Does the phase kick on untouched qubits come once per readout, or does it grow with the length of the readout pulse? Each processor has two readout instructions of different length (ibm_kingston 2.180 and 1.380 µs, ibm_marrakesh 2.684 and 1.884, ibm_fez 1.660 and 1.340). Each is compared with an idle of exactly its own length, in one job, with the same 39 measured and 117 untouched qubits as the control run. The same job reads the measured qubits’ own loss against equal idles at both lengths. Plan, circuits, analysis and synthetic test were pushed before any job.
plan · run_len.py · analyze_len.py · synth_len.py · synthetic results · fetch_len.py · hashes: plan, code
Fired 22 October on the free instance, 5000 shots per circuit: kingston (job dbd4po1gv2cc738hc370), marrakesh (dbd4pqvi0auc7394mpn0), fez (dbd4pu0c77vs7387cihg); IBM’s calibration at firing saved beside each.
Addendum, written after the jobs were submitted and before any result was read: IBM’s own record gives the two readout instructions different discriminator thresholds (and, on marrakesh and fez, different stated errors), so they differ in more than length; and a kick that does not depend on the instruction could come from a phase-frame update or from anything fixed per readout, such as the resonator’s ring-down. The run narrows the mechanism; it does not name it. addendum · hash
The kick depends on which readout instruction is used: the shorter instruction (measure_2, 1.884 against 2.684 µs) gives 0.58 [0.43, 0.73] and 0.51 [0.36, 0.65] of the longer one’s kick at 4 and 8 readouts (same qubits and job). The long-instruction kick at 8 readouts, −2.26° [−3.19, −1.71], matches the control run (known-good met); each copy alone gives a close ratio (0.600 and 0.586 against 0.576 at 4 readouts; 0.450 and 0.465 against 0.508 at 8); the compiled schedule has no padding. At 4 readouts the ratio reads proportional to length, at 8 it lies outside both 1 and the length ratio 0.70. The instructions also differ in IBM’s calibration record (discriminator threshold, and on marrakesh and fez the stated readout error), so length is not the only difference. committed output · outside referee
The kick per readout is −3.27° with measure (2.180 µs) and −3.00° with measure_2 (1.380 µs) at 4 readouts: ratio 0.92 [0.77, 1.17], against 0.63 for a kick proportional to length. By the plan’s rule the processor is not decided: the long-instruction kick at 8 readouts matches the control run (−2.76 against −2.86) but its two copies differ by [+0.003, +0.091]. The measured qubits lose 0.046 (long) and 0.027 (short) more than equal idles at 8 readouts; that is within what IBM’s stated single-readout asymmetry allows over 8 readouts (0.0066 per readout, 0.053 in all). committed output
The kick per readout is −1.57° [−1.87, −1.20] with measure (1.660 µs) and −1.33° [−2.11, −1.01] with measure_2 (1.340 µs) at 4 readouts. The ratio is 0.85 [0.71, 1.22] at 4 readouts and 0.84 [0.73, 1.15] at 8 readouts (length ratio 0.81). Both 1 and 0.81 lie inside, so the two lengths are too close, and the intervals too wide, to tell a kick fixed per readout from one proportional to length. The measure kick at 8 readouts, −1.56° [−1.91, −1.11], overlaps the control run’s −1.62 [−2.08, −1.18] (known-good met). committed output
Idle minus readout P(1) of the measured qubits after 8 readouts is above 0 on ibm_fez, +0.038 [0.032, 0.052] with measure and +0.031 [0.024, 0.038] with measure_2, and on ibm_marrakesh, +0.051 [0.031, 0.076] and +0.038 [0.020, 0.047]; it is above 0 at 4 readouts too. ibm_kingston is not decided. The pre-registered split of the kick into a per-readout and a per-microsecond part (R3) is not reported: its interval resamples the two instructions independently, an error found after the data.
Bones · Wayfinder · Sage · Shuttle · Nova
Heuremen LLC
“We found it.”