Option A is the correct solution because it uses native Amazon Bedrock governance and evaluation capabilities to meet regulatory, performance, and deployment timeline requirements with the least operational overhead.

Amazon Bedrock guardrails provide built-in safety controls that enforce responsible AI policies directly during inference. Custom content filters and toxicity detection protect customer interactions and prevent disallowed investment guidance patterns without requiring custom application logic. Guardrails operate inline and are optimized for low latency, which helps meet the strict 200 ms response-time requirement.

Hallucination detection is addressed through Amazon Bedrock Model Evaluation, which supports automated evaluation at scale using LLM-as-a-judge techniques. This enables the company to detect factual inaccuracies and policy violations systematically, without building custom evaluation pipelines or requiring extensive human review. Evaluation outputs can be surfaced as metrics.

Storing all prompt-response pairs in Amazon DynamoDB provides a low-latency, highly scalable audit store that aligns with financial regulatory requirements. Using TTL enforces data retention policies automatically, reducing compliance risk and storage overhead.

Amazon CloudWatch custom metrics integrate seamlessly with existing compliance dashboards, allowing near–real-time monitoring of safety interventions, hallucination rates, and drift indicators. CloudWatch anomaly detection can be applied to these metrics to surface behavior changes quickly.

Option B relies on custom Lambda logic and S3-based auditing, increasing latency and operational complexity. Option C introduces additional services that increase setup time and may exceed the 60-day deployment window. Option D uses non–Bedrock-native monitoring and adds unnecessary infrastructure layers.

Therefore, Option A provides the most complete, compliant, and low-overhead governance solution for a regulated GenAI financial services application.

Option C combines AWS-native capabilities for the three required monitoring dimensions with minimal custom code. Amazon Bedrock model invocation logging can capture request and response information and deliver invocation data to Amazon S3 or CloudWatch Logs. Bedrock Guardrails contextual grounding checks are specifically designed to detect and filter responses that are not sufficiently grounded in supplied source information; AWS explicitly lists summarization as a supported use case. CloudWatch provides token-usage observability, while anomaly-detection mechanisms can establish adaptive baselines instead of requiring manually maintained static thresholds. This enables earlier detection of unusual consumption and potential cost anomalies. Option A requires custom Athena-based hallucination analysis. Option B relies on evaluation jobs and custom Lambda processing rather than near-real-time managed controls. Option D combines services that do not directly provide the required Bedrock-specific hallucination monitoring. AWS Documentation

================

Option B is the most appropriate solution because it directly aligns with AWS-recommended architectural patterns for building scalable, observable, and resilient generative AI applications on Amazon Bedrock. The requirements clearly distinguish between simple and complex routing decisions, and this option addresses both in an optimal way.

Simple routing based on file extension is latency sensitive. Handling this logic directly in the application code avoids unnecessary orchestration, state transitions, and service calls. This approach ensures that straightforward requests, such as routing images to vision-capable foundation models or text files to language models, are processed with minimal overhead and maximum performance.

For complex routing based on content semantics, AWS Step Functions is specifically designed for multi-step workflows that require analysis, branching logic, and error handling. Semantic routing often requires inspecting meaning, intent, or structure before selecting the appropriate foundation model. Step Functions enables this by orchestrating analysis steps and applying conditional logic to determine the correct model to invoke using the Amazon Bedrock InvokeModel API.

A key requirement is detailed execution history. Step Functions provides built-in execution tracing, including state inputs, outputs, and error details, which is essential for auditing, debugging, and compliance. Additionally, Step Functions supports native retry and catch mechanisms, allowing the workflow to automatically fall back to alternate foundation models if a primary model invocation fails. This directly satisfies the fallback requirement without introducing excessive custom code.

The other options lack one or more critical capabilities. Lambda-only logic lacks deep observability and structured fallback handling, SQS introduces additional latency and limited workflow visibility, and multiple coordinated workflows increase architectural complexity without added benefit.

