EduLogic wants to compare two Bedrock models for answer accuracy and safety before rolling one into production. They need to run automatic, repeatable metrics at scale and also collect human judgments for edge cases. Which architecture best implements both automated evaluation with built-in metrics and human evaluation for sampled outputs?

This option correctly combines automated and human evaluation. Bedrock's evaluation capabilities directly support running models against labeled datasets and computing built-in metrics like BLEU/ROUGE/F1 and safety scores, which are essential for large-scale, repeatable assessments. Persisting artifacts to S3 ensures data governance and auditability. For human evaluation, SageMaker Ground Truth or Amazon A2I are the appropriate services for collecting high-quality human judgments on sampled outputs, especially for nuanced edge cases. This architecture provides a comprehensive and robust evaluation framework. The other options are less suitable: Relying on CloudWatch Logs and Lambda for real-time metric computation is inefficient and complex for comprehensive model evaluation, and a public web form for feedback is not scalable or secure. SageMaker Model Monitor is for monitoring models in production, not for pre-production comparative evaluation, and Bedrock does not automatically create a human review queue for ambiguous cases. Relying solely on human evaluation via Mechanical Turk for all outputs is impractical, expensive, and slow for large datasets, and completely neglecting automated metrics misses valuable, scalable insights.