A system produces a structured claim.
For example: an extracted total, a constrained calculation, or a recovered sequence rule.
Confidence is not evidence. Chiron checks the claims it can prove exactly, refutes what is false, and leaves everything outside that boundary explicitly uncertified.
Try a public, live licensed-engine check below — or use the sequence lab to see held-out proof in action.
Chiron extracts supported, exactly checkable claims and reports each one as VERIFIED, REFUTED, or REFUSED. Free-form remainder is never silently approved.
For supported claims, verdicts are exact. For sequences, a rule is stamped only when it predicts held-out terms it never saw — ==, not a tolerance.
Fed the primes, it can recover a formula and still decline to stamp it. In claim checking, free-form remainder stays explicitly uncertified.
The current frozen external eval has 22 stamped outputs, all 22 externally correct, 0 false stamps, and 12 refusals. Run it yourself →
Chiron is not an audit platform or a general answer judge. It is the small, inspectable gate a workflow can place around one independently checkable output.
For example: an extracted total, a constrained calculation, or a recovered sequence rule.
It verifies exact evidence, refutes a false supported claim, or leaves the rest explicitly uncertified.
A machine-readable evidence record shows VERIFIED, REFUTED, or REFUSED and what was not covered.
A supported verified result can satisfy a defined gate. A refuted or refused result routes to the named human or wider review process.
The restraint is the brand — nothing here claims more than it proves.
Try a non-sensitive claim with the live checker, or a sequence with the browser demo core.
The licensed engine's frozen outputs, graded live against oeis.org — data the author doesn't control.
The real engine over HTTP — certificate out, source never serialized, rate-limited.
The claim checker and “licensed engine” sequence mode call the public HTTP demo. The browser sequence core is prototype/browser_core.py, fetched verbatim and run by CPython-in-WebAssembly. It is strictly weaker by design: five hypothesis families and a more conservative evidence rule. It refuses Catalan, factorials, and Tribonacci — which the licensed engine stamps (see eval/ and BATTERIES.md). The shared contract is narrower and more important: do not stamp what cannot be exactly supported.
prototype/primus_prototype.py enforces the membrane, the deterministic seal, the adjacency law — and a certifier stub that structurally cannot say VERIFIED (that's gate 26). This fetches the real file and runs its battery here.