Cross Check
Multi-source verification for factual claims with confidence scoring and source links.
Workflow
- 1. Optimize query — see
{baseDir}/references/query-optimization.md. - Search first (REQUIRED) — ALWAYS use
web_search as the primary tool. Never skip to web_fetch directly. Run web_search(query, count=10). If T1 found, stop. If only T2-3, run a second query. If contradictions, investigate. - Evaluate sources — classify by tier. See
{baseDir}/references/source-tiers.md. - Direct verify (optional, only after search) — use
web_fetch(url) only to verify specific URLs found via search. Never use web_fetch as a substitute for web_search. - Present — MUST include clickable source URLs for every result:
> [Answer]
> Confidence: [T1-T5] ([%]) | Sources:
name, ... | Verified: [date] | Note: [caveats]
Source URLs
CRITICAL: Every cited source MUST include a direct clickable URL. If URL unavailable: T1 → find alternative source, T2+ → downgrade one tier. Never list a source without its URL.
Error Handling
- - Zero results → broaden query, retry once. Still empty → T1 confidence.
- INLINECODE8 fails → skip, note it, drop confidence one level.
- All T4 sources → report T2, mark Uncertain.
- Contradictions → see
{baseDir}/references/research-patterns.md.
Security & Privacy
This skill uses only OpenClaw's built-in web_search and web_fetch tools. No data is sent to external services beyond what these tools do.
External Endpoints:
- - Web search API (via OpenClaw's built-in websearch, provider depends on user config)
- Web pages (via OpenClaw's webfetch for verification)
Data Handling:
- - No credentials required
- No data stored locally
- Search queries sent through OpenClaw's configured search provider
- No autonomous invocation without user intent - activates when queries contain verification language
Trust Statement:
By using this skill, your search queries are sent through OpenClaw's built-in web search provider. Only install if you trust OpenClaw's data handling practices.
交叉验证
对事实性声明进行多源验证，包含置信度评分和来源链接。
工作流程
- 1. 优化查询 — 参见 {baseDir}/references/query-optimization.md。
- 优先搜索（必需） — 始终使用 websearch 作为主要工具。切勿直接跳过到 webfetch。运行 websearch(query, count=10)。如果找到T1级来源，停止。如果只有T2-3级来源，运行第二次查询。如果存在矛盾，进行调查。
- 评估来源 — 按层级分类。参见 {baseDir}/references/source-tiers.md。
- 直接验证（可选，仅在搜索之后）— 仅使用 webfetch(url) 验证通过搜索找到的特定URL。切勿用 webfetch 替代 websearch。
- 呈现结果 — 每个结果必须包含可点击的来源URL：
> [答案]
> 置信度：[T1-T5]（[%]）| 来源：
名称，... | 验证日期：[日期] | 备注：[注意事项]
来源URL
关键：每个引用的来源必须包含可直接点击的URL。如果URL不可用：T1级 → 寻找替代来源，T2级以上 → 降低一个层级。切勿列出没有URL的来源。
错误处理
- - 零结果 → 扩大查询范围，重试一次。仍为空 → T1级置信度。
- web_fetch 失败 → 跳过，注明情况，降低一个置信度级别。
- 全部为T4级来源 → 报告为T2级，标记为不确定。
- 存在矛盾 → 参见 {baseDir}/references/research-patterns.md。
安全与隐私
本技能仅使用OpenClaw内置的 websearch 和 webfetch 工具。除这些工具的功能外，不会向外部服务发送任何数据。
外部端点：
- - 网络搜索API（通过OpenClaw内置的websearch，提供商取决于用户配置）
- 网页（通过OpenClaw的webfetch进行验证）
数据处理：
- - 无需凭证
- 不本地存储数据
- 搜索查询通过OpenClaw配置的搜索提供商发送
- 无用户意图时不会自动调用——仅在查询包含验证性语言时激活
信任声明：
使用本技能时，您的搜索查询将通过OpenClaw内置的网络搜索提供商发送。仅当您信任OpenClaw的数据处理实践时才安装。
