A discovery company
that uses simulation.

Cogitan works on open questions in the superconducting quantum stack. Simulation is how we work on them, and surrogates are simulation made cheap — that is the entire explanation for why we build them. They are instruments, not the business.

What usually counts as a discovery here is narrower than it sounds. Our strongest results have all had the same shape: a standard method does not measure what it claims, or an approach does not survive contact with data outside the conditions it was developed in. Randomized benchmarking tracking process infidelity at r=0.96 under gate-independent noise and r=0.61 under gate-dependent noise is a result of that kind. So is finding that a decoder trained on real device data buys almost nothing over one trained on simulation.

Those answers are worth less than a breakthrough and cost far less to reach, which is a trade we make deliberately and often. What we are working on now is public, and each question carries the result that would settle it.

We evaluated larger markets — semiconductor design, aerospace — and declined them. The reasons are specific enough to be worth stating, because a company that can say why it passed on a bigger market is easier to believe than one that claims it.

Mature fields have mature tools. Semiconductor EDA is forty years of incumbents calibrated against billions of fabricated devices; aerospace CFD is decades of wind-tunnel correlation. A surrogate entering either has to beat something already proven against physical reality at enormous scale. In superconducting design the incumbent is often a physicist with an electromagnetic solver and a notebook — a beatable bar, not because the people are worse but because the tooling layer was never built.

The ground truth here is still simulable rather than empirical. This is the one that decides it, because it is what makes this method work rather than some other one. Surrogates are cheap simulation, which requires that the expensive simulation exists and that we can run it ourselves — here that is electromagnetic solvers and JoSIM. In the larger markets the binding ground truth is fab or flight data we do not own. The method works here and would not work there.

The field is writing its own methodology right now, and often getting it wrong. Benchmarks reported on random rather than grouped splits. Models answering confidently outside the region they were trained on. Figures of merit that decouple from the thing they are used to decide. In a mature field those conventions are settled and defended; here they are new and cheap to test. That is why a small team with measurement discipline keeps finding real results — the field has not built its immune system yet.

Two curves crossed recently. Operator learning, self-supervised representation methods, and conformal prediction matured over roughly three years, while superconducting devices became complex enough that hand design-space search stopped scaling. Earlier there was not enough complexity to justify a surrogate. Later, incumbents move in. That timing is why this is a company and not a paper.

And the honest cost. An immature field is also a small one, and the market may not reach the size this thesis needs. We would rather say that here than have you work it out later.

Fluxus, Corpus, and Readout Co-Design were built because we needed them for questions we were trying to answer. They are sold because they hold up, not the other way round.

That ordering is not sentiment — it is why we can tell you where each of them stops. A vendor who builds a tool in order to sell it has no incentive to publish its envelope. We publish ours because we ran into it ourselves. Fluxus-S serves three of eight SFQ cell types and refuses the rest, because the envelope was measured rather than assumed. Corpus publishes the honest denominator next to its rule count. RSFQ-JEPA ships with a classical baseline that beats it on Ic margin.

Nothing we serve is validated against fabricated silicon. Every validation figure we publish names a simulated or analytic reference, and the manifest says so in its own first field.

Criteria are fixed before the run, not after the result. Our benchmark question was settled across four pre-registered runs, each hashed before execution — and it needed all four, because three earlier passes at n≈50 scattered across two different verdicts before the answer at n=200 held.

A finding that looked solid at n=40 was withdrawn at n=200. Three times in three days, a result computed at small n reversed, and every time the small-n reading had been seductive and directional. The lesson we took is short enough to repeat: a point estimate without an interval is not a measurement.

What we do not claim is that we always know when we are wrong. Calibrated uncertainty is a property some of our models have and others do not, measured per model, and stated per model. A company-wide version of that claim would be exactly the kind of thing this page exists to avoid.

Questions too small to staff still deserve answering. Those go to Cogitan Commons, a research community in quantum where anyone can propose a group. Every proposal names the result that would end it, and every result gets published — including the ones that do. A project that meets its own criterion and stops is recorded as finished, not failed.

It runs on the same discipline as the rest of this page, which is the only reason we think it is worth other people’s time.

Not a general-purpose AI company — we work one stack, on purpose, and declined the larger ones for the reasons above. Not a foundation model company. Not a consulting firm wrapping existing models in a thin domain layer. Not a hardware maker: we sell software for the superconducting stack, and we do not fabricate anything.

And not self-serve. There is no free tier and nothing on this site can be bought — access is reviewed and granted, partly because this work can touch export-controlled technology.