Topic 01 Practice Test 2 covers GenAI Architecture, Business Requirements, FM Selection, Provider Abstraction, and Resilience in Domain 1 (Tasks 1.1-1.2). For broader exam preparation, review the AIP-C01 Exam Dumps. Every option includes focused technical reasoning tied to the tested decision.
Question 1
A finance assistant will call internal calculation APIs. The business requires deterministic authorization around tool use and independent scaling of the API layer and FM layer. Which architecture principle best fits? Choose ONE.
- Embed all authorization rules only in the system prompt
- Keep authorization and business controls outside the FM
- Expose the calculation APIs directly to end users
- Train the FM to memorize calculation results
Correct Answer: B
Correct Answer
Answer B is correct because Keep authorization and business controls outside the FM fits separation of deterministic application controls. Separating deterministic controls from probabilistic generation lets each layer scale and be governed independently. Here, separation of deterministic application controls is the governing criterion. Keep authorization and business controls outside the FM therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Embed all authorization rules only in the system prompt primarily concerns asking the model to enforce access rules probabilistically. This scenario needs separation of deterministic application controls. Keep authorization and business controls outside the FM meets that requirement. Embed all authorization rules only in the system prompt does not, because its focus remains asking the model to enforce access rules probabilistically.
Answer C is incorrect because Expose the calculation APIs directly to end users primarily concerns removing the controlled application mediation layer. This scenario needs separation of deterministic application controls. Keep authorization and business controls outside the FM meets that requirement. Expose the calculation APIs directly to end users does not, because its focus remains removing the controlled application mediation layer.
Answer D is incorrect because Train the FM to memorize calculation results primarily concerns encoding changing business outputs into model parameters. This scenario needs separation of deterministic application controls. Keep authorization and business controls outside the FM meets that requirement. Train the FM to memorize calculation results does not, because its focus remains encoding changing business outputs into model parameters.
Question 2
A team is comparing two GenAI designs: direct FM prompting and retrieval-augmented generation. The key business question is whether grounded answers justify the added complexity. What should the proof of concept compare? Choose ONE.
- Only the number of AWS services used
- Only developer preference
- Compare answer quality with and without retrieval
- Only maximum output length
Correct Answer: C
Correct Answer
Answer C is correct because Compare answer quality with and without retrieval fits controlled architecture comparison. Holding the evaluation set constant isolates the business value of adding retrieved evidence. Here, controlled architecture comparison is the governing criterion. Compare answer quality with and without retrieval therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Only the number of AWS services used primarily concerns counting components rather than measuring outcome quality. This scenario needs controlled architecture comparison. Compare answer quality with and without retrieval meets that requirement. Only the number of AWS services used does not, because its focus remains counting components rather than measuring outcome quality.
Answer B is incorrect because Only developer preference primarily concerns using opinion instead of empirical evidence. This scenario needs controlled architecture comparison. Compare answer quality with and without retrieval meets that requirement. Only developer preference does not, because its focus remains using opinion instead of empirical evidence.
Answer D is incorrect because Only maximum output length primarily concerns measuring verbosity rather than grounded-answer value. This scenario needs controlled architecture comparison. Compare answer quality with and without retrieval meets that requirement. Only maximum output length does not, because its focus remains measuring verbosity rather than grounded-answer value.
Question 3
An organization wants every GenAI service to emit the same trace identifiers and invocation metadata regardless of the selected model. What is the best platform approach? Choose ONE.
- Document the convention but require manual implementation each time
- Fine-tune all models to mention trace IDs
- Use a shared telemetry-enforcing invocation wrapper
- Use one prompt template for every application
Correct Answer: C
Correct Answer
Answer C is correct because Use a shared telemetry-enforcing invocation wrapper fits standardized technical component for observability. A reusable component can standardize cross-cutting behavior while allowing different models underneath. Here, standardized technical component for observability is the governing criterion. Use a shared telemetry-enforcing invocation wrapper therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Document the convention but require manual implementation each time primarily concerns relying on repeated custom work that can drift. This scenario needs standardized technical component for observability. Use a shared telemetry-enforcing invocation wrapper meets that requirement. Document the convention but require manual implementation each time does not, because its focus remains relying on repeated custom work that can drift.
