AI will confidently tell you it discovered new physics. ClaimGate doesn't argue — it extracts each claim, routes it to a deterministic verification gate, seals an evidence pack with a certificate hash, and tells you exactly what survives.
python -m claimgate claim "Energy equals mass times acceleration and proves a new gravity effect that saves 90% power."| Claim | Gate | Verdict |
|---|---|---|
| Energy = mass × acceleration | UnitGate | DIMENSIONALLY_INVALID |
| proves a new gravity effect | ClaimGate | UNSUPPORTED_CLAIM |
| saves 90% power | ExperimentalGate | NEEDS_DATA |
3 claims extracted · 3 evidence packs sealed · replay reproducible.
AI proposes. Gates verify. Unsupported claims do not survive.
This does not prove scientific truth. It records what was checked, what failed, what needs evidence, and what still requires simulation or experiment. Open problems are refused (UNSUPPORTED_CLAIM), never "proved".
→ 60-second demo · Benchmark docs · Limitations · Contamination policy
python -m claimgate selftest # core self-tests
python -m claimgate run --physics --json # PhysicsGate (dimensional coherence)
python -m claimgate run --replay --json # evidence-pack replay audit
python -m claimgate claim "your claim" --json # full evidence packs
python -m pytest tests/ -q # anti-cheat testsProofBench X is also a math verification benchmark — it tests whether a system can verify exact arithmetic, metamorphic equivalence, counterexamples, certificate stability and replay without trusting a model.
python -m claimgate run --v1 --json # 395 cases
python -m claimgate run --v2 --json # 115 casesThe verifier's integrity scores are perfect (metamorphic, certificate stability, replay, parser, disguise, deep-chain all 1.0); the bundled simulated model fails exactly where untrusted models fail (numeric slips, false identities, lost exactness). That gap is the point. → Math benchmark scoreboard
ClaimGate is the front door. Each gate it routes to is also a small, standalone, MIT-licensed tool you can adopt on its own — pure stdlib, no private code:
| Tool | Does | Repo |
|---|---|---|
| UnitGate | dimensional consistency of LHS = RHS |
kyal102/unitgate |
| EvidencePack | seal a verdict with a deterministic certificate hash | kyal102/evidencepack |
| ReplayGate | re-run a sealed pack, detect drift (safe shell=False allowlist) |
kyal102/replaygate |
| ClaimLint | lint README/docs for over-claims (CLI + GitHub Action) | kyal102/claimlint |
| ClaimStack | end-to-end demo wiring the above into one pipeline | kyal102/claimstack-demo |
Try it in one command:
python -m claimgate --demoProofBench X / ClaimGate is proof-aware verification infrastructure: exact math, symbolic invariance, counterexample witnesses, dimensional (PhysicsGate) and theory (TheoryGate) coherence checks, sealed evidence packs, drift-detecting replay, and claim extraction/routing — all decided by a deterministic verifier, never a model. Powered by the JARVI3 / DTL verification engine.
Prototype status — no public-leaderboard eligibility until an uncontaminated, post-implementation holdout exists (see docs/CONTAMINATION_POLICY.md). The verifier is the final authority.
MIT — see LICENSE. (Change if you prefer a different license.)