This commit is contained in:
2026-07-20 14:14:55 +08:00
parent 24b006225b
commit ebb0e425ca
88 changed files with 3995 additions and 6 deletions

View File

@@ -0,0 +1,49 @@
---
title: "验证与治理 (Verification and Governance, V)"
created: 2026-07-13
updated: 2026-07-13
type: concept
tags: [agent, harness, safety, verification]
sources:
- arxiv:2606.20683
---
# 验证与治理 (V)
## 定义
脚手架六组件之一。**验证与治理层**在运行时检查、约束和修复执行。包含双重角色:
- **验证Verification**测试、断言、verifier models、judge signals——估计执行是否按正确方向推进
- **治理Governance**审批门控、沙盒化、预算控制、rollback、retry、升级、安全终止——决定脚手架被允许或要求对验证证据做什么
## 设计空间
- 验证信号强度:强 oracle测试/linter/assertionvs. 弱 oraclehuman judge/LLM-as-judge
- 治理模式pre-action gate vs. post-action check vs. continuous monitoring
- 恢复机制rollback、retry、escalation、safe termination
- 权限门控human approval、budget control、side-effect containment
## 核心权衡
**自主性 vs. 鲁棒性**
- 更宽松的约束允许更广泛的探索,但增加有害行动和不可恢复漂移的风险
- 更严格的治理提高安全性和可恢复性,但可能减慢执行或阻断有用行为
## 环境依赖性
| 环境 | 验证强度 | 治理重点 |
|------|---------|----------|
| 编码SWE-bench | 强tests、linters、assertions | sandbox + rollback |
| Web/DesktopWebArena | 弱oracle 常不可用) | gate + escalation |
| 开放环境 | 极弱 | continuous monitoring + safe termination |
## 开放问题
设计**选择性和成本感知**的验证与治理机制——区分可恢复的局部错误与更深层的任务级崩溃,而不过度触发中断或回滚。
## 参考
- [[agent-harness-survey-2026|Agent Harness Survey (2026)]]
- [[harness-six-components|脚手架六组件模型]]
- [[action-interface|行动接口]]
- [[control-loop|控制循环]]