Paste a source and a claim. The wheel checks whether the claim is actually grounded in the source, then answers the question that matters: is this safe to serve, to cache, to publish? This page calls the real API, live.
Every verdict names the policy pattern it matched and the evidence per detector — never a bare yes/no. Composite claims are split into atoms and grounded one by one; a failing atom is anchored so you can see which sentence is the problem. Irreversible actions (cache, publish) are judged stricter than serving: when a detector can't sign, the gate refuses. If our detector tier is degraded, the response says so honestly.
The wheel measures grounded-in-source — RAG faithfulness and report-checking, not chat quality. It reduces confident hallucination; it does not certify truth. It grades our own outputs with the same rules.
Use it from your own code: POST /v1/verify with {source, claim, action} — see the docs. Prefer to audit your evals yourself first? Take the verification starter kit — it's a gift, no key needed.
The shared demo key allows 6 checks a minute. · How we treat your data → · Back to overview