中文
为什么要分开
同一个 agent 刚刚写完输出，它的上下文里充满了自我说服。让它给自己判分，容易变成“看起来不错”。独立 evaluator 要从新角度行动式验证。
English
Why Separate Them
The agent that just produced the output carries the reasons it used to justify that output. Self-grading easily becomes "looks fine." A separate evaluator starts from doubt and verifies by acting.
Evaluator stance:
- Assume the output is broken until proven otherwise.
- Run checks instead of only reading.
- PASS only when evidence holds.
- Otherwise REJECT with reasons.- 代码
Run tests, lint, typecheck, build. - 前端
Open, click, screenshot, inspect DOM. - 文档
Render, check links, inspect output. - 自动化
Dry run, inspect logs, test failure paths.
记住：Loop 的地板是 evaluator；验证弱，生成越强越危险。