What we prove — and what we won't

We mapped 178 theories and 285 relations between them through the engine, deterministically, with no model in the loop. Each checkable result is a verdict + a worked trail + a re-checkable seal. The number that matters most: across this assay, 0 sealed claims were false — a measured result on this set, not a promise the engine can never err; when it does, we fix it and re-run. And the part that earns your trust is not the ones it confirmed — it's the ones it refused.

Every checkable claim carries a receipt you can re-fetch by its content hash; re-fetch and the bytes must match, or it is not that record. Don't trust us — re-check it. Point your own agent at the engine →

Sealed — a receipt, not "trust me"

A sample across the fleet. Each links to its permanent, re-verifiable seal.

Try one yourself in the browser → · dozens of domains: physics, medicine, finance, chemistry, law, statistics, cryptography, optics, genetics, music, and more.

Reasoned in chains — composed derivations

The engine doesn't only check single facts; it seals a chain, each step building only on a confirmed prior — lemma to theorem, end to end.

It even runs the code

Beyond formulas: a real function, executed against test cases and sealed only if every case passes.

Refused — the ones it would not seal

These are famous, published, argued-over — and not one is a deterministic computation. So the engine returned INCOMPLETE: honest silence, never a false verdict. A machine that cannot say "I can't seal that" cannot be trusted when it says "this holds." This is the map's honest edge.

Note the sharp line: the engine refuses the general conjecture (Goldbach for all evens) yet could seal any single instance (100 = 47 + 53, both prime). That boundary — between the computed and the merely believed — is the whole point.