Topic 01 Practice Test 3 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 healthcare workflow can tolerate a slower answer when the primary GenAI path is unavailable, but it must never skip mandatory rule-based eligibility checks. Which design is most appropriate? Choose ONE.
- Encode eligibility policy only in a prompt
- Keep mandatory checks outside the FM with degradation
- Disable all processing when the FM is unavailable
- Raise model temperature during outages
Correct Answer: B
Correct Answer
Answer B is correct because Keep mandatory checks outside the FM with degradation fits resilient separation of mandatory rules. The rules remain enforceable even when the generative component changes or falls back. Here, resilient separation of mandatory rules is the governing criterion. Keep mandatory checks outside the FM with degradation therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Encode eligibility policy only in a prompt primarily concerns depending on probabilistic instruction following for mandatory checks. This scenario needs resilient separation of mandatory rules. Keep mandatory checks outside the FM with degradation meets that requirement. Encode eligibility policy only in a prompt does not, because its focus remains depending on probabilistic instruction following for mandatory checks.
Answer C is incorrect because Disable all processing when the FM is unavailable primarily concerns avoiding degraded service even though the business permits it. This scenario needs resilient separation of mandatory rules. Keep mandatory checks outside the FM with degradation meets that requirement. Disable all processing when the FM is unavailable does not, because its focus remains avoiding degraded service even though the business permits it.
Answer D is incorrect because Raise model temperature during outages primarily concerns changing output diversity rather than system resilience. This scenario needs resilient separation of mandatory rules. Keep mandatory checks outside the FM with degradation meets that requirement. Raise model temperature during outages does not, because its focus remains changing output diversity rather than system resilience.
Question 2
A team has a promising demo built on ten handpicked examples. Before approving production, what change makes the feasibility evidence more credible? Choose ONE.
- Test representative normal, edge, and failure cases
- Remove difficult examples from the test set
- Increase temperature to make outputs more varied
- Approve production because the demo succeeded
Correct Answer: A
Correct Answer
Answer A is correct because Test representative normal, edge, and failure cases fits representative feasibility coverage. A broader sample tests whether the design works beyond curated examples and exposes operational limitations. Here, representative feasibility coverage is the governing criterion. Test representative normal, edge, and failure cases therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Remove difficult examples from the test set primarily concerns improving apparent results by hiding realistic failure modes. This scenario needs representative feasibility coverage. Test representative normal, edge, and failure cases meets that requirement. Remove difficult examples from the test set does not, because its focus remains improving apparent results by hiding realistic failure modes.
Answer C is incorrect because Increase temperature to make outputs more varied primarily concerns altering generation randomness rather than improving evaluation coverage. This scenario needs representative feasibility coverage. Test representative normal, edge, and failure cases meets that requirement. Increase temperature to make outputs more varied does not, because its focus remains altering generation randomness rather than improving evaluation coverage.
Answer D is incorrect because Approve production because the demo succeeded primarily concerns treating a narrow demonstration as sufficient evidence. This scenario needs representative feasibility coverage. Test representative normal, edge, and failure cases meets that requirement. Approve production because the demo succeeded does not, because its focus remains treating a narrow demonstration as sufficient evidence.
Question 3
Architecture reviews repeatedly find the same GenAI risks only after implementation. What should the organization add to improve consistency earlier? Choose ONE.
- A larger production quota
- Use a Well-Architected GenAI design checklist
- A mandatory model fine-tune
- A single shared output token limit
Correct Answer: B
Correct Answer
Answer B is correct because Use a Well-Architected GenAI design checklist fits early standardized design review. A repeatable review artifact surfaces known architecture concerns before teams commit to implementation. Here, early standardized design review is the governing criterion. Use a Well-Architected GenAI design checklist therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because A larger production quota primarily concerns adding capacity without improving design discipline. This scenario needs early standardized design review. Use a Well-Architected GenAI design checklist meets that requirement. A larger production quota does not, because its focus remains adding capacity without improving design discipline.
Answer C is incorrect because A mandatory model fine-tune primarily concerns customizing models regardless of risk findings. This scenario needs early standardized design review. Use a Well-Architected GenAI design checklist meets that requirement. A mandatory model fine-tune does not, because its focus remains customizing models regardless of risk findings.
