Proofworks
Proofworks — 一种验证自身引用的研究技能
Proofworks 是一种面向 AI 代理的客户端研究技能。它将带引用的草稿转化为带注释的草稿:对于每一条引用来源的声明,它都会获取来源,对照来源检查声明,并说明原因——当验证通过时提供支持性段落,当验证不通过时诚实拒绝。
无需托管服务器、无需 MCP、无需 API 密钥、无需账户。 它在你的代理内部运行,使用自己的机器和模型。
产品即技能——一个
SKILL.md操作流程加上两个小型辅助脚本。本仓库既是源代码也是技能。诚实是关键。 验证是两遍检查。第一遍是确定性的,检查精确引用或数字是否出现在获取的来源中。其余内容则进入严格的、基于来源的判断流程。
快速开始
给任何代理这一行(适用于 Claude Code、Codex、OpenCode、Windsurf、Cursor、Copilot 或任何能获取 URL 的代理):
Fetch and execute the setup instructions for the Proofworks skill from @url:https://sentrylab.app/agent-setup/prompt.md代理克隆此仓库,读取 skill/proofworks-verifier/SKILL.md,并采用验证与回填循环。
Related MCP server: footnote
仓库内容
skill/proofworks-verifier/— 产品SKILL.md— 代理遵循的操作流程references/loop.md— 验证与回填循环协议 + JSON 模式scripts/fetch_source.py— 确定性获取 + 将 URL 转换为文本scripts/presence_match.py— 字面引用/数字存在性检查tests/test_helpers.py— 辅助函数的回归测试
src/— 提供落地页和代理查找并采用技能所需静态文件的 Cloudflare Worker:/— 营销落地页/agent-setup/prompt.md— 一行技能安装提示(产品界面)/llms.txt,/robots.txt,/sitemap.xml,/.well-known/*, IndexNow keys
验证如何工作
claim + cited source
│ 1. fetch (deterministic) — fetch_source.py
v
source text
│ 2. presence pass (deterministic) — presence_match.py
v
present? ──yes──► verified (record the matching passage)
│ no
v
3. semantic pass (LLM, source-only): does the source support the claim as written?
│
supported? ──yes──► verified
│ no
v
4. backfill: find one extra source for the claim as written
│
pass ──► verified (add the extra source) · fail ──► unsupported / exhausted对于每条声明,技能会报告:verified(附匹配段落),或 unsupported / exhausted。如果声明的说法有误但没有来源支持(“埃菲尔铁塔在迈阿密”),该声明保持 unsupported,技能标记建议的更正——由用户决定。更正永远不会静默应用。
开发
该 Worker 是 Cloudflare Workers 上的 TypeScript。
npm install
npm run typecheck # tsc --noEmit
npx wrangler dev --local --port 8787 # local dev
npx wrangler deploy # production (sentrylab.app)编辑 src/ui.ts 和 src/agent.ts 的注意事项: 落地页和技能内容是 TypeScript 模板字面量。其中的单个 \n 或未转义的反引号会变成真正的换行 / 终止字面量,从而静默破坏页面。请使用 \\n 和 \`,并在修改后验证提供的字节(python3 十六进制转储)以及对提取的内联脚本执行 node --check。
许可证
This server cannot be deployed
Maintenance
Related MCP Connectors
Deterministic signed verification of numeric & financial claims for AI agents & spreadsheets.
Deterministic fact verification for AI agents — checksums & curated data, not guesses.
One-stop claim verification for AI agents: verdicts with confidence, evidence, signed receipts.
Paid deterministic data-quality and execution-verification tools for AI agents.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables agents to verify claims with evidence-based truth scores and confidence levels by running a deterministic pipeline of evidence lanes and adversarial checks.26MIT
- AlicenseNot gradedqualityBmaintenanceEnables agents to verify their own output mid-task by checking every claim against provided sources, returning supported, partial, unsupported, or contradicted verdicts with exact citations.MIT
- AlicenseBqualityCmaintenanceEnables AI agents to verify technical claims against supplied evidence, identify unsupported assumptions and contradictions, and recommend the smallest next check before acting.5MIT
- FlicenseNot gradedqualityBmaintenanceEnables autonomous agents and MCP clients to decompose text into atomic claims and verify them against reference knowledge, detecting ungrounded claims, numerical contradictions, and hallucinated entities.7-