Answer B is incorrect because Fine-tune all models to mention trace IDs primarily concerns changing generated text rather than application telemetry. This scenario needs standardized technical component for observability. Use a shared telemetry-enforcing invocation wrapper meets that requirement. Fine-tune all models to mention trace IDs does not, because its focus remains changing generated text rather than application telemetry.
Answer D is incorrect because Use one prompt template for every application primarily concerns standardizing prompts without enforcing operational metadata. This scenario needs standardized technical component for observability. Use a shared telemetry-enforcing invocation wrapper meets that requirement. Use one prompt template for every application does not, because its focus remains standardizing prompts without enforcing operational metadata.
Question 4
A business case requires the lowest inference cost among models that meet a defined quality threshold. What should the team compare? Choose ONE.
- Compare quality-adjusted cost on the workload
- Raw price per token only
- Model parameter count only
- Training dataset size only
Correct Answer: A
Correct Answer
Answer A is correct because Compare quality-adjusted cost on the workload fits cost-quality trade-off benchmarking. The selection objective is economic efficiency after the minimum quality requirement is satisfied. Here, cost-quality trade-off benchmarking is the governing criterion. Compare quality-adjusted cost on the workload therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Raw price per token only primarily concerns ignoring whether cheaper output meets the required quality. This scenario needs cost-quality trade-off benchmarking. Compare quality-adjusted cost on the workload meets that requirement. Raw price per token only does not, because its focus remains ignoring whether cheaper output meets the required quality.
Answer C is incorrect because Model parameter count only primarily concerns using model size as a proxy for value. This scenario needs cost-quality trade-off benchmarking. Compare quality-adjusted cost on the workload meets that requirement. Model parameter count only does not, because its focus remains using model size as a proxy for value.
Answer D is incorrect because Training dataset size only primarily concerns using a background attribute that does not measure application economics. This scenario needs cost-quality trade-off benchmarking. Compare quality-adjusted cost on the workload meets that requirement. Training dataset size only does not, because its focus remains using a background attribute that does not measure application economics.
Question 5
A company exposes one internal GenAI endpoint to many applications and wants routing changes to occur centrally. Which AWS pattern best fits? Choose ONE.
- Use API Gateway with a centralized routing layer
- Give every application direct credentials for each provider
- Publish model IDs in a spreadsheet
- Require each team to fork the routing code
Correct Answer: A
Correct Answer
Answer A is correct because Use API Gateway with a centralized routing layer fits centralized model-routing facade. A stable front door plus centralized routing isolates consumers from provider changes. Here, centralized model-routing facade is the governing criterion. Use API Gateway with a centralized routing layer therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Give every application direct credentials for each provider primarily concerns distributing provider configuration and access management to every client. This scenario needs centralized model-routing facade. Use API Gateway with a centralized routing layer meets that requirement. Give every application direct credentials for each provider does not, because its focus remains distributing provider configuration and access management to every client.
Answer C is incorrect because Publish model IDs in a spreadsheet primarily concerns sharing configuration manually rather than enforcing it at runtime. This scenario needs centralized model-routing facade. Use API Gateway with a centralized routing layer meets that requirement. Publish model IDs in a spreadsheet does not, because its focus remains sharing configuration manually rather than enforcing it at runtime.
Answer D is incorrect because Require each team to fork the routing code primarily concerns creating multiple independently drifting implementations. This scenario needs centralized model-routing facade. Use API Gateway with a centralized routing layer meets that requirement. Require each team to fork the routing code does not, because its focus remains creating multiple independently drifting implementations.
Question 6
A critical GenAI workflow uses a customized model that cannot be served in the secondary Region. The recovery target requires regional failover. What should the architecture team address? Choose ONE.
- Replicate only application logs to the recovery Region
- Predeploy approved recovery-region model capacity
- Increase primary-Region provisioned throughput
- Change the prompt wording during failover
Correct Answer: B
Correct Answer
Answer B is correct because Predeploy approved recovery-region model capacity fits cross-region model availability planning. Regional resilience requires usable inference capacity in the target region, not just application replication. Here, cross-region model availability planning is the governing criterion. Predeploy approved recovery-region model capacity therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Replicate only application logs to the recovery Region primarily concerns protecting observability data without providing model capacity. This scenario needs cross-region model availability planning. Predeploy approved recovery-region model capacity meets that requirement. Replicate only application logs to the recovery Region does not, because its focus remains protecting observability data without providing model capacity.
Answer C is incorrect because Increase primary-Region provisioned throughput primarily concerns adding capacity without creating a recovery path. This scenario needs cross-region model availability planning. Predeploy approved recovery-region model capacity meets that requirement. Increase primary-Region provisioned throughput does not, because its focus remains adding capacity without creating a recovery path.
Answer D is incorrect because Change the prompt wording during failover primarily concerns altering content without solving model availability. This scenario needs cross-region model availability planning. Predeploy approved recovery-region model capacity meets that requirement. Change the prompt wording during failover does not, because its focus remains altering content without solving model availability.
Question 7
A newly deployed customized model causes a sharp quality regression. Which lifecycle capability should already be in place? Choose ONE.
- Delete all prior model versions before release
- Increase request timeout
- Maintain a tested rollback to the approved version
- Disable evaluation metrics
Correct Answer: C
Correct Answer
Answer C is correct because Maintain a tested rollback to the approved version fits model deployment rollback. Production model lifecycle management should support rapid restoration when a release fails. Here, model deployment rollback is the governing criterion. Maintain a tested rollback to the approved version therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Delete all prior model versions before release primarily concerns removing the known-good artifact needed for rollback. This scenario needs model deployment rollback. Maintain a tested rollback to the approved version meets that requirement. Delete all prior model versions before release does not, because its focus remains removing the known-good artifact needed for rollback.
Answer B is incorrect because Increase request timeout primarily concerns changing patience for responses rather than restoring model quality. This scenario needs model deployment rollback. Maintain a tested rollback to the approved version meets that requirement. Increase request timeout does not, because its focus remains changing patience for responses rather than restoring model quality.
Answer D is incorrect because Disable evaluation metrics primarily concerns removing the evidence that exposed the regression. This scenario needs model deployment rollback. Maintain a tested rollback to the approved version meets that requirement. Disable evaluation metrics does not, because its focus remains removing the evidence that exposed the regression.
Question 8
A team is designing a GenAI search assistant for internal engineering documents. The documents change daily, but the FM itself does not need domain-specific writing style. Which architectural direction is most appropriate? Choose ONE.
- Fine-tune the model every night with all documents
- Increase output token limits
- Retrieve changing knowledge at request time
- Use a larger context window with no retrieval strategy
Correct Answer: C
Correct Answer
Answer C is correct because Retrieve changing knowledge at request time fits externalized changing knowledge. Frequently changing facts are better supplied as current context than repeatedly embedded into model weights. Here, externalized changing knowledge is the governing criterion. Retrieve changing knowledge at request time therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Fine-tune the model every night with all documents primarily concerns using model customization for rapidly changing factual content. This scenario needs externalized changing knowledge. Retrieve changing knowledge at request time meets that requirement. Fine-tune the model every night with all documents does not, because its focus remains using model customization for rapidly changing factual content.
Answer B is incorrect because Increase output token limits primarily concerns allowing longer responses without improving factual freshness. This scenario needs externalized changing knowledge. Retrieve changing knowledge at request time meets that requirement. Increase output token limits does not, because its focus remains allowing longer responses without improving factual freshness.
Answer D is incorrect because Use a larger context window with no retrieval strategy primarily concerns expanding capacity without defining how current evidence is selected. This scenario needs externalized changing knowledge. Retrieve changing knowledge at request time meets that requirement. Use a larger context window with no retrieval strategy does not, because its focus remains expanding capacity without defining how current evidence is selected.
Question 9
A proposed GenAI assistant may reduce case-handling time, but the expected savings are uncertain. Which proof-of-concept measure best addresses business value? Choose ONE.
