← What’s Lost with a Look · all measurements

Controls: decay, idle, drift

Runs. Decay-design and idle-control jobs on IBM processors, 1–7 October 2026 (job IDs in the data files).

What holds

• Idle-control residuals against each qubit’s own decay (ibm_pittsburgh): +0.0017 ± 0.0013, +0.0020 ± 0.0014, +0.0014 ± 0.0014.
• A qubit left in 0 for 300 µs sits 0.04 to 0.80 % above its own job’s zero-wait level in all 26 decay-design jobs on six processors (five of them under two widths, the smallest, +0.04 %, at 0.3 widths); at 600 µs three jobs are at or below zero, so the offset does not grow but is not constant.
• Decay-time copies within jobs: marrakesh 174.0 / 177.3 / 180.7 µs, kingston 260.4–265.6, pittsburgh 293.8–297.1; the largest job-to-job move of any copy is 9.8 µs, and copies within one marrakesh job span 10.4 µs.
• Measured-qubit loss over untouched-qubit time per chip-day: 1.207–1.784.

Run it

Download the script and the files in data/ into one folder (data in data/), then python3 verify_processors_apparatus.py. Standard library only; it prints each figure beside the recomputed value. The compact data were cut from the raw results; the raw files and the reduction scripts are available on request.

Outside referee check, 22 October, after the data (referee_controls.py, output): the 300 µs idle-in-0 offset is positive in all 26 jobs, does not grow by 600 µs and is not constant (mean 600 minus 300: −0.11 %, higher in 9 of 26; three jobs at or below zero at 600 µs); compared like with like, the observed range of three copies (3.3 to 11.5 µs) is inside the simulated 95th percentile of the range of three in 8 of 9 jobs (marrakesh job 301: 10.4 against 8.9 µs; the simulation takes its noise scale from the copies, so it cannot separate shot noise from drift; up to 9.8 µs between jobs and 11.5 µs between copies inside one job). Null check for the non-exponential share (observed 61.5–84.8 %, 26 jobs; a separate computation in the 22 October re-check file; the paper’s figures, recomputed by verify_processors_apparatus.py on this page, are 62.4 to 85.3 %; the baseline for a pure exponential with the same noise is 49 to 54 %, from the re-check of 22 October; the output below is the earlier baseline): null_check.py · output

Files: verify_processors_apparatus.py · data/pittsburgh_park_idle_push.json · data/spectator_and_measured_qubit_sort.json · data/time_runs_ruler_copies.json

Bones · Wayfinder · Sage · Shuttle · Nova
Heuremen LLC
“We found it.”