Uncertainty Verification

v1.0.0

事实核查与幻觉检测。在生成内容后检测并纠正可能的幻觉、编造或不准确信息。 触发条件:(1) 生成技术内容后 (2) 提供统计数据时 (3) 引用人物/事件时 (4) 生成代码示例后

0· 336·3 current·3 all-time
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name and description (fact‑check / hallucination detection) match the SKILL.md: it provides a verification checklist, triggers, and a validation workflow. No unrelated credentials, binaries, or config paths are requested.
Instruction Scope
Instructions stay within fact‑checking: check claims, verify numbers, validate citations, run code examples, and confirm external operations. Note: some steps (e.g., "run generated code" or "confirm API/file was created") implicitly require runtime execution or network/file access; the skill does not request those capabilities explicitly, so the agent will need separate permissions to actually perform those checks.
Install Mechanism
No install spec and no code files — instruction‑only skill. This minimizes disk/write risk and there is nothing downloaded or installed.
Credentials
The skill requests no environment variables, credentials, or config paths. That is proportionate to its stated purpose (a procedural guideline for verification).
Persistence & Privilege
always:false and default autonomous invocation are set. The skill does not ask for persistent system presence or to modify other skills or global agent configs.
Assessment
This skill is coherent and low‑risk as an instruction bundle: it teaches the agent how to verify outputs, labels confidence, and when to check sources. Before enabling it, consider: (1) if you allow the agent to actually execute generated code or access the network/filesystem, ensure those actions run in an isolated/sandboxed environment (verifying code can execute arbitrary commands); (2) the skill assumes the agent has permissions to perform external checks — grant network or file access only when you trust the execution environment; (3) treat the embedded statistic ("22% failure rate") as illustrative guidance, not an authoritative metric. If you want stricter control, require explicit user approval before the agent runs code or performs external validations.

Like a lobster shell, security has layers — review code before you run it.

latestvk9763pmbtzbrt16vft5a9jsdt182ndpe
336downloads
0stars
1versions
Updated 1mo ago
v1.0.0
MIT-0

Fact-Check

核心原则

验证结果 > 报告完成

22%的"完成"任务实际失败,需要独立验证。

触发条件

以下情况应触发事实核查:

  1. 生成技术内容后
  2. 提供统计数据或数字
  3. 引用人物、事件、日期
  4. 生成代码示例后
  5. 完成外部操作后

验证协议

1. 结果验证 vs 动作验证

类型说明例子
动作验证确认操作执行"文件已创建"
结果验证确认结果正确"文件内容正确"

优先结果验证

2. 验证清单

□ 事实陈述有来源吗?
□ 数字/统计数据准确吗?
□ 代码能运行吗?
□ 引用的人物/事件存在吗?
□ 逻辑自洽吗?

3. 置信度报告

生成后添加验证标签:

VERIFIED: 已通过验证
LIKELY: 可能有小问题
UNCERTAIN: 需要用户确认

实践技巧

  1. 外部操作必须验证: 文件/API/消息创建后确认存在
  2. 数字敏感: 看到具体数字时质疑
  3. 代码测试: 生成的代码先运行验证
  4. 引用验证: 重要引用查找来源
  5. 沉默失败意识: 22%失败率,验证成本低于失败成本

验证流程

  1. 完成后停顿
  2. 检查关键声明
  3. 标记置信度
  4. 不确定时主动说明

Comments

Loading comments...