- Track only model parameter count
- Count prompt tokens without a baseline
- Test only a synthetic happy-path question
- Compare task time and acceptable-quality rate
Correct Answer: D
Correct Answer
Answer D is correct because Compare task time and acceptable-quality rate fits business-value validation. The experiment should quantify whether the new approach improves the business process while maintaining required quality. Here, business-value validation is the governing criterion. Compare task time and acceptable-quality rate therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Track only model parameter count primarily concerns measuring model size rather than business impact. This scenario needs business-value validation. Compare task time and acceptable-quality rate meets that requirement. Track only model parameter count does not, because its focus remains measuring model size rather than business impact.
Answer B is incorrect because Count prompt tokens without a baseline primarily concerns collecting cost data without comparing workflow value. This scenario needs business-value validation. Compare task time and acceptable-quality rate meets that requirement. Count prompt tokens without a baseline does not, because its focus remains collecting cost data without comparing workflow value.
Answer C is incorrect because Test only a synthetic happy-path question primarily concerns using evidence too narrow to support the business decision. This scenario needs business-value validation. Compare task time and acceptable-quality rate meets that requirement. Test only a synthetic happy-path question does not, because its focus remains using evidence too narrow to support the business decision.
Question 10
A central platform team wants product squads to follow the same retry, timeout, and logging policy while still choosing different FMs. Which solution is most appropriate? Choose ONE.
- Require identical application codebases
- Use a reusable platform policy component
- Put retry instructions in user prompts
- Use the same model temperature everywhere
Correct Answer: B
Correct Answer
Answer B is correct because Use a reusable platform policy component fits reusable cross-cutting platform control. Common operational behavior can be enforced independently of the model selected by each product. Here, reusable cross-cutting platform control is the governing criterion. Use a reusable platform policy component therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Require identical application codebases primarily concerns coupling products that may have different business logic. This scenario needs reusable cross-cutting platform control. Use a reusable platform policy component meets that requirement. Require identical application codebases does not, because its focus remains coupling products that may have different business logic.
Answer C is incorrect because Put retry instructions in user prompts primarily concerns asking an FM to control network behavior it does not own. This scenario needs reusable cross-cutting platform control. Use a reusable platform policy component meets that requirement. Put retry instructions in user prompts does not, because its focus remains asking an FM to control network behavior it does not own.
Answer D is incorrect because Use the same model temperature everywhere primarily concerns standardizing generation randomness rather than resiliency policy. This scenario needs reusable cross-cutting platform control. Use a reusable platform policy component meets that requirement. Use the same model temperature everywhere does not, because its focus remains standardizing generation randomness rather than resiliency policy.
Question 11
An application must summarize contracts that can exceed the smallest candidate model context window. Which limitation should be evaluated before selection? Choose ONE.
- The number of IAM users in the account
- The S3 storage class of archived logs
- Verify context capacity against expected input size
- The color of the application interface
Correct Answer: C
Correct Answer
Answer C is correct because Verify context capacity against expected input size fits context-capacity limitation analysis. The model must be able to accept the intended prompt and document context without unsupported truncation. Here, context-capacity limitation analysis is the governing criterion. Verify context capacity against expected input size therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because The number of IAM users in the account primarily concerns an identity count unrelated to model input capacity. This scenario needs context-capacity limitation analysis. Verify context capacity against expected input size meets that requirement. The number of IAM users in the account does not, because its focus remains an identity count unrelated to model input capacity.
Answer B is incorrect because The S3 storage class of archived logs primarily concerns a storage decision unrelated to inference context limits. This scenario needs context-capacity limitation analysis. Verify context capacity against expected input size meets that requirement. The S3 storage class of archived logs does not, because its focus remains a storage decision unrelated to inference context limits.
Answer D is incorrect because The color of the application interface primarily concerns a presentation choice unrelated to model capability. This scenario needs context-capacity limitation analysis. Verify context capacity against expected input size meets that requirement. The color of the application interface does not, because its focus remains a presentation choice unrelated to model capability.
Question 12
A routing policy should send simple requests to a lower-cost model and complex requests to a more capable model. The policy may change weekly. Which design best supports this? Choose ONE.
