Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 35

Question 340

A regional model logs abuse-detection data outside permitted scope. Which control is required in this deployment?

Assume prompts never leave the calling Region because the S3 source bucket is local.
Rely on the model provider's brand as a substitute for reviewing current retention policies.
Verify provider-specific storage and cross-Region processing terms before use.
Route through a global profile and inspect only the initiating endpoint's Region.

Explanation

Correct Answer

Inspect exact provider/model/Region-specific input/output abuse-detection storage and cross-Region terms before processing restricted content.

Incorrect Answers

Cross-Region inference and monitoring destinations may differ from source storage.

Different products and versions may have distinct processing/storage terms.

Actual destination Regions and abuse monitoring scope determine exposure.

Question 341

A board wants traceable evidence for GenAI customer decisions. Which governance implementation produces defensible evidence or authorization?

Store only the final prose reply and its timestamp without which sources or model produced it.
Record model, prompt, knowledge source and policy version with decision correlation.
Keep unversioned prompt templates and assume their future contents explain past actions.
Record only a general agent name with no actor/request correlation.

Explanation

Correct Answer

Correlate each decision with the exact model/prompt version, knowledge-source revision, applicable policy and executed result to support replay and review.

Incorrect Answers

A fluent output cannot reconstruct the decision evidence or policy state.

A mutable template can differ from what ran at decision time.

The organization cannot trace the particular customer decision to its accountable execution.

Question 342

A knowledge base loses original document author metadata. Which action resolves this policy-control failure?

Index only the extracted paragraph strings and discard source metadata permanently.
Ask the model to guess the document author from writing style.
Attach an arbitrary generic organization name as author to every citation.
Preserve author/source attribution through ingestion and retrieval output.

Explanation

Correct Answer

Retain source author, document ID, origin and revision metadata through processing and cite authorized provenance for generated claims.

Incorrect Answers

A retrieval hit then cannot show who created or approved the underlying text.

Stylistic inference is not authenticated source attribution.

A false label is not evidence of actual document provenance.

Question 343

A dataset imported into a knowledge base has no lineage. What is the correct compliance or accountability change?

Store only the final vector coordinates and delete transformation inputs and run identifiers.
Use the FM to reconstruct missing ETL transformations after the fact.
Register source provenance and transformation history in Glue Data Catalog.
Label all data simply 'approved' without the underlying source and transformation identifiers.

Explanation

Correct Answer

Register source location, dataset lineage and transformation history in the appropriate catalog/governance system so indexed evidence is auditable.

Incorrect Answers

Vectors cannot prove which original source or cleaning steps produced them.

Generated histories are not reliable execution receipts.

An approval label does not establish traceable lineage.

Question 344

A model-card record is being submitted for a regulated deployment but it lacks intended use, tested populations and risk boundaries. Select THREE disclosures needed to make the model scope and limitations reviewable.

Publish only benchmark accuracy without explaining whom the benchmark represents.
Declare that every possible use case is safe because the model is hosted on AWS.
Document intended model use.
Document evaluated populations and known limitations.
Omit model ID and version, assuming the documentation remains correct after upgrades.
Document version-specific risk and operating boundaries.

Explanation

Correct Answer

Document the evaluated and approved intended tasks and prohibited uses of the model.

Describe tested populations, failure modes and known limitations with data-backed evidence.

Record version-specific risk, required oversight and operating restrictions so owners can assess deployments.

Incorrect Answers

A high aggregate score hides tested populations and known exceptions.

Infrastructure placement does not remove application-specific harms.

Model behavior and constraints change across releases.

Question 345

A compliance review cites only the Bedrock service's security controls. Which action resolves this policy-control failure?

Assume AWS cloud certification automatically certifies the customer's GenAI business workflow.
Separate shared-responsibility service assurances from application-specific compliance evidence.
Present an S3 encryption checkbox as proof that generated output is factually compliant.
Skip organization-specific risk evaluation because a managed FM is used.

Explanation

Correct Answer

Separate AWS's shared-responsibility security assurances from the application's own IAM, PII, consent, retrieval and audit obligations.

Incorrect Answers

Cloud provider controls do not validate application-specific permissions or decisions.

Encryption protects storage; it does not establish correct policy interpretation.