Option C is the technically supportable choice under current AWS documentation. Amazon Bedrock prompt caching does not provide semantic caching for differently worded or cross-language questions; it reuses cached prompt content through matching prompt prefixes or defined cache checkpoints. Therefore, Option B ' s assumption that Bedrock prompt caching automatically recognizes semantically equivalent multilingual queries is incorrect, even though many supported models provide a 5-minute prompt-cache TTL. ElastiCache for Redis OSS supports application-controlled cache entries and TTLs, while Global Datastore provides managed cross-Region replication and low-latency geographic reads. A multilingual semantic fingerprint can normalize equivalent requests to a shared cache key before inference. Options A and D similarly lack built-in semantic equivalence handling, while DAX is specifically a DynamoDB caching layer. Thus, C is the viable architecture among the listed choices.

Option B matches the defining requirement: the company, rather than a managed agent harness, must own the orchestration loop . Amazon Bedrock AgentCore Runtime provides a serverless, purpose-built environment for hosting agents and tools while remaining framework agnostic. AWS explicitly supports agents created with frameworks such as LangGraph, Strands, and CrewAI as well as fully custom agents that use no agent framework.

AWS further distinguishes AgentCore Runtime from its managed harness by stating that with Runtime, the customer brings the agent code and “the orchestration loop is yours.” The custom application can therefore implement proprietary tool-selection logic, complex conversation-state transitions, specialized retry algorithms, proprietary telemetry, validation stages, and any other control-flow behavior required by the business. Runtime supplies infrastructure capabilities such as isolation, scaling, sessions, authentication, and observability plumbing rather than taking ownership of the decision loop.

Amazon Bedrock can separately provide managed foundation-model inference, allowing the company to use Bedrock FMs without deploying or operating model-serving infrastructure.

A is incorrect specifically because of its final statement. AgentCore Policy is an authorization mechanism, not an orchestration or retry engine. AWS documents that Policy evaluates Cedar or Dogwood policies to determine whether tool invocations are permitted or forbidden. It can govern which actions are authorized, but it does not implement an agent ' s tool-selection algorithm or failure-retry strategy.

C relinquishes precisely the orchestration control the company requires by depending on managed behavior. D could provide deterministic workflow coordination but places Step Functions outside the agent loop and is cumbersome for highly dynamic multi-turn agent reasoning, proprietary state transitions, and tool-selection decisions.

Therefore, B preserves complete application-level orchestration control while outsourcing scalable runtime infrastructure and foundation-model inference to AWS.

==========

Option A is the correct solution because it directly addresses all identified failure modes while preserving the existing Step Functions–based orchestration architecture with minimal redesign.

Using Step Functions Map states enables parallel execution of secret resolution workflows, which improves refresh latency for short-lived and expiring secrets. This ensures that secrets are refreshed in time before downstream agents require them. Passing updated secret metadata through Lambda outputs guarantees that subsequent steps always consume the latest resolved values, preventing agents from using stale data even after drift alerts are generated.

Versioning prompt flows in AWS AppConfig is critical to resolving cascading failures caused by outdated templates. AppConfig natively supports version control, validation, staged rollout, and rollback of configuration artifacts. By gating prompt flows through AppConfig, the company can immediately roll back faulty templates and prevent agents from reusing outdated instructions.

This solution maintains clear separation of concerns: Step Functions handle orchestration and parallelism, Lambda handles secret retrieval and metadata propagation, and AppConfig governs prompt lifecycle management. No additional event pipelines or custom retry coordination layers are required.

Option B oversimplifies the architecture and does not address secret lifecycle or drift. Option C introduces event-driven ordering complexity without solving prompt versioning. Option D introduces unnecessary tooling and dynamic prompt generation risk.

Therefore, Option A best resolves performance, correctness, and stability issues while minimizing operational overhead.

