Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 4

Question 30

Two Bedrock foundation models have very different parameter counts. The proposed procurement decision is based only on which has more parameters. What evaluation is needed?

Select the model with the largest published parameter count and skip task testing.
Use different prompts and workloads for each model and compare the cheapest request.
Compare the number of available model cards without invoking either version.
Compare both available model versions on the same representative labeled tasks, using quality, p95 latency, token cost, model restrictions and permitted Regions.

Explanation

Correct Answer

Parameter count does not directly rank accuracy, cost or latency on a particular workload. A comparable evaluation population and regional/model constraints determine feasibility and business utility.

Incorrect Answers

Size is not evidence of superior task accuracy or an acceptable latency/cost envelope.

Changing evaluation inputs confounds the comparison and hides accuracy, throughput and completion-cost differences.

Catalog breadth does not demonstrate whether either candidate meets the application acceptance thresholds.

Question 31

A frontend hardcodes six foundation-model identifiers; deployments now require UI code changes when a model version or inference profile is rotated. Which integration boundary fixes this?

Copy every possible model identifier into a larger frontend lookup table.
Move model IDs, allowed inference profiles and routing rules into an approved server-side configuration, with a stable application gateway.
Add retry loops without changing how clients choose model IDs.
Use a vector index containing model descriptions as a substitute for a governed routing configuration.

Explanation

Correct Answer

Server-side governed routing decouples model selections from frontend binaries, permits policy-checked rotations and avoids exposing unauthorized model IDs to users.

Incorrect Answers

The UI still controls provider routing and still needs redeployment when approved versions change.

Retries cannot fix hardcoded routing or enforce an approved model catalog.

Semantic retrieval of a model name does not authorize inference destinations or maintain deployment configuration.

Question 32

A previously tested Bedrock model is unavailable after a Regional failure. An approved compatible alternate exists in another permitted Region. Select TWO runtime mechanisms needed to detect the outage and safely maintain service.

Rely on the model catalog listing alone as a real-time health signal, without evaluating recent inference calls.
Monitor actual primary model/Region inference health and errors with a time-bounded failover threshold.
Switch to a previously evaluated alternate model route and expose any documented change in supported behavior.
Increase response-token limits on the unavailable primary model until the Region recovers.
Assume global inference profiles meet every regional restriction without checking destinations.

Explanation

Correct Answer

Correct: recent request failures and latency/availability thresholds let the application detect a genuine primary failure before rerouting.

Correct: a prequalified alternate provides continuity without inventing capability equivalence; model differences must be assessed and communicated.

Incorrect Answers

A listing proves a model can be offered, not that its current regional endpoint is handling requests successfully.

Token limits do not restore an unreachable regional endpoint or create new serving capacity.

Global routing may include destinations outside the approved geography; inference profiles must be checked from the caller Region and allowed by policy.

Question 33

An EU-residency rule permits processing only in approved EU Regions. A proposed GLOBAL Bedrock inference profile might send prompts outside that geography. Which change directly satisfies the processing-location restriction?

Allow the global profile in IAM while leaving the profile's possible execution geography unchanged.
Replace it with an available EU geographic profile and verify its actual execution destinations comply with the residency boundary.
Keep the global profile but store the prompt and returned answer in an EU S3 bucket.
Encrypt the prompts during transit and assume routing beyond EU borders is irrelevant.

Explanation

Correct Answer

EU geographic inference profiles are designed to keep processing inside their supported geography, subject to verifying the precise model/Region and organizational permissions.

Incorrect Answers

An allowed IAM call does not constrain the destinations of a global profile; authorization and residency are distinct tests.

Where the output is stored is not the same as where inference was performed.

Transport encryption protects contents but does not make processing outside the allowed geography compliant.

Question 34

A Bedrock request using a specific geographic inference profile fails with an authorization error. The caller has permission to invoke the profile, but an AWS Organizations SCP denies one of its possible destination Regions. What must be checked to restore policy-compliant routing?

Allow the profile and necessary model invoke actions for all required destinations, or apply a correctly scoped inference-profile exception to the Region-deny policy.
Change the prompt wording so the model avoids destinations blocked by the SCP.
Permit only the source Region in IAM and assume destination controls are never evaluated.
Change the S3 bucket policy but leave the model invoke/SCP denial unchanged.

Explanation

Correct Answer

A profile can evaluate IAM and SCP decisions for candidate destination Regions. The policy must allow those destinations or a documented scoped inference-profile exception.

Incorrect Answers

The model has no ability to override an AWS Organizations explicit deny by following a prompt.

Destination-model authorization is also evaluated; source-only permission is not sufficient under cross-Region routing.

An S3 permission does not correct a denied Bedrock inference profile or model-invoke action.

Question 35

Production requests intermittently receive Bedrock throttling responses; some requests may trigger a downstream case update after model completion. What retry design is safe?

Use bounded exponential backoff with jitter for retryable inference errors, respect quotas and make downstream case updates idempotent before retries.
Retry every exception immediately in an unbounded loop until it succeeds.
Re-run the entire conversation including any successful case-update tool call without a deduplication key.
Raise the model's temperature to make rate-limit responses less frequent.

Explanation

Correct Answer

Retryable throttling requires limits and jitter to prevent a retry storm. Inference retries and transaction execution are different boundaries, so business side effects must use deduplication or idempotency keys.

Incorrect Answers

Tight retries amplify throttling and create unpredictable queues and costs.

A repeated tool side effect can modify the business record more than once even when the inference retry was legitimate.

Sampling settings do not increase service quotas or solve throttling.

Question 36

An organization selected a Bedrock model version that is not callable through direct on-demand inference in its required Region. It still needs a supported production invocation path. What must the developer do before deployment?

Keep the unavailable model identifier and retry it until Bedrock provisions direct on-demand access automatically.
Request a larger token quota and assume it changes the model's permitted deployment modes.
Identify a supported inference profile or offered deployment/capacity mode for that exact model version and Region.
Use InvokeModel and Converse interchangeably without checking the model's supported API interfaces.

Explanation

Correct Answer

Deployment capabilities vary by model/version and Region, and an inference profile may be the supported on-demand path.

Incorrect Answers

Retries do not turn an unsupported direct invocation mode into a supported deployment.

Quota controls allowed traffic quantities; it cannot create a serving mode that the model does not offer.

API interface compatibility also varies. An arbitrary API swap will not solve a missing serving mode.

Question 37

A team has only a handful of verified examples but wants to fine-tune a model every day as policies change. What should be evaluated before paying for customization?

Fine-tune on the handful of examples and skip held-out testing because the data are recent.
Enlarge the temperature range to encode new policy facts permanently into weights.
Store confidential policies in an unrestricted public prompt template to avoid building retrieval.
Benchmark few-shot prompting and permission-aware RAG against the baseline, then consider supported model customization only if task evaluation shows a durable gap.

Explanation

Correct Answer

Frequent source-fact changes are ordinarily better handled by current retrieval than repeated weight updates. Adaptation requires enough suitable examples and a supported model/version plus measurable gains.

Incorrect Answers

Too few examples and no independent evaluation cannot establish a generalizable improvement.

Sampling parameters affect response selection but cannot update training knowledge.

This introduces disclosure risk and does not provide source-level authorization or currency controls.

Question 38

A custom classification model has passed offline evaluation, but the deployment process cannot identify the exact approved artifact or roll it back. What governance mechanism is required?

Deploy from a mutable local filename without recording a model package version.
Register only a free-text model name without approval state or artifact identity.
Increase serving instance count before recording artifact lineage.
Register a versioned model package with its metrics, lineage and approval state in SageMaker Model Registry, and deploy only the approved version.

Explanation

Correct Answer

SageMaker Model Registry tracks model package versions, metadata/lineage and approval status so CI/CD can promote and restore a known evaluated artifact.

Incorrect Answers

A mutable path cannot prove what was approved or enable reliable reproducible rollback.

A label alone does not bind a specific model package to evaluated metrics and release governance.

Throughput does not solve provenance or production promotion control.

Question 39

A LoRA adapter package is ready for production with a designated base model. Select TWO controls that prove the adapter/base pairing and make that validated pair recoverable after a faulty release.

Verify the LoRA adapter matches the exact base-model revision and passes load/behavior checks as that pair.
Approve any adapter that has the same architecture label, even when its base checkpoint revision is unknown.
Version and retain both adapter and base artifacts with a tested rollback to their known-good pairing.
Increase SageMaker instance count instead of recording which adapter and base were promoted.
Rely on general text-output schema validation alone to establish adapter weight compatibility.

Explanation

Correct Answer

Correct: an adapter is a model-specific weight delta; verifying the base checkpoint and successful evaluation is essential to compatibility.

Correct: immutable linked versions and a tested rollback restore the precise adapter/base combination, not an arbitrary older adapter.

Incorrect Answers

Architecture-family similarity cannot prove the exact checkpoint or target layer shapes match the adapter.

Capacity can improve serving throughput but cannot repair an incompatible adapter or prove a known-good rollback path.

A valid JSON output says nothing about whether the adapter weights match the base architecture and revision.

Question 40

A just-deployed custom model begins producing unsafe finance recommendations; the prior approved model package passed the same tests. What immediate release action is appropriate?

Increase serving instances to make the new model respond faster.
Mark the failing package as approved because its artifact can load.
Route production traffic back to the last validated model package and investigate the failing safety-regression cohort before another promotion.
Raise response-token limits to suppress unsafe advice.

Explanation

Correct Answer

A model-quality regression requires restoring the verified version rather than merely scaling. Versioned SageMaker deployment and evaluation evidence allow safe rollback and root-cause analysis.

Incorrect Answers

Additional capacity does not correct unsafe generated content.

Loading successfully is not the same as passing safety acceptance; a failing model should not be promoted.

Output length does not enforce safe recommendations or resolve a detected behavioral regression.

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

Question 30

Explanation

Correct Answer

Incorrect Answers

Question 31

Explanation

Correct Answer

Incorrect Answers

Question 32

Explanation

Correct Answer

Incorrect Answers

Question 33

Explanation

Correct Answer

Incorrect Answers

Question 34

Explanation

Correct Answer

Incorrect Answers

Question 35

Explanation

Correct Answer

Incorrect Answers

Question 36

Explanation

Correct Answer

Incorrect Answers

Question 37

Explanation

Correct Answer

Incorrect Answers

Question 38

Explanation

Correct Answer

Incorrect Answers

Question 39

Explanation

Correct Answer

Incorrect Answers

Question 40

Explanation

Correct Answer

Incorrect Answers

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