← What’s Lost with a Look
Control run, 21 October 2026
Order of events, from the two servers’ own records: plan and circuits pushed to GitHub at 23:00:47 UTC; IBM created the first job (ibm_kingston, dbck8c9gv2cc73eahbc0) at 23:01:05; the analysis script was pushed at 23:02:19; that first job finished at 23:02:31; ibm_marrakesh (dbck95c24mrc73cjqgi0) was created at 23:02:46 and ibm_fez (dbck9v424mrc73cjqiig) at 23:04:28. IBM charged 69, 71 and 68 seconds.
plan · run.py · SHA-256 of both · analyze.py (committed before any result was opened; SHA-256)
Results, read by the committed analysis (full output)
Every reading below agreed between the two copies unless marked; four outcomes allowed; 95 % intervals over qubits or pairs.
- Kick: −1.62° fez, −2.86° kingston, −2.39° marrakesh per readout, all clear of 0. The same size as earlier on 21 October (1.73 / 2.41 / 2.13°); the sign is the opposite only because this script reads the Y axis with the opposite convention (checked from the two circuit definitions), so the sign is not a finding. Untouched qubits precessing in either direction get it.
- Number read out: with 1 of the 39 qubits read out, the kick is at most 0.15° on fez and 0.09° on kingston (95 % bounds; −0.08 fez, +0.01 kingston, −0.03° marrakesh; copies disagree on marrakesh, cannot tell); reading out all 39 gives the full kick; with 10, −0.13, −0.28, −0.23° (8 to 10 % of the full kick, less than the proportional 26 %; fez: copies disagree, cannot tell; marrakesh copies read 0.32 and 0.48°, the pooled median smaller than either), so the kick is not proportional to the number read out (neither pre-registered alternative). The subsets of 1 and of 10 are one random draw (seed 20261021) and the single qubit is one position.
- Slope on precession rate: slope of kick on precession rate consistent with 0 on all three (one of kingston’s two cross-copy fits is below 0). The plan’s stretch of 9.1 % is excluded (below) on fez and kingston in both cross-copy fits and cannot tell on marrakesh (one fit consistent, one below); it is not used in the paper, where the extra-time bounds (check D below) are.
- Superposition cost (coherence after readouts minus after equal idle): marrakesh −0.011 (2 readouts), −0.016 (8); kingston −0.009 (2), −0.007 (4), 8 cannot tell (copies disagree); fez consistent with 0 at 2 and 4; where the committed rule reads copies as disagreeing (fez 8, kingston 8, marrakesh 4) a copy-gap check defined after the data finds no shift, and the pooled readings are fez −0.0061 [−0.0131, −0.0019] (below 0), kingston −0.0067 [−0.0144, +0.0015] (consistent with 0), marrakesh −0.0107 [−0.0168, −0.0062] (below 0) (script, output).
- Entanglement cost (concurrence after readouts minus after equal idle, per Bell pair): kingston −0.011 (2), −0.010 (4); marrakesh −0.009 (2), −0.016 (4); fez −0.005 at 4 readouts (at 2 consistent with 0). Below 0 on all three processors.
- Leakage test: invalid as designed. The pre-registered quantity reads +0.51 to +0.56 on all three, but a measured qubit in |0⟩ that is misread as 1 also reads 1 after the flip, so this number is dominated by readout misassignment and does not measure leakage. Leakage is neither ruled in nor out by this run.
- Fraction of 1 outcomes in the mid-circuit readouts: 0.016 fez, 0.032 kingston, 0.035 marrakesh (includes misassignment).
Data: kingston · marrakesh · fez job records; raw counts kingston · marrakesh · fez; IBM’s stated calibration at firing: kingston · marrakesh · fez.
Two checks made after the data (rules and script committed before running; output)
- Stuck measured qubits: measured qubits read 1 on two consecutive readouts in 0.4–1.1 % of shots, almost all of it persistent (excited or leaked, not chance misreads). The worst-case bound on what such neighbours can cost is larger than the measured loss for both coherence and concurrence on all three processors: not excluded. Untouched qubits with no measured qubit wired to them (31 per processor) show no resolved coherence loss.