Answer D is incorrect because A single shared output token limit primarily concerns standardizing one parameter while leaving broader risks unchecked. This scenario needs early standardized design review. Use a Well-Architected GenAI design checklist meets that requirement. A single shared output token limit does not, because its focus remains standardizing one parameter while leaving broader risks unchecked.
Question 4
A team is choosing between a general-purpose FM and a smaller specialist model for a narrow extraction task. Both meet functional requirements. What should the team evaluate next? Choose ONE.
- Always choose the general-purpose model
- Compare task accuracy, latency, and cost
- Always choose the smaller model
- Choose whichever has more documentation pages
Correct Answer: B
Correct Answer
Answer B is correct because Compare task accuracy, latency, and cost fits workload-specific model fit. The best fit depends on measured performance and economics for the actual task, not on general model prestige. Here, workload-specific model fit is the governing criterion. Compare task accuracy, latency, and cost therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Always choose the general-purpose model primarily concerns assuming breadth is always superior for a narrow task. This scenario needs workload-specific model fit. Compare task accuracy, latency, and cost meets that requirement. Always choose the general-purpose model does not, because its focus remains assuming breadth is always superior for a narrow task.
Answer C is incorrect because Always choose the smaller model primarily concerns assuming lower size always means acceptable quality. This scenario needs workload-specific model fit. Compare task accuracy, latency, and cost meets that requirement. Always choose the smaller model does not, because its focus remains assuming lower size always means acceptable quality.
Answer D is incorrect because Choose whichever has more documentation pages primarily concerns using documentation volume rather than operational evidence. This scenario needs workload-specific model fit. Compare task accuracy, latency, and cost meets that requirement. Choose whichever has more documentation pages does not, because its focus remains using documentation volume rather than operational evidence.
Question 5
An enterprise allows only two approved model providers and wants teams to use logical model tiers such as FAST and HIGH_QUALITY instead of provider names. Which design is best? Choose ONE.
- Expose raw provider model IDs in all application code
- Use prompt comments to describe the tier
- Create a separate user account for each tier
- Map logical tiers to concrete models centrally
Correct Answer: D
Correct Answer
Answer D is correct because Map logical tiers to concrete models centrally fits logical-model indirection. Applications depend on business-facing tiers while the platform controls which concrete model implements each tier. Here, logical-model indirection is the governing criterion. Map logical tiers to concrete models centrally therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Expose raw provider model IDs in all application code primarily concerns coupling every consumer to specific model identifiers. This scenario needs logical-model indirection. Map logical tiers to concrete models centrally meets that requirement. Expose raw provider model IDs in all application code does not, because its focus remains coupling every consumer to specific model identifiers.
Answer B is incorrect because Use prompt comments to describe the tier primarily concerns documenting intent without controlling routing. This scenario needs logical-model indirection. Map logical tiers to concrete models centrally meets that requirement. Use prompt comments to describe the tier does not, because its focus remains documenting intent without controlling routing.
Answer C is incorrect because Create a separate user account for each tier primarily concerns using identity objects instead of model-selection configuration. This scenario needs logical-model indirection. Map logical tiers to concrete models centrally meets that requirement. Create a separate user account for each tier does not, because its focus remains using identity objects instead of model-selection configuration.
Question 6
An assistant depends on a premium model for rich summaries, but during capacity pressure the business prefers a concise extractive response over an error. What should the design include? Choose ONE.
- Silently drop the request
- Increase sampling randomness
- Queue interactive requests for several days
- Provide an approved degraded-service response mode
Correct Answer: D
Correct Answer
Answer D is correct because Provide an approved degraded-service response mode fits business-approved graceful degradation. The system can preserve essential user value when full generative capability is temporarily unavailable. Here, business-approved graceful degradation is the governing criterion. Provide an approved degraded-service response mode therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Silently drop the request primarily concerns avoiding compute use by providing no service. This scenario needs business-approved graceful degradation. Provide an approved degraded-service response mode meets that requirement. Silently drop the request does not, because its focus remains avoiding compute use by providing no service.
Answer B is incorrect because Increase sampling randomness primarily concerns changing response variability without reducing dependency on constrained capacity. This scenario needs business-approved graceful degradation. Provide an approved degraded-service response mode meets that requirement. Increase sampling randomness does not, because its focus remains changing response variability without reducing dependency on constrained capacity.
