21 October 2026, plan hashed locally before the jobs, not externally timestamped (plan SHA-256 9c985e3c…, plan). The superposition run found that untouched qubits lose a little more coherence during the first readouts than during an equal idle. An ordinary explanation: the measured qubits, held in 0, are occasionally left in 1, and an excited neighbour dephases its partners through ZZ coupling. This run tests that explanation directly.
Same 39 measured and 117 untouched qubits; all untouched qubits on the equator; 0, 4 or 8 steps of four kinds: idle with the measured qubits in 0 (D0), readout with them in 0 (M0), idle with them deliberately in 1 (D1), readout with them in 1 (M1). Every mid-circuit readout is recorded, so the accidental-excitation rate is measured directly. Readout excess X = extra coherence loss after 8 readouts versus 8 idles; excitation cost Y = extra loss with a neighbour held fully in 1 for the same time; the ordinary explanation predicts X = (excitation rate) × Y.
The planned test could not do its job. It predicted the readout excess as (excitation rate) × Y, where Y is the coherence cost of a neighbour held fully in 1. But a neighbour held in 1 applies a fixed phase, and coherence |⟨X⟩ + i⟨Y⟩| ignores a fixed phase, so Y ≈ 0 and the prediction is zero by construction. The declared output (here) is kept as it came out.
Corrected after the data. A neighbour that is excited only in some shots applies its phase in those shots only, which does cost coherence: about p (1 − cos θ), with p the excitation rate measured from the recorded mid-circuit readouts and θ the phase measured in the idle-in-1 arm. reanalysis_random_phase.py · output
| chip | excitation rate | untouched qubits | coherence excess after 8 readouts | from random excitation |
|---|---|---|---|---|
| ibm_marrakesh | 4.8 % | next to a measured qubit (86) | +0.025 [0.012, 0.036] | +0.0008 |
| ibm_marrakesh | no measured neighbour (31) | +0.010 [0.000, 0.019] | 0 | |
| ibm_kingston | 3.1 % | next to a measured qubit (86) | +0.004 [−0.003, 0.014] | +0.0005 |
| ibm_kingston | no measured neighbour (31) | +0.009 [−0.006, 0.018] | 0 |
On ibm_marrakesh the excess next to a measured qubit is far above what random excitation of the measured qubits can produce (+0.025 against +0.0008), and on qubits with no measured neighbour, where that model predicts zero, it is +0.010. On ibm_kingston the excess is +0.004 [−0.003, 0.014] against +0.0005, an interval that contains both 0 and the raw-readout run’s kingston cost. Neither reading of the planned prediction (zero by construction) tests excitation. Caveats: the excitation rate includes readout misassignment (which makes the prediction larger, not smaller); one kingston measured qubit read 1 in 75 % of shots. Kick in 1 minus kick in 0 from this run, re-read after the data (check_j_gap.py, output): +0.204 [−0.092, +0.320] on kingston (larger in 1 by at most 0.32°; consistent with zero) and −0.673 [−0.982, −0.403] on marrakesh (smaller in 1). Holding the measured qubits in 1 without any readout shifts the untouched qubits’ phase by 3.10° (kingston) and 3.12° (marrakesh) per step (2.5 to 3.4° on three processors in the 22 October deciding runs; phase_by_arm.py, output).
Files: plan (SHA-256 9c985e3c…, local, not externally timestamped) · run.py · analyze.py · job kingston · job marrakesh · raw: marrakesh, kingston · submission log. analyze.py reads the qubit roles from countwide_roles_<chip>.json.
Bones · Wayfinder · Sage · Shuttle · Nova
Heuremen LLC
“We found it.”