- Where the kick lands: with 10 or 39 read out, qubits not wired to any readout qubit get the kick too (10: −0.11, −0.20, −0.13°; 39: −2.1, −4.3, −2.1° on fez, kingston, marrakesh): not only wired qubits. In this layout every untouched qubit is within 2 couplings of a measured one; the deciding runs show it 3 to 4 couplings away, and the reach run reads where it stops. With 1 read out the script prints NEIGHBOUR-LOCAL, but that rests on one neighbour with no interval; by the committed rule it is cannot tell.
Three more checks after the data (rules and scripts committed before each ran)
- Clean shots (rule, script, output): keeping only shots in which every measured neighbour read 0 on every readout (96–99 % kept), the coherence loss survives on all three (fez +0.005 / +0.010 at 2 / 8 readouts; kingston +0.008 / +0.007 at 2 / 4; marrakesh +0.010 / +0.016 / +0.014 at 2 / 4 / 8). Concurrence loss survives on fez (+0.004 at 4) and kingston (+0.010 / +0.009 at 2 / 4); on marrakesh it is no longer resolved. Not excluded by this check: a neighbour leaked to a level that reads as 0.
- Extra elapsed time (rule, script, output): bounding hidden extra time per readout by the stretch fit and IBM’s T2, it is excluded on kingston (2 and 4 readouts), not on fez or marrakesh, where the bound is too loose.
- Wired vs unwired (rule, script, output): any extra time or chip-wide cause acts on all untouched qubits; on fez, qubits wired to a measured qubit lose more than unwired ones in the same circuits (2 and 8 readouts), so such causes are excluded there; kingston and marrakesh cannot tell (about 30 unwired qubits each).
- Dose vs coherence (rule, script, output): at the same schedule, the coherence loss with 1 qubit read out is consistent with 0 on all three, and with 39 read out it is +0.006 (fez), +0.016 (marrakesh); the difference is clear of 0 on fez (+0.010) and marrakesh (+0.012), so extra elapsed time per step is excluded there; kingston cannot tell on this check (excluded by the extra-time bound above).
- Distance and readout count (rule, script, output; this run and the raw-readout run, same qubits; the two runs are pooled, as printed in the paper). Kick per readout after 8 readouts: 1.58, 2.80, 2.07° (fez, kingston, marrakesh). Kick per readout after 8 minus after 2: +0.036, −0.008, +0.294°, consistent with 0; it changes by at most 0.227, 0.166, 0.493° (14, 6, 24 % of the kick). Straight line through the phase gap at 2, 4, 8 readouts: intercept consistent with 0 on fez and kingston; −0.769° on marrakesh, a fixed offset of about a third of one readout’s kick. Kick size next to a measured qubit minus two steps away: −0.612, −1.599, +0.332°, not resolved on any processor (fez and kingston point larger two steps away). Superposition cost after 8 readouts minus after 2 (all shots): +0.0043 fez and +0.0120 marrakesh (above zero); kingston +0.0047, not resolved. Marrakesh clean shots: +0.0145. An earlier pass (check H, output) compared the two runs’ labels instead of their difference and read the sign of the distance test backwards; check I replaces its reporting.
- Backward spinners (rule, committed before computing; script, output; this run and the raw-readout run): untouched qubits spinning backward (spin rate ≤ −1° per step) get a kick of −1.640, −3.361, −2.278° per step here and −1.827, −3.340, −2.537° in the raw-readout run (fez, kingston, marrakesh; every interval below 0), the same direction as qubits spinning forward; backward minus forward is consistent with 0 everywhere. In this analysis’s Y convention most untouched qubits read as spinning backward; the 21 to 36 per processor spinning the other way (the minority the paper quotes) get −1.177, −3.032, −3.367° here and −0.960, −2.798, −1.523° in the raw-readout run. Extra time per step would give the two groups opposite signs.