Answer C is incorrect because Queue interactive requests for several days primarily concerns preserving full processing but violating the intended user experience. This scenario needs business-approved graceful degradation. Provide an approved degraded-service response mode meets that requirement. Queue interactive requests for several days does not, because its focus remains preserving full processing but violating the intended user experience.
Question 7
A customized model needs a phased rollout so only a small share of traffic sees the new version before full promotion. Which lifecycle practice is most appropriate? Choose ONE.
- Replace all production traffic instantly
- Disable monitoring during the rollout
- Use a staged rollout with rollback criteria
- Randomly change prompts at the same time
Correct Answer: C
Correct Answer
Answer C is correct because Use a staged rollout with rollback criteria fits phased model deployment. Limited exposure reduces release risk and provides evidence before broader promotion. Here, phased model deployment is the governing criterion. Use a staged rollout with rollback criteria therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Replace all production traffic instantly primarily concerns maximizing blast radius before validation. This scenario needs phased model deployment. Use a staged rollout with rollback criteria meets that requirement. Replace all production traffic instantly does not, because its focus remains maximizing blast radius before validation.
Answer B is incorrect because Disable monitoring during the rollout primarily concerns removing the signals needed for promotion decisions. This scenario needs phased model deployment. Use a staged rollout with rollback criteria meets that requirement. Disable monitoring during the rollout does not, because its focus remains removing the signals needed for promotion decisions.
Answer D is incorrect because Randomly change prompts at the same time primarily concerns confounding model-version evaluation with unrelated changes. This scenario needs phased model deployment. Use a staged rollout with rollback criteria meets that requirement. Randomly change prompts at the same time does not, because its focus remains confounding model-version evaluation with unrelated changes.
Question 8
A GenAI application will be used by three business units with different data sources but the same authentication, logging, and invocation patterns. Which design reduces duplicated engineering while preserving unit-specific integrations? Choose ONE.
- Build three unrelated stacks with copied code
- Use shared platform components with configurable adapters
- Use one hard-coded connector for all units
- Fine-tune a separate model for each unit by default
Correct Answer: B
Correct Answer
Answer B is correct because Use shared platform components with configurable adapters fits reusable standardized architecture components. Common controls can be standardized while source-specific behavior remains replaceable at the edges. Here, reusable standardized architecture components is the governing criterion. Use shared platform components with configurable adapters therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Build three unrelated stacks with copied code primarily concerns duplicating common implementation and operational controls. This scenario needs reusable standardized architecture components. Use shared platform components with configurable adapters meets that requirement. Build three unrelated stacks with copied code does not, because its focus remains duplicating common implementation and operational controls.
Answer C is incorrect because Use one hard-coded connector for all units primarily concerns removing the flexibility required by different source systems. This scenario needs reusable standardized architecture components. Use shared platform components with configurable adapters meets that requirement. Use one hard-coded connector for all units does not, because its focus remains removing the flexibility required by different source systems.
Answer D is incorrect because Fine-tune a separate model for each unit by default primarily concerns customizing models without evidence that model behavior differs. This scenario needs reusable standardized architecture components. Use shared platform components with configurable adapters meets that requirement. Fine-tune a separate model for each unit by default does not, because its focus remains customizing models without evidence that model behavior differs.
Question 9
A platform group wants all GenAI teams to handle configuration, secrets references, telemetry, and error envelopes consistently while allowing application-specific business logic. Which approach provides the best balance? Choose ONE.
- Provide a standardized reusable GenAI service template
- Mandate one prompt for every application
- Require every team to copy a sample repository once
- Select one model provider permanently
Correct Answer: A
Correct Answer
Answer A is correct because Provide a standardized reusable GenAI service template fits standardized deployment scaffolding. A shared template makes operational controls repeatable without forcing identical business behavior. Here, standardized deployment scaffolding is the governing criterion. Provide a standardized reusable GenAI service template therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Mandate one prompt for every application primarily concerns standardizing content instead of cross-cutting implementation. This scenario needs standardized deployment scaffolding. Provide a standardized reusable GenAI service template meets that requirement. Mandate one prompt for every application does not, because its focus remains standardizing content instead of cross-cutting implementation.
