A. Incorrect.
CloudWatch alarms observe metrics. They are not the primary service for filtering uploaded image and text content.

B. Incorrect.
Training custom moderation models adds significant model management overhead when managed services already meet the need.

C. Incorrect.
Pre-authentication checks run during sign-in, not during every content upload and generation workflow.

D. Correct.
Step Functions can orchestrate managed Bedrock, Comprehend, and Rekognition controls around the existing S3 workflow with low infrastructure management.

A. Incorrect.
Setting every category to high and blocking all PII can create excessive false positives and does not apply different handling strategies.

B. Incorrect.
Chaining many guardrails increases complexity and is not needed when one guardrail can define multiple policy types and actions.

C. Correct.
A single well-tuned guardrail can combine denied topics, content filters, PII masking or blocking, and custom handling for both inputs and outputs.

D. Incorrect.
Grounding checks are useful for factuality, but they do not replace topic denial, harmful-content filtering, or PII controls.

A. Incorrect.
The problem is overly broad retrieval, and the scenario states that additional model tuning is not expected to help.

B. Incorrect.
A custom migration can work but is a larger architecture change than needed.

C. Correct.
Metadata-aware filtering constrains retrieval by fields such as content type, publication date, source, or category while preserving the existing Bedrock Knowledge Bases architecture.

D. Incorrect.
Replacing the application with Amazon Q Business is unnecessary when the current knowledge base can use metadata filters.

A. Incorrect.
S3 object tags alone do not provide the purpose-built prompt workflow, parameterization, and approval governance required.

B. Correct.
Bedrock Prompt Management is built for prompt templates, variables, versioning, governance, and integration with IAM and CloudTrail.

C. Incorrect.
SageMaker Canvas is not the right service for enterprise prompt lifecycle governance in Bedrock workflows.

D. Incorrect.
A custom Lambda and DynamoDB approval system adds operational overhead when Bedrock provides prompt management features.

A. Incorrect.
A custom routing layer increases complexity and duplicates behavior available through Bedrock inference profiles.

B. Correct.
Geographic cross-Region inference profiles can absorb Regional pressure while keeping traffic inside the allowed geography.

C. Incorrect.
Sending European data to North America can violate the data residency requirement.

D. Incorrect.
Lowest-latency global routing alone does not enforce the European data boundary.

A. Incorrect.
Separate authentication and custom I/O formats increase inconsistency across environments.

B. Correct.
The Bedrock Converse API standardizes conversational message structure across SDKs, IAM roles provide consistent authentication, and prior messages preserve context.

C. Incorrect.
Process memory is not reliable across Lambda invocations or container restarts and does not standardize API use.

D. Incorrect.
Hard-coded credentials and public normalization endpoints are poor security and unnecessary.