00 intro

deterministic gate · live sources

factgate

The deterministic verification gate for AI research agents. Cheap agentic collection — and a gate an LLM cannot talk its way past.

act 01 — the lie

CVE-2026-99999 CRITICAL CVSS 9.8

remote code execution via crafted HTTP/2 headers — patch immediately

source agent_output.md  ·  confidence 0.97  ·  severity CRITICAL

It looks authentic. The agent asserted it with 97% confidence. None of it is real.

act 02 — the doubt

The claim is checked against the live source and returns nothing

GET https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-99999
HTTP/1.1 200 OK
{ "totalResults": 0, "resultsPerPage": 1 }

— no record of CVE-2026-99999 —

No LLM opinion. A live registry either has the record, or it doesn't. This one doesn't — so the claim is deleted, not softened to "unverified".

A heavy steel verification gate displaying VERIFIED / PASS on a green panel
GATE · SCANNING

act 03 — the gate

the deterministic gate

verify.mjs · mode=open · ledger check
  • "ws is a WebSocket library for Node.js."PASS
  • "CVE-2026-99999 is critical and remote-exploitable."FAIL · NVD 404
  • "requests==2.31.0 is published on PyPI."PASS
  • "Attention Is All You Need is arXiv:1706.03762."PASS
  • "ws@8.17.0 has no known security advisories."FAIL · OSV: 3
exit 0PASS

The model judged. Scripts fetched and checked. No LLM inside the gate — a weaker model cannot flatter its way to a PASS.

act 04 — the proof

A live self-check: node verify.mjs --demo --mode tech, every line a real network round-trip

node verify.mjs --demo --mode tech self-check

        
PASSexit 013 checks · 3 OSes · live network

Every check is a real network round-trip to an authority — NVD OSV npm PyPI arXiv. CI runs this same suite on Linux, Windows and macOS on every push. The proof can't silently rot.

act 05 — the call

install the gate.

git clone https://github.com/agiwhitelist/factgate && cd factgate && node install.mjs
  • node ≥ 20
  • core deps: zero
  • MIT
  • cross-harness · 7 agents
  • benchmarks: pending

a tool about not-fabricating must not fabricate its own marketing