Answer C is incorrect because Require every team to copy a sample repository once primarily concerns creating independent copies that can drift after cloning. This scenario needs standardized deployment scaffolding. Provide a standardized reusable GenAI service template meets that requirement. Require every team to copy a sample repository once does not, because its focus remains creating independent copies that can drift after cloning.
Answer D is incorrect because Select one model provider permanently primarily concerns standardizing a vendor choice rather than reusable architecture controls. This scenario needs standardized deployment scaffolding. Provide a standardized reusable GenAI service template meets that requirement. Select one model provider permanently does not, because its focus remains standardizing a vendor choice rather than reusable architecture controls.
Question 10
A model produces excellent prose but frequently violates a required response schema. Another model is slightly less eloquent but reliably follows the schema. The schema is mandatory. Which model is the better fit? Choose ONE.
- The more eloquent model regardless of schema failures
- Whichever model has higher temperature defaults
- Choose the schema-reliable model
- Whichever model was adopted by another team
Correct Answer: C
Correct Answer
Answer C is correct because Choose the schema-reliable model fits constraint-first capability selection. A hard integration contract outweighs stylistic quality that does not meet the application interface requirement. Here, constraint-first capability selection is the governing criterion. Choose the schema-reliable model therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because The more eloquent model regardless of schema failures primarily concerns prioritizing style over a mandatory integration contract. This scenario needs constraint-first capability selection. Choose the schema-reliable model meets that requirement. The more eloquent model regardless of schema failures does not, because its focus remains prioritizing style over a mandatory integration contract.
Answer B is incorrect because Whichever model has higher temperature defaults primarily concerns using a generation setting unrelated to schema compliance. This scenario needs constraint-first capability selection. Choose the schema-reliable model meets that requirement. Whichever model has higher temperature defaults does not, because its focus remains using a generation setting unrelated to schema compliance.
Answer D is incorrect because Whichever model was adopted by another team primarily concerns copying a decision made for a different workload. This scenario needs constraint-first capability selection. Choose the schema-reliable model meets that requirement. Whichever model was adopted by another team does not, because its focus remains copying a decision made for a different workload.
Question 11
A platform team must rotate from one model provider to another during an incident, but it cannot rebuild client applications. Which prerequisite is most important? Choose ONE.
- Clients embed the provider hostname in every request
- Users manually rewrite prompts during incidents
- Every client stores its own fallback table
- Have clients call a stable internal contract
Correct Answer: D
Correct Answer
Answer D is correct because Have clients call a stable internal contract fits stable consumer contract for failover. A consistent consumer interface makes provider switching an implementation detail of the platform. Here, stable consumer contract for failover is the governing criterion. Have clients call a stable internal contract therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Clients embed the provider hostname in every request primarily concerns making provider identity part of each client implementation. This scenario needs stable consumer contract for failover. Have clients call a stable internal contract meets that requirement. Clients embed the provider hostname in every request does not, because its focus remains making provider identity part of each client implementation.
Answer B is incorrect because Users manually rewrite prompts during incidents primarily concerns changing content instead of infrastructure routing. This scenario needs stable consumer contract for failover. Have clients call a stable internal contract meets that requirement. Users manually rewrite prompts during incidents does not, because its focus remains changing content instead of infrastructure routing.
Answer C is incorrect because Every client stores its own fallback table primarily concerns duplicating routing policy across applications. This scenario needs stable consumer contract for failover. Have clients call a stable internal contract meets that requirement. Every client stores its own fallback table does not, because its focus remains duplicating routing policy across applications.
Question 12
A failover design has two model endpoints, but both depend on the same regional service path. Which review finding is most important? Choose ONE.
- The endpoint names are too similar
- Remove the shared regional failure domain
- The prompts should use more examples
- The model outputs should be longer
Correct Answer: B
Correct Answer
Answer B is correct because Remove the shared regional failure domain fits failure-domain independence. Duplicating endpoints inside one failure domain does not satisfy a regional-disruption requirement. Here, failure-domain independence is the governing criterion. Remove the shared regional failure domain therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because The endpoint names are too similar primarily concerns treating naming style as the main resilience concern. This scenario needs failure-domain independence. Remove the shared regional failure domain meets that requirement. The endpoint names are too similar does not, because its focus remains treating naming style as the main resilience concern.
