fact-gate

AI coding agents round numbers, invent dates and overstate test counts in the text they write about their own changes, and every existing claim-verifier on the market runs a second model to check the first one, adding cost and its own failure mode instead of removing one.

A deterministic, zero-dependency checker: it extracts checkable claims from text, compares each one against the diff, a measured command/file/JSON value, or the test suite's own output, and exits non-zero on a contradiction, no second model, no API key. Published on npm and the VS Code Marketplace, with 90+ tests and a 100/100 pass rate on its own evaluation suite.

• A command-line tool, a GitHub Action and a Claude Code hook.
• No second model, no API key, zero dependencies.

Built with: Node.js, CLI, GitHub Actions

1. 01

   An AI agent writes the description of its own change

   Runs alone
2. 02

   Every checkable claim is extracted from the text

   Runs alone
3. 03

   Each claim is compared with the diff, a measured value or the test output

   Runs alone
4. 04

   A contradiction blocks the merge until a person fixes the text

   Hands over to a person

• 0 dependencies.

Related service: Services  →