- Duplicate the entire application for each model
- Let end users edit model SDK calls
- Use a fixed provider chosen at compile time
- Use externally configured policy-based model routing
Correct Answer: D
Correct Answer
Answer D is correct because Use externally configured policy-based model routing fits dynamic policy-based model selection. The application can classify requests and change model mapping without modifying business workflows. Here, dynamic policy-based model selection is the governing criterion. Use externally configured policy-based model routing therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Duplicate the entire application for each model primarily concerns creating separate stacks instead of one controlled routing decision. This scenario needs dynamic policy-based model selection. Use externally configured policy-based model routing meets that requirement. Duplicate the entire application for each model does not, because its focus remains creating separate stacks instead of one controlled routing decision.
Answer B is incorrect because Let end users edit model SDK calls primarily concerns exposing implementation details and weakening governance. This scenario needs dynamic policy-based model selection. Use externally configured policy-based model routing meets that requirement. Let end users edit model SDK calls does not, because its focus remains exposing implementation details and weakening governance.
Answer C is incorrect because Use a fixed provider chosen at compile time primarily concerns preventing the requested dynamic routing behavior. This scenario needs dynamic policy-based model selection. Use externally configured policy-based model routing meets that requirement. Use a fixed provider chosen at compile time does not, because its focus remains preventing the requested dynamic routing behavior.
Question 13
A Step Functions workflow calls an FM and then three downstream business services. When the FM is unhealthy, the team wants controlled retries followed by a fallback branch. Which design is most appropriate? Choose ONE.
- Model retries and fallback explicitly in Step Functions
- Ask the FM to decide whether its own endpoint is healthy
- Remove all timeouts
- Use a single unbounded retry loop
Correct Answer: A
Correct Answer
Answer A is correct because Model retries and fallback explicitly in Step Functions fits orchestrated failure handling. Orchestration can make the failure policy deterministic and visible rather than relying on ad hoc caller behavior. Here, orchestrated failure handling is the governing criterion. Model retries and fallback explicitly in Step Functions therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Ask the FM to decide whether its own endpoint is healthy primarily concerns delegating infrastructure health logic to the failing dependency. This scenario needs orchestrated failure handling. Model retries and fallback explicitly in Step Functions meets that requirement. Ask the FM to decide whether its own endpoint is healthy does not, because its focus remains delegating infrastructure health logic to the failing dependency.
Answer C is incorrect because Remove all timeouts primarily concerns allowing calls to hang instead of enforcing recovery behavior. This scenario needs orchestrated failure handling. Model retries and fallback explicitly in Step Functions meets that requirement. Remove all timeouts does not, because its focus remains allowing calls to hang instead of enforcing recovery behavior.
Answer D is incorrect because Use a single unbounded retry loop primarily concerns risking request storms and excessive recovery delay. This scenario needs orchestrated failure handling. Model retries and fallback explicitly in Step Functions meets that requirement. Use a single unbounded retry loop does not, because its focus remains risking request storms and excessive recovery delay.
Question 14
A customized model version is no longer approved because its underlying policy domain has changed. What is the appropriate lifecycle action? Choose ONE.
- Leave the version routable indefinitely
- Rename the endpoint only
- Increase the context window
- Retire the obsolete version and migrate traffic
Correct Answer: D
Correct Answer
Answer D is correct because Retire the obsolete version and migrate traffic fits controlled model retirement. Model lifecycle management includes controlled retirement so outdated versions stop serving production requests. Here, controlled model retirement is the governing criterion. Retire the obsolete version and migrate traffic therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Leave the version routable indefinitely primarily concerns keeping an unapproved model available to production traffic. This scenario needs controlled model retirement. Retire the obsolete version and migrate traffic meets that requirement. Leave the version routable indefinitely does not, because its focus remains keeping an unapproved model available to production traffic.
Answer B is incorrect because Rename the endpoint only primarily concerns changing a label without enforcing retirement. This scenario needs controlled model retirement. Retire the obsolete version and migrate traffic meets that requirement. Rename the endpoint only does not, because its focus remains changing a label without enforcing retirement.
Answer C is incorrect because Increase the context window primarily concerns altering inference capacity rather than version approval. This scenario needs controlled model retirement. Retire the obsolete version and migrate traffic meets that requirement. Increase the context window does not, because its focus remains altering inference capacity rather than version approval.