Answer C is incorrect because The prompts should use more examples primarily concerns changing prompt quality rather than infrastructure independence. This scenario needs failure-domain independence. Remove the shared regional failure domain meets that requirement. The prompts should use more examples does not, because its focus remains changing prompt quality rather than infrastructure independence.
Answer D is incorrect because The model outputs should be longer primarily concerns changing response size rather than availability. This scenario needs failure-domain independence. Remove the shared regional failure domain meets that requirement. The model outputs should be longer does not, because its focus remains changing response size rather than availability.
Question 13
A team wants every customized model release to pass the same evaluation and approval checks before deployment. What should be standardized? Choose ONE.
- Developer memory of the preferred steps
- The model temperature for all use cases
- The number of endpoint tags only
- Gate model promotion on required evaluations
Correct Answer: D
Correct Answer
Answer D is correct because Gate model promotion on required evaluations fits standardized model promotion gates. Consistent release criteria reduce ad hoc decisions and make model deployment auditable. Here, standardized model promotion gates is the governing criterion. Gate model promotion on required evaluations therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Developer memory of the preferred steps primarily concerns relying on informal knowledge rather than an enforced process. This scenario needs standardized model promotion gates. Gate model promotion on required evaluations meets that requirement. Developer memory of the preferred steps does not, because its focus remains relying on informal knowledge rather than an enforced process.
Answer B is incorrect because The model temperature for all use cases primarily concerns standardizing one inference parameter rather than release controls. This scenario needs standardized model promotion gates. Gate model promotion on required evaluations meets that requirement. The model temperature for all use cases does not, because its focus remains standardizing one inference parameter rather than release controls.
Answer C is incorrect because The number of endpoint tags only primarily concerns recording metadata without enforcing quality gates. This scenario needs standardized model promotion gates. Gate model promotion on required evaluations meets that requirement. The number of endpoint tags only does not, because its focus remains recording metadata without enforcing quality gates.
Question 14
A team must decide whether to use a managed FM API or self-host a model. The workload has unpredictable bursts, a small operations team, and no requirement for custom inference code. Which factor should carry the most weight? Choose ONE.
- Maximum GPU utilization at steady state
- Number of prompt templates
- Prioritize operational burden and elasticity
- Length of the model name
Correct Answer: C
Correct Answer
Answer C is correct because Prioritize operational burden and elasticity fits deployment-model fit to operating constraints. Managed consumption better matches bursty demand and limited infrastructure ownership when custom serving is unnecessary. Here, deployment-model fit to operating constraints is the governing criterion. Prioritize operational burden and elasticity therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Maximum GPU utilization at steady state primarily concerns optimizing a self-hosting metric that the workload does not require. This scenario needs deployment-model fit to operating constraints. Prioritize operational burden and elasticity meets that requirement. Maximum GPU utilization at steady state does not, because its focus remains optimizing a self-hosting metric that the workload does not require.
Answer B is incorrect because Number of prompt templates primarily concerns measuring application content rather than deployment operations. This scenario needs deployment-model fit to operating constraints. Prioritize operational burden and elasticity meets that requirement. Number of prompt templates does not, because its focus remains measuring application content rather than deployment operations.
Answer D is incorrect because Length of the model name primarily concerns using a label that has no architectural significance. This scenario needs deployment-model fit to operating constraints. Prioritize operational burden and elasticity meets that requirement. Length of the model name does not, because its focus remains using a label that has no architectural significance.
Question 15
A team wants to reduce vendor lock-in but also needs a model feature available from only one provider today. What is the best selection posture? Choose ONE.
- Reject the required feature to keep every provider identical
- Hard-code provider details throughout the application
- Meet the capability behind a provider abstraction
- Choose providers randomly per request
Correct Answer: C
Correct Answer
Answer C is correct because Meet the capability behind a provider abstraction fits capability selection with portability awareness. The immediate business requirement is met while architecture limits unnecessary coupling. Here, capability selection with portability awareness is the governing criterion. Meet the capability behind a provider abstraction therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Reject the required feature to keep every provider identical primarily concerns sacrificing the stated business requirement for theoretical symmetry. This scenario needs capability selection with portability awareness. Meet the capability behind a provider abstraction meets that requirement. Reject the required feature to keep every provider identical does not, because its focus remains sacrificing the stated business requirement for theoretical symmetry.
Answer B is incorrect because Hard-code provider details throughout the application primarily concerns meeting today’s need while maximizing future switching cost. This scenario needs capability selection with portability awareness. Meet the capability behind a provider abstraction meets that requirement. Hard-code provider details throughout the application does not, because its focus remains meeting today’s need while maximizing future switching cost.
Answer D is incorrect because Choose providers randomly per request primarily concerns adding variability without a validated routing policy. This scenario needs capability selection with portability awareness. Meet the capability behind a provider abstraction meets that requirement. Choose providers randomly per request does not, because its focus remains adding variability without a validated routing policy.
Question 16
A provider-specific authentication mechanism changes, but application teams should not update their code. Where should the change be handled? Choose ONE.
- Handle authentication inside the provider connector
- In every consuming application separately
- In the user prompt sent to the model
- In the application user interface
Correct Answer: A
Correct Answer
Answer A is correct because Handle authentication inside the provider connector fits provider-specific authentication isolation. The connector owns provider-specific authentication details while consumers keep using the stable internal contract. Here, provider-specific authentication isolation is the governing criterion. Handle authentication inside the provider connector therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because In every consuming application separately primarily concerns spreading provider authentication changes into each business application. This scenario needs provider-specific authentication isolation. Handle authentication inside the provider connector meets that requirement. In every consuming application separately does not, because its focus remains spreading provider authentication changes into each business application.
Answer C is incorrect because In the user prompt sent to the model primarily concerns placing transport authentication concerns inside model content. This scenario needs provider-specific authentication isolation. Handle authentication inside the provider connector meets that requirement. In the user prompt sent to the model does not, because its focus remains placing transport authentication concerns inside model content.
Answer D is incorrect because In the application user interface primarily concerns changing presentation code rather than the provider integration boundary. This scenario needs provider-specific authentication isolation. Handle authentication inside the provider connector meets that requirement. In the application user interface does not, because its focus remains changing presentation code rather than the provider integration boundary.
Question 17
A fallback model is available, but it lacks a capability required for one high-risk transaction type. What should the resilience policy do? Choose ONE.
- Send all requests to the fallback model anyway
- Hide the transaction type from routing logic
- Restrict fallback to capability-compatible request classes
- Increase output tokens on the fallback model
Correct Answer: C
Correct Answer
Answer C is correct because Restrict fallback to capability-compatible request classes fits capability-aware fallback policy. Resilience must not bypass a hard functional or risk requirement simply to keep every request flowing. Here, capability-aware fallback policy is the governing criterion. Restrict fallback to capability-compatible request classes therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Send all requests to the fallback model anyway primarily concerns preserving availability by violating a mandatory capability constraint. This scenario needs capability-aware fallback policy. Restrict fallback to capability-compatible request classes meets that requirement. Send all requests to the fallback model anyway does not, because its focus remains preserving availability by violating a mandatory capability constraint.
Answer B is incorrect because Hide the transaction type from routing logic primarily concerns preventing the system from enforcing the known restriction. This scenario needs capability-aware fallback policy. Restrict fallback to capability-compatible request classes meets that requirement. Hide the transaction type from routing logic does not, because its focus remains preventing the system from enforcing the known restriction.
Answer D is incorrect because Increase output tokens on the fallback model primarily concerns changing response length without adding the missing capability. This scenario needs capability-aware fallback policy. Restrict fallback to capability-compatible request classes meets that requirement. Increase output tokens on the fallback model does not, because its focus remains changing response length without adding the missing capability.
Question 18
A replacement model is approved, but several applications still reference the retiring model ID directly. Which lifecycle issue should be addressed before retirement? Choose ONE.
- Delete the old model immediately
- Raise quotas on the retiring model
- Add the old model ID to prompts
- Move consumers to the replacement through abstraction
Correct Answer: D
Correct Answer
Answer D is correct because Move consumers to the replacement through abstraction fits dependency-aware model replacement. Retirement is safe only when dependent applications no longer require the obsolete concrete model identifier. Here, dependency-aware model replacement is the governing criterion. Move consumers to the replacement through abstraction therefore satisfies the stated requirement directly.
