement? Choose ONE.
- Fairness metric by protected cohort
- Bias drift monitoring
- Versioned guardrail promotion
- AWS Glue Data Catalog source registration
Correct Answer: C
Correct Answer
Answer C is correct because Versioned guardrail promotion: creates a stable guardrail version from a tested draft so production applications can reference an approved configuration snapshot. T11P2Q17 needs separate experimental policy edits from the configuration enforced in production; it directly fits because it will create and explicitly deploy a tested Bedrock Guardrail version.
Incorrect Answers
Answer A is incorrect because Fairness metric by protected cohort: measures outcome or quality differences across relevant groups instead of averaging all users together. Best used when bias can be hidden by a strong overall score; T11P2Q17 instead needs separate experimental policy edits from the configuration enforced in production, so it misses the deciding constraint.
Answer B is incorrect because Bias drift monitoring: tracks fairness-related measurements over time so changing traffic, data, or model behavior can reveal emerging disparities. Best used when a system passed fairness evaluation initially but its population mix and prompts change over months; T11P2Q17 instead needs separate experimental policy edits from the configuration enforced in production, so it misses the deciding constraint.
Answer D is incorrect because AWS Glue Data Catalog source registration: registers governed datasets and metadata in a searchable catalog so data sources have stable identities and descriptions. Best used when teams need an enterprise record of which governed source fed a GenAI workflow; T11P2Q17 instead needs separate experimental policy edits from the configuration enforced in production, so it misses the deciding constraint.
Question 18
A claims assistant can recommend standard approvals, but policy requires a human reviewer for claims above a fixed monetary threshold. Which design most directly addresses the constraint? Choose ONE.
- Human review for high-impact exception
- Automated governance alert and remediation
- Fairness slice intersection analysis
- Source-attribution metadata tags
Correct Answer: A
Correct Answer
Answer A is correct because Human review for high-impact exception: routes cases beyond automated policy confidence or authority to an accountable human reviewer before a consequential action is taken. T11P2Q18 needs keep high-impact exceptions under human accountability; it directly fits because it will route above-threshold cases to a human approval step before action.
Incorrect Answers
Answer B is incorrect because Automated governance alert and remediation: connects detected violations to notifications, containment, rollback, or human review so governance findings produce timely action. Best used when a policy breach must trigger an operational response rather than only create a dashboard entry; T11P2Q18 instead needs keep high-impact exceptions under human accountability, so it misses the deciding constraint.
Answer C is incorrect because Fairness slice intersection analysis: evaluates intersections of attributes when a broad group metric could hide a problem affecting a smaller subgroup. Best used when responsible-AI review needs sufficient granularity to detect localized disparities; T11P2Q18 instead needs keep high-impact exceptions under human accountability, so it misses the deciding constraint.
Answer D is incorrect because Source-attribution metadata tags: carries source identifiers or provenance metadata with content so downstream retrieval and generation can report where evidence originated. Best used when generated content must be traceable to the specific documents that supported it; T11P2Q18 instead needs keep high-impact exceptions under human accountability, so it misses the deciding constraint.
Question 19
A disputed response occurred two months ago, after several policy revisions, and auditors need to know exactly which rules evaluated it. What is the most appropriate implementation? Choose ONE.
- AWS CloudTrail API audit history
- Prompt A/B fairness test
- Policy version correlation
- Token-level sensitive-data redaction
Correct Answer: C
Correct Answer
Answer C is correct because Policy version correlation: records the exact policy or guardrail version alongside each decision so later audits can reproduce which rules were active. T11P2Q19 needs bind each logged interaction to the precise control version in force at that time; it directly fits because it will store the policy or guardrail version identifier with the decision record.
Incorrect Answers
Answer A is incorrect because AWS CloudTrail API audit history: records supported AWS API activity with caller, time, source, request context, and affected resources for security and compliance review. Best used when investigators need to know which identity changed a governed AI resource; T11P2Q19 instead needs bind each logged interaction to the precise control version in force at that time, so it misses the deciding constraint.
Answer B is incorrect because Prompt A/B fairness test: runs controlled prompt variants against the same evaluation set so teams can measure whether wording changes improve fairness without guessing. Best used when a prompt change is proposed specifically to reduce disparate behavior; T11P2Q19 instead needs bind each logged interaction to the precise control version in force at that time, so it misses the deciding constraint.
Answer D is incorrect because Token-level sensitive-data redaction: removes or masks sensitive spans before they are persisted or exposed, reducing the amount of regulated content in governance logs and outputs. Best used when compliance requires detailed logs without retaining raw secrets or personal identifiers; T11P2Q19 instead needs bind each logged interaction to the precise control version in force at that time, so it misses the deciding constraint.
Question 20
A new assistant version passes unit tests but has not completed fairness review or documented its model limitations. Which solution best fits this production requirement? Choose ONE.
- Bedrock LLM-as-a-judge evaluation
- AI response logging with correlation IDs
- CloudWatch decision logging
- Predeployment responsible-AI approval gate
Correct Answer: D
Correct Answer
Answer D is correct because Predeployment responsible-AI approval gate: requires documented review of model limitations, fairness evidence, safety controls, and ownership before a production release is authorized. T11P2Q20 needs prevent technical deployment success from bypassing responsible-AI governance; it directly fits because it will block production promotion until the required responsible-AI approvals and evidence are complete.
Incorrect Answers
Answer A is incorrect because Bedrock LLM-as-a-judge evaluation: uses a separate evaluator model to score generator responses against selected evaluation criteria and provide scoring explanations. Best used when large-scale qualitative assessment needs automated comparative scoring with a defined rubric; T11P2Q20 instead needs prevent technical deployment success from bypassing responsible-AI governance, so it misses the deciding constraint.
Answer B is incorrect because AI response logging with correlation IDs: records model outputs together with request, model, policy, and trace identifiers so reviewers can reconstruct a production interaction. Best used when an incident team must connect a disputed answer to the exact invocation and control configuration; T11P2Q20 instead needs prevent technical deployment success from bypassing responsible-AI governance, so it misses the deciding constraint.
Answer C is incorrect because CloudWatch decision logging: collects application decision records and operational evidence that can be queried for review, alerts, and compliance reporting. Best used when the application must retain why a policy path accepted, blocked, or escalated an AI request; T11P2Q20 instead needs prevent technical deployment success from bypassing responsible-AI governance, so it misses the deciding constraint.
