← What’s Lost with a Look
Deciding runs, 22 October 2026
Six readings from our earlier record, each read again on IBM processors beside the ordinary causes that could make it. Plan, circuits and analysis pushed before any job: plan version 3 was committed at 02:22:06 UTC and pushed before any job; IBM created the first job at 02:23:24 UTC. The analysis was tested on synthetic data with planted effects and with no effect (floor: 10 of 210 null readings resolved, 4.8 %, as recorded in the plan, the saved null outputs not regenerated for this count; on hardware, copy 0 minus copy 1 of the control run resolves 2.0 % of 147 readings with the median and 4.8 % with a 20 % trimmed mean). Every reading is a difference between two arms of the same job, with a 95 % interval; every circuit ran twice in its job; chips are combined by value and width.
plan · runs.py (uses run.py) · analyze.py · synth.py · synthetic test · plan SHA-256 · code SHA-256 · full output (with ibm_phoenix; output before phoenix returned)
The bootstrap draws come from one seeded random stream, so adding ibm_phoenix shifted them for every later reading. Two single-processor labels changed (fez neighbour 1 minus 0: below 0 → consistent; marrakesh read order slot 2: cannot tell → consistent); no across-processor label changed. An outside referee re-ran the order, neighbour and gate readings over 40 seeds (script, output): every across-processor label was stable 40 of 40; fez neighbour 1 minus 0 came out below 0 in 31 of 40 seeds (the printed “consistent” is the minority outcome), marrakesh neighbour + minus 0 is cannot tell in 24 and consistent in 16, kingston gate cost above 0 in 6, fez gate cost minus stated below 0 in 6. The printed outputs stand as committed; no seed was re-picked.
Results (fez, kingston, marrakesh; ibm_phoenix in the gate and revival runs)
- How far the kick reaches. With the 39 measured qubits packed in one region: untouched qubits 3 to 4 couplings from the nearest one get 0.8068, 1.766, 1.298° per readout (fez, kingston, marrakesh; 15 qubits, every interval above 0). Within 2 couplings: 3.458° on kingston; fez copies shifted, marrakesh not resolved (10 qubits). Over all 92 qubits 5 or more away the median is +0.0007822, +0.004896, +0.006988° (intervals up to +0.07055), consistent with zero on all three (+0.0039 [−0.029, +0.037] combined; that pool is dominated by qubits six or more couplings away, and at exactly 5 the reach run reads +0.055 [+0.002, +0.108] combined, at the false-reading floor). Where it stops is read in the reach run.
- No wear-out is resolved over 72 readouts against equal idles on any of three processors. Restart run (plan committed 02:48:46 UTC and pushed before any job, IBM jobs created from 02:49:41; freshclock.py, analyze_freshclock.py, synthetic test, output): the untouched qubits sit in 0 while the measured qubits do 0, 27 or 72 readouts or 72 idles, then are prepared fresh on the equator and take 9 readouts against 9 idles. Window coherence 0.528 to 0.756 (phases are readable). Kick size after 72 readouts minus after none: +0.01532 fez, +0.02869 kingston, −0.1306° marrakesh; intervals: fez [−0.07602, +0.1185], kingston [−0.01897, +0.07381], marrakesh [−0.2059, −0.03681] on a kick of 2.569°. The chips disagree (spread chi2/dof 5.33), so the combined value (−0.004662 [−0.04219, +0.03286]) is not used. Against no preload, marrakesh reads 0.13° less [−0.2059, −0.03681] of 2.569°, with window coherence 0.528 versus 0.602. Against 72 equal idles (the matched comparison), after 72 readouts minus after 72 idles: +0.02776 [−0.0377, +0.08501] fez, +0.005744 [−0.03408, +0.05893] kingston, −0.03605 [−0.1184, +0.1038] marrakesh, consistent with zero on all three (the fez 72-idle arm had shifted copies).
- Why the count run could not answer this. Readouts 1 to 9 kick 1.559, 2.819, 2.844° each (offset-free, rule, output), and readouts 27 to 81 read 0.6338, 0.7258, 0.4261°. But by 54 to 81 readouts most untouched qubits have no coherence left to read a phase from: 67 to 97 of 117 sit below 5 times the shot noise, and the phase noise per qubit is 14 to 30° (script, output, computed after the data). A phase step that is pure noise has a median of 0, so qubits with no coherence left alone make the late segment read small. The early segment stands; the fade is not established.
- Does the superposition cost compound per readout? Cannot tell: the processors disagree at 9 and at 27 readouts, and past 27 most untouched qubits are at the noise floor. Per readout, as a fraction of the coherence left: 0.12, 0.17 and 0.53 % at 9 readouts and 0.13, 0.25 and 0.39 % at 27 (fez, kingston, marrakesh); the processors disagree (chip-to-chip χ²/dof 7.17 and 2.86), so no combined value is quoted. At 27 to 81 readouts the coherence of most untouched qubits is at the noise floor, where its magnitude is biased upward in both arms and the ratio is pulled toward 1; the per-readout share at 81 and the sum-form comparison are not read.