Question 15
An executive sponsor wants a GenAI assistant but has not defined acceptable answer quality, response latency, or cost per request. What should the technical lead establish before finalizing the solution architecture? Choose ONE.
- Select a provider based on brand recognition
- Lock the production instance size
- Create a rollback plan only
- Define measurable quality, latency, and cost criteria
Correct Answer: D
Correct Answer
Answer D is correct because Define measurable quality, latency, and cost criteria fits measurable nonfunctional acceptance criteria. Architecture trade-offs cannot be evaluated consistently until business success metrics and technical thresholds are explicit. Here, measurable nonfunctional acceptance criteria is the governing criterion. Define measurable quality, latency, and cost criteria therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Select a provider based on brand recognition primarily concerns choosing technology before objective acceptance measures exist. This scenario needs measurable nonfunctional acceptance criteria. Define measurable quality, latency, and cost criteria meets that requirement. Select a provider based on brand recognition does not, because its focus remains choosing technology before objective acceptance measures exist.
Answer B is incorrect because Lock the production instance size primarily concerns committing capacity before workload and service targets are understood. This scenario needs measurable nonfunctional acceptance criteria. Define measurable quality, latency, and cost criteria meets that requirement. Lock the production instance size does not, because its focus remains committing capacity before workload and service targets are understood.
Answer C is incorrect because Create a rollback plan only primarily concerns preparing for deployment failure without defining success. This scenario needs measurable nonfunctional acceptance criteria. Define measurable quality, latency, and cost criteria meets that requirement. Create a rollback plan only does not, because its focus remains preparing for deployment failure without defining success.
Question 16
A multilingual support bot must perform well in Japanese and German. Which evaluation provides the strongest selection evidence? Choose ONE.
- Benchmark target-language support conversations
- Use only an English benchmark
- Count the number of languages in marketing material
- Select by maximum output length
Correct Answer: A
Correct Answer
Answer A is correct because Benchmark target-language support conversations fits language-specific capability evaluation. Language performance should be measured directly in the languages and tasks the application will serve. Here, language-specific capability evaluation is the governing criterion. Benchmark target-language support conversations therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Use only an English benchmark primarily concerns assuming English performance transfers to every target language. This scenario needs language-specific capability evaluation. Benchmark target-language support conversations meets that requirement. Use only an English benchmark does not, because its focus remains assuming English performance transfers to every target language.
Answer C is incorrect because Count the number of languages in marketing material primarily concerns using a broad claim instead of workload evidence. This scenario needs language-specific capability evaluation. Benchmark target-language support conversations meets that requirement. Count the number of languages in marketing material does not, because its focus remains using a broad claim instead of workload evidence.
Answer D is incorrect because Select by maximum output length primarily concerns measuring response capacity rather than target-language quality. This scenario needs language-specific capability evaluation. Benchmark target-language support conversations meets that requirement. Select by maximum output length does not, because its focus remains measuring response capacity rather than target-language quality.
Question 17
A provider changes an inference field name. The organization wants only one component to require modification. What architecture provides that containment? Choose ONE.
- Use a provider adapter behind a common interface
- Direct provider calls in every microservice
- A shared prompt file with no adapter
- A DNS alias alone
Correct Answer: A
Correct Answer
Answer A is correct because Use a provider adapter behind a common interface fits localized provider-change isolation. Schema translation is localized so upstream application services remain unchanged. Here, localized provider-change isolation is the governing criterion. Use a provider adapter behind a common interface therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Direct provider calls in every microservice primarily concerns requiring many services to change when the provider contract changes. This scenario needs localized provider-change isolation. Use a provider adapter behind a common interface meets that requirement. Direct provider calls in every microservice does not, because its focus remains requiring many services to change when the provider contract changes.
Answer C is incorrect because A shared prompt file with no adapter primarily concerns centralizing text but not API schema translation. This scenario needs localized provider-change isolation. Use a provider adapter behind a common interface meets that requirement. A shared prompt file with no adapter does not, because its focus remains centralizing text but not API schema translation.
Answer D is incorrect because A DNS alias alone primarily concerns changing endpoint resolution without normalizing request semantics. This scenario needs localized provider-change isolation. Use a provider adapter behind a common interface meets that requirement. A DNS alias alone does not, because its focus remains changing endpoint resolution without normalizing request semantics.
