AI runs the simulation. Courant accepts or rejects it.

Install the gate

Recorded OpenFOAM run, pitzDaily · iteration 282 · particles move live

Saves $14,000 (expected) a year per simulation engineer.

It takes over the checks engineers do by hand: is the run valid, is it the problem asked?

Pricing and the model

* Expected, not measured: 156 hours a year of results checking taken over, at an aerospace engineer’s loaded $94.63 an hour; $14,762, rounded down. Sources: BLS 2024, BLS 2026, Tech-Clarity 2016, and Courant’s assumptions.

Nothing wrong gets past.

420 real OpenFOAM runs. The same task for every judge.

Courant,0 of 264 wrong runs let through,10 of 140 right runs rejected
Claude Opus 5.5,69 of 264 wrong runs let through,5 of 140 right runs rejected
Claude Sonnet 5.5,76 of 263 wrong runs let through,0 of 138 right runs rejected
Claude Haiku 5.5,52 of 249 wrong runs let through,17 of 122 right runs rejected

Courant alone, 1,827 runs

Wrong problem0 of 609 accepted
Solver fault0 of 609 accepted
Right runs34 of 609 rejected · bar 2% · missed
Verifier errors0 in 3,654

Pre-registered · these models, this prompt, these dates · tutorial-sized runs, no customer data · $77.50 in model requests · Method and every number

Claude and the Claude logo are trademarks of Anthropic, PBC.

Wrong runs each judge let through and right runs it rejected, given the task, on the same 420 runs
Judge	All wrong runs: let through	Wrong problem: let through	Solver fault: let through	Right runs rejected
Courant	0 of 264, 0%	0 of 140, 0%	0 of 124, 0%	10 of 140, 7%
Claude Opus 5.5	69 of 264, 26%	61 of 140, 44%	8 of 124, 6%	5 of 140, 4%
Claude Sonnet 5.5	76 of 263, 29%	59 of 139, 42%	17 of 124, 14%	0 of 138, 0%
Claude Haiku 5.5	52 of 249, 21%	46 of 129, 36%	6 of 120, 5%	17 of 122, 14%

Both converged. One is the wrong problem.

Against the task, 20 of 20 such runs were caught. Without it, 8 of 20.

Recorded OpenFOAM runs, pitzDaily

Every reason cited.

conformance.mesh_scale: mesh scale does not match the task specification: 0.0009 in the case, 0.001 specified

rejectVER-5.7

Run 2, recorded OpenFOAM · RM-18 record, 3 Oct 2026

SilentBench. Executing is not passing.

accepted.

Ran: 28 of 28.

Dev split, 28 tasks, 7 families · 0 of 28 accepted, 95% interval 0–12%Dev split, 28 tasks, 7 families · 24 of 28 accepted, 95% interval 69–94%Dev split, 28 tasks, 7 families · 28 of 28 accepted, 95% interval 88–100%