- Cost of an extra X gate. With the decay shape held equal (X X X against X on the same waits), an extra X gate costs 0.0001643 [0.00005906, 0.0002695] combined over four processors (fez, kingston, marrakesh, phoenix), above 0. Against IBM’s stated X error the difference is −0.0002279 [−0.0003419, −0.0001139] combined; on marrakesh alone it is below the stated value, on the other three consistent with it. This train does not see everything the stated error includes, so “below” is not a claim that the gates are better than stated. On fez and kingston, where both were measured, the record’s per-gate residual after spaced X trains (−0.0006 to −0.0020 per gate on four processors) is opposite in sign to the extra-gate cost and at least twice its upper edge (0.00027), so it is not a gate cost there; no extra-gate test exists on boston or pittsburgh.
- A neighbour in superposition does not raise gate error. Error with neighbours in + over neighbours in 0: 1.000, 0.995, 0.999 (an earlier reading in our record put this ratio at 1.7); + minus the mean of 0 and 1: −0.00000129 [−0.00001327, +0.00001069]. A neighbour in 1 rather than 0 lowers it slightly: −0.0000162 [−0.0000305, −0.0000019] combined.
- Read order. Excited population read after other groups minus read alone at the same moment: slot 1 +0.00002755, slot 2 −0.0000217, slot 3 −0.00001549, every interval inside ±0.00018. Per processor up to ±0.00033 (kingston slot 1 copies shifted). Reading all qubits together rather than one alone raises the excited population of qubits held in 0 by +0.000166 [+0.000025, +0.000306] (three processors combined; each processor’s interval touches zero). The qubit read last sits below the qubit read first on marrakesh (−0.000375 [−0.0005, −0.000167]); fez and kingston include zero, and the chip-to-chip χ²/dof is 3.05, so no combined value is quoted. These single-processor labels are sensitive to the bootstrap seed.
- The kick is not bigger with the measured qubits in 1 (claim 4; plan pushed 02:37 UTC before the jobs, excitation.py, analyze_excitation.py, output). Kick in 1 minus kick in 0, per step: +0.178 fez, +0.06599 kingston, +0.9226° marrakesh [kingston: its kick in 0 had shifted copies (cannot tell), and the difference, −0.005397 to +0.229, is consistent with zero; no bound taken from it]. The kick is negative in this convention (in 0: −1.522 fez, −2.967° marrakesh), so a positive difference means a SMALLER kick in 1; the plan’s line “above 0 means bigger in 1” assumed a positive kick and is read here with the sign carried. Same pattern as the earlier two-chip run (check J: kingston no difference, marrakesh smaller). Holding the measured qubits in 1 without any readout shifts the phase by −2.531, −3.241, −3.362° per step.
- GHZ flatness. Not run in these deciding runs: P(000) + P(111) with τ/3 offsets is exactly constant (1/16) in textbook quantum mechanics, so its flatness ratio is infinite in theory; a ratio measured in an earlier run would show how close the hardware comes to an identity, not a physical effect.
- Coherence comes back on ibm_phoenix. On ibm_phoenix (square lattice), untouched qubits’ coherence falls to the measurement floor by about 26 µs when their neighbours are held in superposition and then returns briefly at 58–78 µs (+0.039 [+0.020, +0.046] above the 26–50 µs level, n = 60; 48 of 60 qubits); on ibm_fez the same quantity keeps falling (−0.168 [−0.204, −0.086], n = 92); neighbour on the equator minus neighbour in 0 is +0.243 [+0.211, +0.263] on phoenix and +0.002 [−0.017, +0.025] on fez. On phoenix, with neighbours in 0 the coherence only decays (−0.204 [−0.221, −0.187]), and with neighbours in 1 the copies disagree (cannot tell). ibm_fez shows no come-back within 150 µs in any state (−0.1725, −0.2389, −0.1678). An outside referee’s model after the data, using IBM’s calibrated per-edge ZZ scaled by a fitted factor of about 0.5 on phoenix, tracks the curves better than shuffled couplings do (0 of 60 shuffles); whether phoenix’s calibrated value is stated under a different convention is not known from these data; the pre-registered flat-beat run later found the beat does not scale with the calibrated value (slope 0.19 against 1), which a single convention factor would not produce. The processors disagree far beyond their widths and are not combined. Outside referee’s recomputation and model: load.py, model.py, model2.py. A free fit by the referee after the data (not pre-registered, not in the paper) puts the beat near 12.5 kHz, which would return near 80 µs, where this come-back sat; see the one-neighbour run.
Job records (job_*.json), IBM calibration at firing (props_*.json), raw bits (raw_*.npz) and fire logs are in this folder.