We executed the released circuit again on ibm_phoenix with a hard 60-second QPU execution cap. The job completed with 19 QPU seconds.
The new hardware job
On 25 September 2026 we sent the released 61-qubit, 36-cycle circuit to ibm_phoenix. It used the same logical gates and published physical layout, with 918 CZ gates after compilation. One IBM SamplerV2 job contained ten blocks of 100,000 shots. Measurement twirling was enabled; max_execution_time was set to 60 seconds. Job dar90plvr3kc73eij3vg finished with one million outcomes and 19 seconds of QPU usage. The circuit-execution field reported 15.127 seconds; submission-to-completion took more than seven minutes. These are different clocks.
What do the outcomes contain?
Each outcome is one string of 61 zeros and ones. All million observed strings were distinct. They contained an average of 30.440323 ones. The paper’s published million strings contained 30.423606 ones on average. The total-variation distance between the two histograms of the total number of ones was 0.003272. Across the 61 individual bit positions, the root-mean-square difference in P(bit=1) was about 0.002092.
Those are concrete agreements in simple statistics. They do not tell us whether outcomes with high ideal probabilities occur more often. Two very different 61-bit distributions can share the same number-of-ones histogram.
What we did and did not repeat
We repeated the scale, circuit, shot count and 19-second QPU timing. We did not run a full new mirror-and-patch series. We therefore do not assign the paper’s approximate 0.0023 fidelity to our new million bitstrings. A separate 21-qubit patch pilot gave IBM XEB about 0.313 and a Fire Opal point estimate about 0.385; that smaller patch does not measure full-circuit quality.
The paper repository contains the original circuit and published raw measurements. Our new job ID and figures are recorded above; the new raw outcomes remain in our local qlab project.
Project overview and article source
Edukaizen · 25 september 2026 / 25 September 2026