Every answer has a bound. We hold AI to it.
AI agents now set up and run engineering simulations, and the vendors that ship them grade their own work. Courant is the independent check.
Code decides. A model only drafts.
Every verdict, every number and every rank is decided by deterministic code.
A language model may draft a specification or prose. Nothing else.
- No model decides: deterministic code decides every verdict, number and rankEnforced by an architecture lintSource Security
- Deterministic: the same inputs give byte-identical verdicts and packs31 of 31 identicalSource Evidence
Certified means measured.
By fault injection on real solver runs, with the error rate published.
No regulator, standards body or accreditor is involved.
What we will not do.
Host your runs, give the gate an account, or count your use: it runs where the run is.
Print a number without its denominator.
- No hosted service, no accounts, no telemetry: the gate runs where the run isBy designSource Security
- Nothing leaves the machine: no mesh, geometry or field arrays in any artefact; names can be pseudonymisedSource Security
- Every public number generated, with its denominatorBENCHMARKS.md fails on a hand editSource Evidence