iqm_dla_window_variability_epoch_amendment_2026-09-04.md

IQM DLA window variability: calibration-epoch amendment

Status: prospective amendment after observation window 7

Original preregistration: iqm_dla_window_variability_prereg_2026-07-22.md

Effective for provider decisions: before any observation-window 8 submission

Why this amendment exists

The frozen protocol used a wall-clock gap of at least 12 hours to define a new window. That gap is an operational spacing rule, not evidence that two hardware observations are statistically independent. The retained calibration evidence makes the issue concrete: observation windows 2 and 3 were approximately 12 hours apart but both used calibration set c2097be4-1e23-49bc-adaa-8e8c01df6223.

This amendment does not rewrite the preregistration or its completed observation windows 1 through 7 primary result. It adds a mandatory, separately labelled sensitivity analysis and changes how subsequent observation windows 8 through 10 are counted for inference.

Units and counting rules

1. A nominal window remains one complete, crash-safe 36-circuit execution. It remains the unit of provider custody, budgeting, and the original frozen analysis.
2. A calibration epoch is the set of nominal windows carrying the same exact provider calibration_set_id. Windows with an identical ID MUST belong to the same epoch regardless of elapsed wall-clock time.
3. A changed calibration-set ID establishes a distinct operational epoch. It is not claimed to prove complete physical independence; slow drift can cross calibration boundaries. The epoch analysis is therefore a sensitivity analysis, not a retroactive replacement primary endpoint.
4. Multiple nominal windows inside one epoch are technical replicates. Pool their raw leaked and total counts by arm before estimating each depth's even-minus-odd difference and binomial variance. They contribute one epoch, hence one inverse-variance observation and no extra heterogeneity degree of freedom.
5. Report the original window-level result and the epoch-pooled result together. Neither may be selected or suppressed based on which p-value is preferable.

Prospective observation-window 8 through 10 scheduling

The fixed 12-hour delay is removed as a sufficiency claim. A fully prepared window may run whenever its simulator, custody, calibration, layout, budget, claim, and explicit owner-GO gates pass. If its calibration-set ID matches an existing epoch, it is recorded as a technical replicate and does not increase the inferential epoch count. If the scientific purpose is to add a between-epoch observation, wait for evidence of an actual calibration-set change rather than waiting an arbitrary number of hours.

The original cap of ten nominal windows remains. Do not spend credits merely to make an epoch count equal ten, and do not exceed ten nominal windows or the 35-credit cap without a new prospective protocol. If observation window 10 closes with fewer than ten distinct epochs, report that fact directly.

Epoch-pooled sensitivity analysis

At each depth, technical replicates in epoch e pool to arm proportions

p_even(d,e) = sum(leaked_even) / sum(shots_even) and p_odd(d,e) = sum(leaked_odd) / sum(shots_odd).

The epoch effect is Delta(d,e) = p_even(d,e) - p_odd(d,e), with the same plug-in binomial variance formula as the frozen analysis using the pooled arm totals. The companion analyser then applies:

• d10 Cochran Q across calibration epochs at alpha 0.05;
• d4, d8, and d12 epoch-level Q tests with Holm correction;
• descriptive DerSimonian-Laird tau estimates by depth; and
• descriptive d4 sign stability across epochs.

The epoch analysis becomes interpretable at six distinct epochs, matching the original minimum-unit threshold. A more elaborate nested random-effects fit is not promoted at the current data shape: through observation window 7 only one epoch has more than one nominal window, so a separate within-epoch random-effect variance is weakly identified. Raw-count pooling is the prespecified conservative treatment until enough repeated epochs exist to support a new prospective hierarchical model.

Current frozen mapping through observation window 7

The seven nominal windows form six calibration epochs. Observation windows 2 and 3 are the only technical-replicate pair; observation window 1 and observation windows 4 through 7 each have distinct calibration-set IDs. This mapping and both analysis tracks are materialised by scripts/analyse_iqm_dla_window_variability_epochs.py in data/iqm_paper_replication/iqm_dla_window_variability_calibration_epoch_sensitivity_through_observation_window_07_2026-09-04.json.

All interpretation remains confined to device-noise variability. This amendment does not create a coherent-dynamics, causal-calibration, hardware-performance, or quantum-advantage claim.