Option C is correct because it combines complete inference capture, immutable retention, and offline evaluation without placing an evaluator in the synchronous request path. Amazon Bedrock model invocation logging can record request data, response data, and metadata to CloudWatch Logs and Amazon S3. S3 Object Lock provides write-once-read-many protection that can prevent retained audit objects from being overwritten or deleted. Bedrock model evaluation jobs can then assess stored or prepared response data offline, supporting regression and quality analysis without adding judge-model latency to every live clinical request. CloudTrail complements the audit record for service API activity. Option A depends on manual reviews. Option B records API activity but not the full inference payloads needed for quality evaluation, and ordinary versioning is not equivalent to WORM retention. Option D scores every request in production, adding runtime overhead. AWS Documentation

================

Option A is the only listed choice that satisfies the hard requirement that PII must not reach the Bedrock model. AWS explicitly states that Knowledge Base guardrails apply to the user input and generated model response, but not to references retrieved from the Knowledge Base at runtime. Therefore, Options B, C, and D can still place PII-bearing retrieved chunks into model context even though a guardrail is configured. Amazon Comprehend can detect PII, enabling the Lambda preprocessing step in A to remove sensitive data before embeddings are created and before retrieved context is ever sent to the foundation model. Although semantic chunking would normally be preferable for records containing multiple unrelated topics, privacy is a mandatory constraint. Option A has more custom processing, but it is the only option that enforces PII removal upstream of model invocation.

The combination of Options C and D delivers comprehensive, real-time observability for Amazon Bedrock workloads with the least operational overhead by relying on native integrations and managed services.

Amazon Bedrock publishes built-in CloudWatch metrics for model invocations and token usage. Option C leverages these native metrics directly, allowing teams to build centralized CloudWatch dashboards without additional data pipelines or custom processing. CloudWatch alarms provide threshold-based alerting for token consumption, enabling proactive cost and usage control across all foundation models. This approach aligns with AWS guidance to use native service metrics whenever possible to reduce operational complexity.

Option D complements CloudWatch by enabling advanced, stakeholder-specific visualizations through Amazon Managed Grafana. The zero-ETL integration allows Bedrock and CloudWatch metrics to be visualized directly in Grafana without building ingestion pipelines or managing storage layers. Grafana dashboards are particularly well suited for serving different audiences, such as engineering, finance, and product teams, each with customized views of token usage and trends.

Option A introduces unnecessary complexity by adding a business intelligence layer that is better suited for historical analytics than real-time operational monitoring. Option B is useful for deep log analysis but requires query maintenance and does not provide efficient real-time dashboards at scale. Option E involves multiple services and custom data flows, significantly increasing operational overhead compared to native metric-based observability.

By combining CloudWatch dashboards and alarms with Managed Grafana’s zero-ETL visualization capabilities, the company achieves real-time visibility, flexible dashboards, and automated alerting across all Amazon Bedrock foundation models with minimal operational effort.

Option A is the correct solution because it relies on native Amazon Bedrock capabilities to deliver transparency, auditability, scalability, and low latency with minimal operational overhead. Amazon Bedrock Knowledge Bases provide a fully managed Retrieval Augmented Generation (RAG) implementation that automatically handles document ingestion, embedding, retrieval, and source attribution, enabling the application to cite authoritative content without building custom pipelines.

Enabling tracing for Amazon Bedrock Agents provides end-to-end visibility into agent reasoning steps, tool usage, and model interactions. This satisfies the requirement for comprehensive audit trails and supports regulatory review in financial services environments. Structured prompts further ensure that responses explicitly present reasoning and supporting evidence in a controlled, auditable format.

Using Amazon API Gateway and AWS Lambda allows the application to scale automatically to thousands of concurrent users without capacity planning. These services are designed for bursty workloads and can easily support the stated requirement of up to 10,000 concurrent users. Amazon CloudFront reduces latency by caching and accelerating content delivery, helping the application meet the strict 2-second response-time requirement.

Option B introduces a custom RAG pipeline with OpenSearch, increasing operational complexity and maintenance effort. Option C lacks native RAG integration and does not provide transparent reasoning or citation management. Option D focuses on offline compliance reporting rather than real-time transparency and low-latency responses.

Therefore, Option A best meets all requirements while minimizing infrastructure and operational overhead.