Incorrect Answers
Answer A is incorrect because Delete the old model immediately primarily concerns breaking consumers that still depend on the retired identifier. This scenario needs dependency-aware model replacement. Move consumers to the replacement through abstraction meets that requirement. Delete the old model immediately does not, because its focus remains breaking consumers that still depend on the retired identifier.
Answer B is incorrect because Raise quotas on the retiring model primarily concerns extending capacity rather than removing dependency. This scenario needs dependency-aware model replacement. Move consumers to the replacement through abstraction meets that requirement. Raise quotas on the retiring model does not, because its focus remains extending capacity rather than removing dependency.
Answer C is incorrect because Add the old model ID to prompts primarily concerns spreading the obsolete dependency further into application content. This scenario needs dependency-aware model replacement. Move consumers to the replacement through abstraction meets that requirement. Add the old model ID to prompts does not, because its focus remains spreading the obsolete dependency further into application content.
Question 19
A platform uses AWS Lambda to dispatch requests to several approved FMs. Operations wants to disable one provider immediately without redeploying Lambda. Which change best supports that? Choose ONE.
- Control enabled providers through AWS AppConfig
- Hard-code the provider list and redeploy Lambda
- Put provider availability instructions in the system prompt
- Rename the public API endpoint when a provider is disabled
Correct Answer: A
Correct Answer
Answer A is correct because Control enabled providers through AWS AppConfig fits runtime provider disablement without code deployment. Operations can update external configuration and the router can honor it at runtime. Here, runtime provider disablement without code deployment is the governing criterion. Control enabled providers through AWS AppConfig therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Hard-code the provider list and redeploy Lambda primarily concerns requiring a code release for an operational routing change. This scenario needs runtime provider disablement without code deployment. Control enabled providers through AWS AppConfig meets that requirement. Hard-code the provider list and redeploy Lambda does not, because its focus remains requiring a code release for an operational routing change.
Answer C is incorrect because Put provider availability instructions in the system prompt primarily concerns asking generation instructions to control infrastructure routing. This scenario needs runtime provider disablement without code deployment. Control enabled providers through AWS AppConfig meets that requirement. Put provider availability instructions in the system prompt does not, because its focus remains asking generation instructions to control infrastructure routing.
Answer D is incorrect because Rename the public API endpoint when a provider is disabled primarily concerns changing a consumer-facing label without altering provider eligibility. This scenario needs runtime provider disablement without code deployment. Control enabled providers through AWS AppConfig meets that requirement. Rename the public API endpoint when a provider is disabled does not, because its focus remains changing a consumer-facing label without altering provider eligibility.
Question 20
A fine-tuned model artifact is approved for production, but the team cannot reproduce which base model and adaptation configuration produced it. What should be fixed? Choose ONE.
- Record lineage for base, adaptation, and release
- Change the endpoint DNS name
- Increase logging of user prompts only
- Select a different AWS Region
Correct Answer: A
Correct Answer
Answer A is correct because Record lineage for base, adaptation, and release fits customized-model lineage. Reproducible lifecycle management requires knowing exactly how an approved customized model was derived. Here, customized-model lineage is the governing criterion. Record lineage for base, adaptation, and release therefore satisfies the stated requirement directly.
Incorrect Answers
Answer B is incorrect because Change the endpoint DNS name primarily concerns altering routing without reconstructing model provenance. This scenario needs customized-model lineage. Record lineage for base, adaptation, and release meets that requirement. Change the endpoint DNS name does not, because its focus remains altering routing without reconstructing model provenance.
Answer C is incorrect because Increase logging of user prompts only primarily concerns capturing runtime input without model-build lineage. This scenario needs customized-model lineage. Record lineage for base, adaptation, and release meets that requirement. Increase logging of user prompts only does not, because its focus remains capturing runtime input without model-build lineage.
Answer D is incorrect because Select a different AWS Region primarily concerns moving deployment location without solving reproducibility. This scenario needs customized-model lineage. Record lineage for base, adaptation, and release meets that requirement. Select a different AWS Region does not, because its focus remains moving deployment location without solving reproducibility.
