Your Fault-Tolerance Numbers Might Be Off by 10x, and the Simulator Won’t Say So
Here’s an uncomfortable possibility for anyone building a fault-tolerant quantum computer. The simulation telling you you’re below threshold might be scoring a device that doesn’t exist. Not yours. A cleaner, better-behaved cousin of yours.
That’s the claim at the center of Plaquette, the design platform QC Design just detailed in a paper on arXiv. The number attached to it isn’t small. Clifford-only simulation, the workhorse of the field, can be optimistic about logical error rate by more than an order of magnitude.
Call it the abstraction gap. The distance between the noise your simulator assumes and the noise your hardware actually makes.
Give the standard tools their due. Fast stabilizer simulators are the reason large-scale error-correction studies are possible at all. They run at the scale of real architectures, they’re quick, and they let a team rank which imperfections to fix first. Genuinely useful work. What they assume is where it gets fragile: stochastic Pauli noise, a tidy statistical picture where errors are discrete bit and phase flips that strike at random.
Real qubits didn’t sign that contract. Superconducting transmons leak out of the computational space entirely. Neutral-atom gates scatter through intermediate states. Trapped ions heat as their motional modes soak up phonons, silicon spin qubits leak into valley states, and a miscalibrated control line over-rotates coherently and not at random. None of that is Pauli noise. Some of it isn’t even stochastic.
So the field patches. Pauli twirling, depolarizing stand-ins, hand-built noise models tuned per device and per error process. These are directionally correct, and for some questions they hold up fine. But each one demands expert effort, and each one certifies the abstraction and not the device. You get a trustworthy answer about a machine that was never going to be built.
Why today’s fault-tolerance simulations can miss the mark
Plaquette’s move is to close the distance at the source. A team specifies its hardware error model once, as Kraus operators, as Hamiltonian-Lindblad dynamics, or as an experimentally reconstructed channel. The platform compiles that description into whatever representation four samplers need: Pauli-twirled stabilizer simulation for the easy cases, a new XPauli sampler for leakage and environment effects, near-Clifford samplers for coherent errors, and full-state simulation as an exact reference. The claimed ceiling is tens of thousands of qubits.
The validation is the part worth analyzing. QC Design checked its XPauli and near-Clifford samplers against full-state simulation and reports agreement within statistical uncertainty, including in regimes where Pauli twirling breaks down. They ran it on three concrete failure modes: leakage in superconducting qubits, intermediate-state scattering in neutral atoms, and heating in trapped ions. Different platforms, different physics, one framework.
CEO Ish Dhand frames the pitch around three questions every hardware team is already asking: “Is my device below threshold, and by how much? Which imperfection is most important to suppress? What logical error rate will my FTQC deliver, and at what overhead?” His argument is that Pauli approximations can answer all three wrong, and by a wide margin.
The evidence – and the problems
A note of caution before anyone treats this as settled. This is a company publishing the case for its own flagship product, and the size of the discrepancy varies by platform and noise process, so it won’t always reach an order of magnitude. At the very least, though, the paper makes a claim the field can check, because the reference point is full-state simulation rather than another approximation. If the numbers hold, a fair number of published threshold results deserve a second look.