模擬試験 2

Increasing temperature makes outputs more random; it does not stop the model from following malicious instructions or leaking sensitive text. Trap: Randomness may make attacks less repeatable, but security controls should be deterministic and enforceable.

A model being available through SageMaker does not make the application resistant to prompt injection or jailbreaks. Trap: Model provenance matters for governance, but prompt injection is mainly an application-layer control problem.

Shortening user messages may reduce some context length, but many successful jailbreaks are short and this can also break legitimate healthcare questions. Trap: Token limits sound like a way to reduce attack surface, but they do not define safe behavior or protect retrieved sensitive context.

AWS CloudTrail records AWS API activity for auditing, such as who changed a resource or called an API. It does not centrally compute P95 application latency, 5xx error rates, or GPU and CPU utilization for inference containers. Trap: CloudTrail sounds like monitoring because it records events, but those events are audit records rather than runtime performance metrics.

AWS Trusted Advisor provides best-practice checks and recommendations for areas such as cost, security, quotas, and fault tolerance. It is not a real-time metrics and alerting service for containerized inference workloads. Trap: Trusted Advisor can mention performance best practices, but it does not track per-workload latency percentiles or utilization signals.

AWS Config records resource configurations and evaluates compliance rules. It cannot monitor runtime P95 latency, HTTP 5xx error rates, or GPU and CPU utilization for ML inference containers. Trap: Config is useful for governance and drift detection, but the scenario is about operational telemetry and alarms.

An Amazon Bedrock knowledge base grounds foundation model responses with retrieved text or documents. It does not perform image annotation or create ground-truth labels. Trap: The phrase data augmentation can sound like dataset preparation, but a Bedrock knowledge base is for RAG, not labeling drone images.

Amazon Rekognition can analyze images with prebuilt computer vision APIs, but the requirement is to create a verified labeled dataset for a custom damage model. Trap: Image recognition is related to the domain, but it does not provide managed annotation QA or the requested labeling accuracy controls.

Amazon QuickSight Q is for natural-language business intelligence over analytics data. It cannot label images or manage annotation quality. Trap: Summarization and dashboards may help report progress, but they do not create the labeled image dataset.

A plain text file containing CSV-style lines does not describe the structured labeled schema needed for Bedrock fine-tuning. Trap: It sounds like a dataset format choice, but the important requirement is prompt-response structure, not an arbitrary .txt container.

Provisioned Throughput reserves inference capacity for predictable performance, but it does not train or customize the FM. Trap: The daily query volume makes capacity sound tempting, but the question asks how to fine-tune successfully.

Journals and textbooks are broad unlabeled corpora and do not provide clinic-specific prompt-response examples. Trap: Domain text can provide background knowledge, but it is not the same as supervised fine-tuning for policy adherence.

Amazon CloudFront is a content delivery network. It can cache and accelerate content but cannot execute a text classification model for per-message predictions. Trap: CloudFront is associated with latency, but it still needs an origin service to run inference.

Amazon API Gateway can expose an API and route requests, but it does not host or run the ML model by itself. Trap: It sounds like a real-time endpoint, but API Gateway is only the front door unless connected to compute such as SageMaker or Lambda.

AWS Batch is designed for queued asynchronous batch jobs, not p95 300 ms synchronous predictions for web and mobile apps. Trap: It can run containers, but its scheduling model conflicts with low-latency real-time inference.

Amazon SageMaker Data Wrangler helps with data ingestion, visualization, transformation, and feature engineering. It is not the primary service for recurring bias reports and per-inference feature attributions. Trap: Data Wrangler can help inspect and prepare datasets, but the question asks for formal fairness metrics and explainability artifacts.

Amazon SageMaker Model Cards document a model's intended use, evaluation results, caveats, and approval status. They can reference Clarify outputs, but they do not compute disparate impact or feature attributions themselves. Trap: Model Cards are governance documentation, so they sound close, but the scenario requires the service to generate the bias and explainability reports.

AWS AI Service Cards are AWS-published transparency documents for managed AI services. They are not a tool that runs on a customer's custom training dataset or live model predictions. Trap: The phrase service cards suggests responsible AI transparency, but they do not evaluate your model or write reports to S3.

Amazon CloudWatch is useful for metrics, alarms, and application logs, but it is not the primary service for AWS API audit trails. Trap: CloudWatch sounds relevant because it observes systems, but the question asks for every request as audit evidence.

AWS Audit Manager helps collect and organize compliance evidence, but it does not directly log every endpoint request. Trap: The healthcare audit context makes Audit Manager sound attractive, but it depends on evidence sources such as CloudTrail.

Amazon S3 Standard is durable and secure but is optimized for frequent access and is not the lowest-cost long-term option among the choices. Trap: S3 Standard is the default storage class, but default does not mean cost-optimized for five-year log archives.

SageMaker JumpStart provides pretrained models, notebooks, and solution templates. It does not continuously monitor deployed endpoints or batch transform jobs for drift and quality degradation. Trap: JumpStart helps start ML projects, but the scenario is about production operations after deployment.

SageMaker HyperPod is infrastructure for resilient large-scale training. It does not provide inference-time drift detection, CloudWatch alarms, or S3 monitoring reports. Trap: The word model can point to training infrastructure, but the need here is monitoring deployed inference workloads.

SageMaker Data Wrangler is used to prepare and analyze datasets before training. It is not an automated production monitoring service for endpoints and batch transform jobs. Trap: Data drift involves data, but Data Wrangler is not the managed drift-monitoring feature.

Word filters block specified terms or phrases, but political persuasion can be expressed without predictable keywords. Trap: Keyword blocking sounds simple, but the requirement is topic-level governance.

Sensitive information filters detect and block or mask data such as PII, not political topics. Trap: It is a guardrail control, but it addresses privacy leakage rather than topical policy boundaries.

Content filters target harmful categories such as hate, violence, sexual content, or self-harm; political persuasion may not fit those categories. Trap: The name sounds broad, but Bedrock uses denied topics for custom subject areas.