# Engineering Eval Dashboard · 工程验证仪表盘

> Use when agentic coding increases output faster than review capacity. The goal is not "more tests";
> it is a dashboard that makes correctness, blast radius, and human ownership visible.

## 1. Scope

| Field | Fill |
|---|---|
| Project / repo |  |
| Change surface |  |
| Human owner |  |
| Agent role |  |
| Release path |  |
| Non-goals |  |

## 2. Correctness Contract

| Layer | Machine-checkable rule | Command / verifier | Fails block merge? | Owner |
|---|---|---|---|---|
| Format / lint |  |  | yes / no |  |
| Type / schema |  |  | yes / no |  |
| Unit behavior |  |  | yes / no |  |
| Integration path |  |  | yes / no |  |
| Security / policy |  |  | yes / no |  |
| Product acceptance |  |  | yes / no |  |

## 3. Delegate / Review / Own

| Task | Reversibility | Blast radius | Tier | Why |
|---|---|---|---|---|
|  | high / medium / low | low / medium / high | delegate / review / own |  |

Rule: implementation inside a bounded module usually delegates; boundaries, trust zones, irreversible
data changes, and user-visible policy stay with a named human.

## 4. Eval Readout

| Signal | Baseline | Current | Threshold | Action if breached |
|---|---:|---:|---:|---|
| Test pass rate |  |  |  |  |
| Flaky failures |  |  |  |  |
| Review defects per PR |  |  |  |  |
| Escalations to human |  |  |  |  |
| Rollbacks / hotfixes |  |  |  |  |

## 5. Escalation

- Stop the agent when: 
- Require human review when: 
- Auto-merge is allowed only when: 
- The next verifier to add is: