## Eval Studio

Write evals, golden sets, and regression tests for AI agents. Run them, see what breaks.

### Summary
- **Eval suites:** 5 active
- **Total goldens:** 310
- **Pass rate:** 90.3% (+1.4)
- **Avg runtime:** 1m 41s

### Running
/run

### Regressions caught
- **Today:** 2
- **Caught before merge**

### Eval Suites
Each suite is a versioned set of goldens. Pass requires all hard rubrics to clear and 0 regressions vs the last green run.

#### Suite Details
- **Agents** `support-agent / golden / v3`
  - Goldens: 119
  - Pass: 4
  - Fail: 1
  - Last Run: 38s

- **RAG** `rag-recall / chunk-strategy`
  - Goldens: 65
  - Pass: 12
  - Fail: 3
  - Last Run: 2m 12s

- **Safety** `prompt-injection / red-team`
  - Goldens: 26
  - Pass: 5
  - Fail: 1
  - Last Run: 1m 04s

- **Tool Use** `tool-call / arg-validation`
  - Goldens: 56
  - Pass: 0
  - Fail: 0
  - Last Run: 22s

- **Voice** `voice-agent / latency budget`
  - Goldens: 14
  - Pass: 2
  - Fail: 2
  - Last Run: 4m 30s

### Regressions caught today
Cases that were passing on the last green run and are now failing. The grader explains *why*.
- **`rag-recall`** · long-table-mid-page regression
  - Chunk size raised from 800→1200 broke table boundaries
- **`support-agent`** · angry-billing-tier-2 regression
  - Escalation threshold drift in v3 prompt

### Error
Invalid domain for site key.