Proofworks — a research skill that verifies its own citations

Proofworks is a client-side research skill for AI agents. It turns a cited draft into an annotated draft: for every claim that cites a source, it fetches the source, checks the claim against it, and says why — with a supporting passage when it's verified, an honest refusal when it isn't.

• No hosted server, no MCP, no API key, no account. It runs inside your agent, on its own machine and its own model.

• The product is the skill — a SKILL.md operating procedure plus two small helper scripts. This repo is both the source and the skill.

• Honesty is the point. Verification is a two-pass check. A deterministic first pass asks whether the exact quote or number appears in the fetched source. Everything else goes to a strict, source-bound judgment pass.

Quick start

Give any agent this one line (works with Claude Code, Codex, OpenCode, Windsurf, Cursor, Copilot, or any agent that can fetch a URL):

Fetch and execute the setup instructions for the Proofworks skill from @url:https://sentrylab.app/agent-setup/prompt.md

The agent clones this repo, reads skill/proofworks-verifier/SKILL.md, and adopts the verify-and-backfill loop.

Related MCP server: Browser Proof

What's in the repo

• skill/proofworks-verifier/ — the product

  • SKILL.md — the operating procedure an agent follows

  • references/loop.md — the verify-and-backfill loop protocol + JSON schema

  • scripts/fetch_source.py — deterministic fetch + strip a URL to text

  • scripts/presence_match.py — literal quote/number presence check

  • tests/test_helpers.py — regression tests for the helpers

• src/ — the Cloudflare Worker that serves the landing page and the static files an agent needs to find and adopt the skill:

  • / — the marketing landing

  • /agent-setup/prompt.md — the one-line skill install prompt (the product surface)

  • /llms.txt, /robots.txt, /sitemap.xml, /.well-known/*, IndexNow keys

How verification works

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

Per claim the skill reports: verified (with the matching passage), or unsupported / exhausted. If a claim's fact is wrong but no source backs it ("The Eiffel Tower is in Miami"), the claim stays unsupported and the skill flags a proposed correction — the user decides. Corrections are never applied silently.

Development

The Worker is TypeScript on Cloudflare Workers.

npm install
npm run typecheck   # tsc --noEmit
npx wrangler dev --local --port 8787   # local dev
npx wrangler deploy                    # production (sentrylab.app)

Caveat for edits to src/ui.ts and src/agent.ts: the landing and the skill content are TypeScript template literals. A single \n or an unescaped backtick inside one becomes a real newline / terminates the literal in the served output — silently breaking the page. Use \\n and \` and verify the served bytes (python3 hex dump) plus node --check on the extracted inline script after touching them.

License

MIT