Managed infrastructure still leaves owner responsibilities for data and output controls.

Question 346

An organization has no owner for model rollout risk. What is the correct compliance or accountability change?

Let the generation model approve its own release after scoring itself highly.
Assign named business and technical approvers with recorded promotion decisions.
Treat a deployment pipeline success event as an authorized business promotion decision.
Record a department label but no actual approver or timestamp for model changes.

Explanation

Correct Answer

Assign named business risk owners and technical approvers who record acceptance/rejection of each promoted model/prompt version.

Incorrect Answers

A model is not an independent accountable risk owner.

Technical execution cannot replace risk signoff.

Unattributed labels are insufficient for audit or accountability.

Question 347

An agent caused a business mutation but its logs do not identify who asked for it or what the backend actually committed. Select TWO attribution and execution-receipt records required for accountability.

Store only the natural-language model reply and discard the executed tool's request ID.
Bind execution to authenticated initiating principal.
Capture actual tool arguments and authoritative outcome receipt.
Execute every tool with a shared administrator role and omit the actual actor from audit logs.
Trust the model's account name assertion instead of an authenticated initiator token.

Explanation

Correct Answer

Bind every tool action to the verified initiating identity and authorization context.

Capture the actual parameters, execution correlation and authoritative resulting state/receipt to reconstruct what happened.

Incorrect Answers

Prose is not evidence of which operation committed.

The action cannot be attributed to the user who authorized it.

An unverified model field cannot identify who requested the operation.

Question 348

A compliance audit finds that the citations displayed by a policy-answering application omit the effective date and document revision, even though retrieval already selects the correct current policy. Which evidence-format change makes its released answers auditable?

Use vector similarity alone and ignore the retrieved document's effective-date metadata.
Overwrite every source version and keep only the latest page URL without preserving citation lineage.
Ask the FM to infer the revision date from surrounding language.
Attach the retrieved policy revision and effective-date metadata to the source evidence and user-visible citations.

Explanation

Correct Answer

This failure is in answer evidence lineage rather than retriever selection: even correctly filtered current policies must expose their revision/effective-date metadata in the citation so auditors can reconstruct the source used.

Incorrect Answers

Relevance does not establish whether a policy revision is in force.

Without version references, earlier decisions become unreplayable.

Generated dates cannot replace authoritative source metadata.

Question 349

A firm needs to show who changed a guardrail policy. What is the correct compliance or accountability change?

Store only the current guardrail filter text and erase deployment history.
Rely on generation output logs to infer who created a Guardrail version.
Use CloudTrail/control-plane audit events and policy version history.
Mark every guardrail update as made by the service account without recording the initiating user.

Explanation

Correct Answer

Use AWS CloudTrail control-plane records and immutable guardrail version history to trace which authorized principal changed the moderation policy.

Incorrect Answers

Current configuration cannot identify the earlier actor or modification.

Inference activity is distinct from administrative policy-change events.

Shared labels obscure actual operator accountability.

Question 350

An application needs to identify drift in retrieval relevance. Which governance implementation produces defensible evidence or authorization?

Use vector index uptime alone as proof of retrieval relevance.
Compare different unrelated test questions between versions and attribute score differences to drift.
Refresh source documents without recording which version/index produced prior benchmark results.
Monitor golden-set recall trends and index/model changes with alerts.

Explanation

Correct Answer

Track golden-set retrieval recall/precision across source/index and embedding-model versions, alerting when reproducible quality declines.

Incorrect Answers

A healthy index can return systematically wrong documents.

Confounded populations do not establish regression.

Lack of versioned evidence prevents diagnosis of relevance changes.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 340

Explanation

Correct Answer

Incorrect Answers

Question 341

Explanation

Correct Answer

Incorrect Answers

Question 342

Explanation

Correct Answer

Incorrect Answers

Question 343

Explanation

Correct Answer

Incorrect Answers

Question 344

Explanation

Correct Answer

Incorrect Answers

Question 345

Explanation

Correct Answer

Incorrect Answers

Question 346

Explanation

Correct Answer

Incorrect Answers

Question 347

Explanation

Correct Answer

Incorrect Answers

Question 348

Explanation

Correct Answer

Incorrect Answers

Question 349

Explanation

Correct Answer

Incorrect Answers

Question 350

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests