The design that survives is the one whose corpses are documented.
The slow readout cavity died first.
v1 used κ/2π = 1.5 MHz and managed only 97.5% fidelity — below spec. The fix was a fast
cavity (4 MHz) with a Purcell filter to protect T1, which the filter then had to earn (see below).
A 0.9% “leakage” that wasn’t.
Three calibration rounds of the 27-dim CZ model plateaued at 99.0% with stubborn 0.9% leakage.
It was the metric: projecting onto bare states counts the qubit–coupler static dressing
((95/1150)² ≈ 0.7%) as loss. In the dressed basis — what a lab actually calibrates — true
leakage is 1.3×10⁻⁵.
The original frequency plan was killed by Monte Carlo.
A/B separation of 250 MHz sits 34 MHz from the |11⟩↔|02⟩ collision at |α| = 284 MHz;
at σEJ = 0.5% chip yield collapsed to 4.5%. Moving to 200 MHz separation with
sweet-spot operation and junction trimming recovered 99.6%. Bonus finding: “pull every qubit
to its exact target” is unworkable — flux-noise Tφ pays for every MHz of pull.
g was unmeasurable until the qubit crossed the cavity.
Dispersive-regime mode pulls (~15–30 MHz) drowned in ±7 MHz of per-design adaptive-mesh
noise; a three-parameter fit returned nonsense. Sweeping LJ through the crossing
and reading the minimum splitting (2g ≈ 350 MHz signal) made it robust in one shot.
Four dead κ variants, three separate bugs.
Eigenmode terminations absorbed nothing (Q ≈ 5×10⁶) until: ① sheet resistors sharing mere
edges with metal never mesh-bond — overlap them 1 µm into the metal; ② one feedline end was
accidentally drawn shorted, making the line resonant and un-drivable at 7 GHz; ③ an R-gate
(rerun with R=50 Ω vs 5 kΩ; Q must move ~50×) — borrowed from a sister project’s pitfall
table — is now mandatory for any lossy boundary.
Side-coupled Purcell filter: rejected by 800×.
The elegant side-coupled λ/2 filter measured Qf = 22 877 against a target of 28.
No amount of tuning closes three orders of magnitude — the topology was wrong. A direct-tap
filter (Qf = (π/2)/sin²(πx/L)) measured 16.5 at the first try and trims to 28.
An HFSS driven-modal hang that ate four suspects.
After a host reboot, every driven solve of the etched-CPW models spun forever at pass 1
(1 core, flat 183 MB, zero result files) while eigenmode solved fine. Dual-instance lock
contention, scheduler environment, Fast-sweep type and MPI were each accused and acquitted by
single-variable experiments. Root cause still open; the eigenmode method family became the
workaround — and delivered everything driven mode owed.
Seven GDS render iterations.
Resonator blocks overrunning feedlines, a missing shared-slot cut, two control lanes merged
into one slot (a short), bond pads stacked on top of each other, and the witness structures
colliding with two different feedlines. Render → look → fix, seven times, then DRC-lite clean.
