Evidence

Validation you can check yourself.

The hardware does not exist yet. What is validated today is the physics and the design, in simulation, end to end. Every number on this page is reproducible: the figures below come from open code and pinned data, and the SDK that powers them is on PyPI right now. We would rather find our own hard numbers than have a reviewer find them for us.

Fault tolerance
Surface-code logical error rate versus physical loss per cycle, full circuit-level noise on the left panel and the phenomenological control on the right; the logical curves cross at the threshold.
threshold.png · stim + PyMatching · left: full circuit-level noise, right: phenomenological control

Fault-tolerance threshold, full circuit-level noise

Under realistic circuit-level noise, with noisy syndrome extraction decoded by stim and PyMatching, the surface-code logical error rate crosses at 0.79 to 0.90 percent loss per cycle. The right panel is the easier phenomenological control, shown for contrast: the circuit-level number is the honest, harder one. We operate the machine at 0.4 to 0.5 percent per cycle to keep margin below it. Finding our own harder threshold before a reviewer did is the entire point of running the full noise model.

Independent reproduction

Independent cross-check, two codebases

Cat fidelity computed two ways: once in QuTiP, once in a fully independent Strawberry Fields implementation. The two agree to about 1e-5, far inside the 1e-2 target. This is independent reproduction across separate quantum-optics stacks, not a single-codebase claim that could hide a shared bug. When two unrelated tools land on the same number, the number is real.

Cat-state fidelity computed by QuTiP and by an independent Strawberry Fields implementation, overlaid; the two curves agree to about 1e-5.
crosscheck.png · QuTiP vs Strawberry Fields · agreement ~1e-5, target 1e-2
Manufacturability
Predicted resonance-spec yield from a 3D electromagnetic and fabrication Monte Carlo as a function of thermal tuner range; yield reaches about 99 percent with plus or minus 2 to 3 nm of tuner authority.
fab-yield.png · 3D EM + fabrication Monte Carlo · yield vs thermal tuner range

Manufacturability, 3D EM + fabrication Monte Carlo

A 3D electromagnetic model fed through a fabrication Monte Carlo predicts resonance-spec yield of about 99 percent once the thermal tuners have +/- 2 to 3 nm of authority on the thick film. The key design finding: the yield-recovery knob is tuner range, not tighter lithography. You buy yield with control authority you already have, not with a process node you do not.

Tapeout-grade

Tapeout-grade test chip, QF-TC1

QF-TC1 is a 5x5 mm validation chip with 22 structures, DRC clean at 0 violations. Each structure maps to a specific simulation claim it is built to measure on silicon, so the chip is a checklist for the physics, not a generic test die. It is drawn on a foundry-neutral stand-in rule deck, and the layout has since been retargeted onto the production foundry PDK (GlobalFoundries 45SPCLO) with the design rules already satisfied.

Layout of the QF-TC1 test chip: a 5 by 5 mm tile holding 22 validation structures, DRC clean with zero violations on a foundry-neutral stand-in deck.
qf-tc1.png · 5x5 mm, 22 structures · DRC clean, 0 violations, foundry-neutral stand-in deck
Reproduce it

Run the model yourself.

The same resource and noise model that powers the figures above ships in the SDK. Install it and print the device spec for an H-cat circuit.

Install
pip install dyber
Build a circuit and read its resources
from dyber import Dyber, Circuit

c = Circuit(2); c.h(0); c.cx(0, 1); c.measure_all()
dy = Dyber()
job = dy.backend("local_simulator").run(c, shots=1000)
print(job.result().resources)

At the 0.5 percent per-cycle loss operating point this prints code distance 53 and 5,617 cats per logical, the same model that powers the Resource Analyzer on this site. The resource and noise models are open source in the dyberforge package.

Methods and tooling

The actual stack.