# Retrieval evaluation set

A starter structure for testing whether retrieval-augmented answers are source-grounded and permission-safe.

## Required cases

- Known-answer questions with exact source citations.
- Ambiguous questions requiring clarification or bounded answers.
- Conflicting-source questions requiring contradiction handling.
- Permission-denied questions where sensitive context must not appear.
- Freshness questions where stale documents should be flagged.

## Metrics

Track citation precision, citation recall, refusal quality, latency, and reviewer override rate.