Question 18
A team uses adapter-based tuning for several departments and wants to preserve a common base model while loading department-specific adaptations. Which approach best fits? Choose ONE.
- Merge every department into one untracked artifact
- Use DNS records as adapter versions
- Store adapter choice only in user messages
- Version base models and adapters separately
Correct Answer: D
Correct Answer
Answer D is correct because Version base models and adapters separately fits adapter lifecycle management. Adapter-based customization works best when base and adaptation versions can be tracked and promoted deliberately. Here, adapter lifecycle management is the governing criterion. Version base models and adapters separately therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Merge every department into one untracked artifact primarily concerns eliminating separation and version traceability. This scenario needs adapter lifecycle management. Version base models and adapters separately meets that requirement. Merge every department into one untracked artifact does not, because its focus remains eliminating separation and version traceability.
Answer B is incorrect because Use DNS records as adapter versions primarily concerns using networking configuration as model metadata. This scenario needs adapter lifecycle management. Version base models and adapters separately meets that requirement. Use DNS records as adapter versions does not, because its focus remains using networking configuration as model metadata.
Answer C is incorrect because Store adapter choice only in user messages primarily concerns letting request content determine unmanaged lifecycle state. This scenario needs adapter lifecycle management. Version base models and adapters separately meets that requirement. Store adapter choice only in user messages does not, because its focus remains letting request content determine unmanaged lifecycle state.
Question 19
A regulated workflow requires a model that can be used in an approved AWS Region. Which selection check is mandatory? Choose ONE.
- Assume every model is offered in every Region
- Verify model availability in the required Region
- Select first and request an exception later
- Use a larger prompt to compensate
Correct Answer: B
Correct Answer
Answer B is correct because Verify model availability in the required Region fits regional-availability model selection. Regional availability is a hard deployment constraint for the regulated workload. Here, regional-availability model selection is the governing criterion. Verify model availability in the required Region therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Assume every model is offered in every Region primarily concerns ignoring service and model regional differences. This scenario needs regional-availability model selection. Verify model availability in the required Region meets that requirement. Assume every model is offered in every Region does not, because its focus remains ignoring service and model regional differences.
Answer C is incorrect because Select first and request an exception later primarily concerns treating a known compliance boundary as optional. This scenario needs regional-availability model selection. Verify model availability in the required Region meets that requirement. Select first and request an exception later does not, because its focus remains treating a known compliance boundary as optional.
Answer D is incorrect because Use a larger prompt to compensate primarily concerns changing input text without solving availability. This scenario needs regional-availability model selection. Verify model availability in the required Region meets that requirement. Use a larger prompt to compensate does not, because its focus remains changing input text without solving availability.
Question 20
A model deployment pipeline can promote versions but does not record which evaluated artifact reached production. What improvement is most important? Choose ONE.
- Add more examples to the system prompt
- Tie promotion to a versioned registry record
- Increase endpoint instance count
- Change the endpoint URL for every request
Correct Answer: B
Correct Answer
Answer B is correct because Tie promotion to a versioned registry record fits traceable model release provenance. The production release should be traceable to the exact evaluated and approved artifact. Here, traceable model release provenance is the governing criterion. Tie promotion to a versioned registry record therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Add more examples to the system prompt primarily concerns improving prompt behavior without release traceability. This scenario needs traceable model release provenance. Tie promotion to a versioned registry record meets that requirement. Add more examples to the system prompt does not, because its focus remains improving prompt behavior without release traceability.
Answer C is incorrect because Increase endpoint instance count primarily concerns adding capacity without identifying the deployed artifact. This scenario needs traceable model release provenance. Tie promotion to a versioned registry record meets that requirement. Increase endpoint instance count does not, because its focus remains adding capacity without identifying the deployed artifact.
Answer D is incorrect because Change the endpoint URL for every request primarily concerns creating routing churn rather than lifecycle evidence. This scenario needs traceable model release provenance. Tie promotion to a versioned registry record meets that requirement. Change the endpoint URL for every request does not, because its focus remains creating routing churn rather than lifecycle evidence.
