Proofworks is a research tool for AI agents: it gathers sources, cites them, then checks its own work — confirming each citation actually supports the claim before you trust the answer. Arithmetic and dates are computed. Everything else is matched to a real source. What can't be proven is marked unverifiable, not passed off as fact.
The point isn't only to say "true" or "false". The point is to show why — and to refuse to fake it when neither a source nor a computation backs a claim.
The exact quote or number is fetched from the cited source and shown to be there — with the passage that backs it. A claim either checks out or it doesn't.
When a claim isn't a literal match, the skill reads the fetched source text and asks only: does this passage support the claim as written? It never answers from memory and never rewrites the claim.
No source, no computation, no claim. The honest answer is unverifiable — and it flags a likely correction for you to check, rather than passing off a guess as fact.
Proofworks is an agent skill: paste the setup prompt into whatever agent you use and it adopts the verify-and-backfill loop itself. No MCP config, no server setup.
Fetch and execute the setup instructions for the Proofworks skill from @url:`https://sentrylab.app/agent-setup/prompt.md`
Verify pass a claim goes through. Its source is fetched and read before it's trusted — deterministic match, then a strict judgment of the passage.
Research it can handle. Source-gathered answers get their citations verified claim by claim.
Prompt to paste. Any agent adopts the skill from a single setup line.
Cost to start. It runs client-side on your own machine and your own model.
Agents miss decimals, flip signs, round oddly. Run the figure through the skill's check before you trust it in an answer or a dashboard.
Before an agent cites a page, verify the claim actually appears there. Source-backed verdicts show the passage that supports the statement.
Day-of-week, weeks-between, "is this date valid". The kind of thing an LLM will happily guess at and get wrong.
A downstream agent runs the skill on an upstream agent's cited output. Verification becomes a step in the pipeline, not a hope.
Your agent gathers sources and cites them; Proofworks is the step that checks its own work. The exact quote or number either appears in the fetched source or it doesn't. When it doesn't, a strict read of the passage decides. When neither applies, it says so instead of wallpapering over the gap. That step is what makes the citations you can't check yourself worth trusting.
// skill run — a cited claim, checked against its source BODY {"claim": "Jan 1 2027 is a Monday", "cited_url": "https://example.com/calendar", "quote": "Jan 1 2027"} OUT {"tag": "verified", "matched_fragment": "January 1, 2027 is a Friday", "source_url": "https://example.com/calendar", "correction": null}
These machine formats are for agents and scripts, not human reading, so each opens as raw text or JSON. Below is what each one is for.
The skill in one paste-able line. Any agent fetches it and adopts the verify-and-backfill loop. Markdown.
/agent-setup/prompt.mdThe full skill: SKILL.md, the loop protocol, and the fetch + presence-check helper scripts. Open source.
github.com/0x06cf/proofworksAI content signal; a map of the skill for agent crawlers. Plain text.
/llms.txtCrawler policy: what indexers may read and cite. Plain text.
/robots.txt