Cogitan

Quantum discovery · Applied AI

We find out what is true
in the superconducting stack.

Simulation is the method. Surrogates are simulation made cheap — we built ours to do our own research, and sell them because they work.

What “simulation made cheap” means — RSFQ verification, MIT-LL SFQ5ee+

verify.py
import cogitan

# load your cell library and run verification
results = cogitan.verify(
    path = "./layouts/sfq5ee+",
    pdk = "mit-ll-sfq5ee+",
    batch = True,
)

for cell, r in results.items():
    print(
        f"{cell:<14} {r.delay:>5.1f}
        f" ± {r.sigma:.1f} ps  {r.status}"
    )
$ cogitan verify ./layouts/sfq5ee+

SPLIT_v1_a     17.2 ± 1.4 psSPLIT_v2_b     18.9 ± 1.1 psDFF_x3_c       22.1 ± 5.8 ps   → spice
AND2_v1_d      14.4 ± 1.6 ps...

─────────────────────────────────
1,000 cells  4.8s    (SPICE: ~85 min)
997 pass  ·  3 → spice fallback

Wall-clock time · chip-scale verification

1K cells~100× speedup
SPICE
5–10 min
Fluxus
5s
10K cells~120× speedup
SPICE
~90 min
Fluxus
45s
100K cells~140× speedup
SPICE
8–10 hrs
Fluxus
4–5 min

The analytic engine is validated against JoSIM. See the instruments →

What we are working on now

Open questions, each with the result that would settle it. Stating the settling condition first is how we decide what to stop.

What must a device paper report for its claims to be independently checkable?

The median published device reaches 7–8 rules of our set. Saturated with what its own field already knows how to publish, it reaches 28 — so the binding constraint is reporting convention, not physics or tooling. Growing the database from 16 to 60 superconducting rules barely moved the number.

Settled if

A minimal reporting set lifts median reach from ~8 toward 28 across a large corpus — and closing it changes verdicts, not just counts.

How tight is the structural floor on logical error for dissipative cats?

The ideal-code floor is a valid lower bound but runs 6–14× low against the dissipative cat. The missing factor is set by the confinement ratio κ₂/γ and buffer adiabaticity g₂/κ_b.

Settled if

An effective gap for the dissipative cat closes the floor-to-rate distance across the κ₂/γ sweep — or is shown not to exist.

Does the readability limit extend to codes read by canonical phase measurement?

Our preprint bounds logical decoherence from a code's own structure, but the framework cannot express canonical phase measurement — which is how number-phase codes are conventionally read. So the result carries a scope restriction we state rather than hide.

Settled if

An extension of the bound to phase measurement, or a proof that the restriction is essential.

All research →

Cogitan Commons

The frontier is worth more open than closed.

We run 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 stops on its own criterion is recorded as finished, not failed, because quantum has nowhere else to put that.

commons.cogitan.ai →

How we work

01

We publish the baseline that beats us

RSFQ-JEPA ships with a gradient-boosted-tree baseline that wins on Ic margin — 0.006 against our 0.093. It is in the repo's own results table, not a footnote.

02

Our tools refuse questions they cannot answer

Fluxus-S serves 3 of 8 SFQ cell types and declines the rest, because the envelope was measured rather than assumed. A surrogate that answers everywhere is a surrogate you cannot trust anywhere.

03

We withdraw results that do not survive more data

A finding that looked solid at n=40 was retracted at n=200. Three times in three days, a result computed at small n reversed. A point estimate without an interval is not a measurement.

Built for you and your agent

Designed to be read by a program
before it's called by one.

An agent driving design tooling is dangerous for one specific reason: it will report a number from outside a model's valid range with exactly the same confidence as one from inside it, and the engineer reading the output cannot tell which happened.

So the engine is built to be interrogated first. Everything it knows about its own limits is machine-readable, and it would rather refuse than guess.

A manifest, not just a spec

One call returns, for every function: what the ground truth behind it actually is, its measured held-out error and the reference that error was measured against, the input envelope it's trusted inside, and the conditions under which it refuses. An OpenAPI document says what can be called. This says what you may believe.

Refusals an agent can repair

Ask for something outside the validated envelope and you get the field, the value we received, and the range we allow — as data, not prose. Your agent corrects the call and continues. It never has to parse an English sentence to work out which parameter was wrong.

It abstains instead of guessing

Outside its trained region a prediction comes back flagged, carrying the analytic oracle's answer rather than a confident extrapolation. Every result names the engine that produced it, so a surrogate answer is never mistaken for a solver answer.

Including the numbers that don't flatter us

The manifest publishes the classical baseline each surrogate was measured against — and where the baseline wins, it says so in the same field your agent reads before choosing. On RSFQ margins, a gradient-boosted tree is the more accurate model. You should know that from us, not from your own benchmark.

An MCP server ships with access, so the same functions appear as tools in Claude and any other MCP client — each one carrying that function's fidelity, envelope and refusal conditions in its description.

Start here — costs nothing

curl https://api.cogitan.ai/v1/fluxus/capabilities \
  -H "Authorization: Bearer cog_sk_..."

If you are building in this stack, we would like to hear what you are stuck on.