Elizabeth R. S. Burnim

What this is

I built a GPU solver for the quasiclassical theory of superconductivity — the physics of Josephson junctions and proximity structures — with every term separately gated and the transport verified against references independent of the code.

For a team scaling superconducting readout, what exists today is a verified design tool for junction and proximity physics, and a design partnership on the first component.

The first component will be chosen by what the teams scaling those machines cannot buy. Read the research →

824commits, each attributed

43gates that print a verdict

3references independent of the code

9,736lines of solver

The record

What has been shown, dated.

Each entry is dated by the record that holds it and is held by a gate that prints it or a run record that names it. The open findings are listed with the rest, because they are part of the record too.

2026-08-21
Transport solved on GPU at production resolution. The two device gates print their first pass on hardware; the 1200-element run converges.
Pass
2026-08-21
Agreement with an independent integrator. A four-rung mesh ladder against an adaptive Runge–Kutta reference: relative L₂ disagreement at or below 6.5 × 10⁻⁹.
Pass
2026-08-23
Second-order convergence, pre-registered. Under a protocol registered before the run, the measured order of the transport solve is p = 2.00, forty checks.
Pass
2026-09-11
Reflecting walls, on device. The specular boundary condition passes 226 checks on hardware after 218 across nine configurations on host; walled runs converge in the same iteration counts as unwalled ones.
Pass
2026-09-11
The unit-disk excess. Along wall-normal trajectories the computed amplitude leaves its physical bound at the nodes where the pairing steps — present in every inhomogeneous run, its size tracking the element aspect ratio, its mechanism not established.
Open
2026-09-12
Equilibrium current, conserved. The current observable is gated at twenty named checks: the closed form at roundoff, conservation to better than 5 × 10⁻³ on converged transport at two resolutions.
Pass
2026-09-16
Free energy confirmed on hardware. Both device gates for the free-energy functional pass; the current-reduction kernel is exact on device at 1.7 × 10⁻¹⁵.
Pass
2026-09-16
One device gate has never passed. Its single hardware run failed with non-finite fields at near-grazing angles of a complex twisted state; the host reproduces none of it; the device confirmation is unrun.
Open
2026-09-17
Continuous verification. A fresh-clone battery — every check, every host gate — now runs on every push.
Pass

Verification first

Every term of the discretization is gated separately; every number in the record regenerates from a clean checkout; a commit that turns the battery red cannot land. The solver was built with AI-assisted development under that discipline — every change gated, every commit attributed to the model that made it, every result adjudicated against a protocol registered before the run.

Direction

The readout chain of a superconducting quantum processor still depends on discrete ferrite isolators and circulators on every line. Before choosing what to build, the work is customer discovery with the teams that scale these systems: what they cannot buy, what its absence costs, and who would fabricate a first device.

Cryogenic components for superconducting quantum processors.

Every readout line still carries ferrite isolators between the qubit and its amplifier.

What has been shown, dated.

Verification first

Direction

Three doors, each with the questions I will ask.