Anthropic Evals and Observability

When to use

Invoke when an Anthropic-backed capability (Claude via the Messages API) needs measurable regression gates, production telemetry, prompt/model versioning, or operational monitoring, and the success criteria and regression bar come from ai-architecture.md or a quality-engineering handoff.

Do not use to define the model contract, success criteria, or eval bar from scratch, nor for RAG, tool, or agent implementation.

Inputs

Required: