IQM Garnet Parity-Asymmetry Window-Variability Preregistration (FU-W)

Date: 2026-07-22

This preregistration commits the design, decision rules, and budget for the window-variability follow-up to the executed d10 sign-replication campaign (iqm_dla_d10_retest_prereg_2026-07-22.md). It is committed and pushed BEFORE any of its own hardware data exist (per the standing rule that the freeze must be git-provable). Execution requires a separate explicit owner GO per window submission.

Motivating (already-collected) evidence

The 2026-07-22 d10 sign-replication campaign found that the negative depth-10 asymmetry observed on 2026-07-21 does NOT replicate in a fresh execution window: the even-minus-odd leakage difference moved from −0.0237 (2026-07-21 powered block) to +0.0133 (2026-07-22 retest), a cross-window shift of 0.0370 with z = 2.74, two-sided p = 0.006, on the same pinned layout and the same calibration set. The depth-10 anomaly is therefore attributed to window-to-window drift, not to a stable depth-dependent sign change.

That finding converts drift from a confound into the object of study:

Q1: Does the parity-asymmetry difference drift between execution windows by more than binomial shot noise, at each depth? Q2: How large is the between-window standard deviation τ per depth, and is the shallow-depth (d4) strong positive asymmetry stable in sign across windows while deeper values fluctuate?

The 2026-07-21/22 data are design input here and are never re-used as confirmatory evidence for this campaign's endpoints; every window analysed below is newly collected.

Circuit Matrix (per window)

Field	Value
Device	IQM Garnet (20 qubits, square lattice) via Resonance
n	4
Circuit family	committed iqm_dla_pinned_n4_d{4,8,10,12}_{even,odd} builders (same code path as all prior campaigns)
States	0011 even, 0001 odd
Depths	4, 8, 10, 12
Repetitions	4 per state/depth (execution-order replicates inside the batch)
Main shots	1024 per repetition → 4096 per state/depth per window
Readout states	0011, 0001, 0000, 1111 at 2048 shots, run EVERY window (fresh per-window readout snapshot is part of the design)
Layout	[2, 7, 12, 13] pinned (identical to all 2026-07 campaigns); fallback [9, 4, 3, 8] recorded before submission if unavailable

Per window: 4 × 2 × 4 = 32 main + 4 readout = 36 circuits, 32,768 main + 8,192 readout = 40,960 shots.

Batching disclosure (frozen): each window submits in one pass — all 32 mains batch into one job and the 4 readout states into a second job (batch_all, the d10-retest pattern). Within a window the repetitions are execution-order replicates, NOT independent windows; the independent unit of this campaign is the WINDOW. Expected cost ≈ 2 jobs ≈ 2 credits per window at the observed ~1 credit/job billing.

Windows (frozen definition)

• A window is one full 36-circuit pass submitted after an explicit owner GO for that window.
• Consecutive windows are separated by ≥ 12 h wall clock.
• Target W = 10 windows; the campaign is analysable and publishable from W = 6 onward (the test degrees of freedom adapt to the achieved W; fewer than 6 windows → report as incomplete, no confirmatory claim).
• Windows MAY span calibration-set changes; the active calibration set and a live calibration snapshot (dump-calibration) are recorded per window and used as descriptive covariates only.

Endpoints and Decision Rules (frozen)

Per window w and depth d, the analysis pools the 4 repetitions into leak_even(d, w) and leak_odd(d, w) (4,096 shots per arm) and forms the difference Δ(d, w) = leak_even − leak_odd with binomial variance se²(d, w) = p_e(1−p_e)/4096 + p_o(1−p_o)/4096.

• Primary (depth-10 heterogeneity): Cochran's Q homogeneity test of Δ(10, w) across the executed windows against the shot-noise-only null (Q = Σ_w (Δ_w − Δ̄)² / se²_w with the inverse-variance weighted mean, Q ~ χ²(W−1) under H0), α = 0.05. Rejection ⇒ between-window drift at depth 10 exceeds binomial shot noise. Failure to reject ⇒ bounded null: at W = 10 and the anticipated se ≈ 0.011 per window arm pair, the design detects a between-window standard deviation τ ≥ 0.019 with ≈ 90 % power (scaled-χ² approximation); the 2026-07-21→22 shift (0.0370 between two windows, implying τ ≈ 0.026 under a two-draw model) is detected with > 95 % power.
• Secondary S1 (per-depth heterogeneity): the same Q test at depths 4, 8, and 12, Holm–Bonferroni adjusted across the three secondary depths.
• Secondary S2 (drift magnitude profile): DerSimonian–Laird method-of-moments point estimate τ̂(d) per depth, reported descriptively (no CI claim) together with the shot-noise se so the ratio is auditable.
• Secondary S3 (shallow-anchor sign stability): fraction of windows with Δ(4, w) > 0, with an exact (Clopper–Pearson) 95 % interval. The 2026-07-21 d4 asymmetry (+0.230 relative) predicts near-universal positivity; instability here would undercut the pooled-direction claim of the powered block and MUST be reported prominently if observed.
• Secondary S4 (calibration covariates, descriptive only): per-window scatter of Δ(10, w) against the recorded per-window CZ fidelity and readout error on the pinned layout; no test, no claim beyond description.
• Per-window drift table (per-repetition rows inside each window) as in all prior campaigns.
• Readout handling: four-state exact-state parity-confusion correction per window using THAT window's readout block, identical method to the powered block; raw-count endpoints remain primary exactly as before.

Either primary outcome is publishable: heterogeneity confirmed quantifies drift as a real noise floor for single-window depth-profile claims (and retroactively explains the d10 anomaly); the bounded null shows single-window profiles are shot-noise limited at the committed power. No coherent-dynamics claim is available in any branch (exact statevector baseline pins noiseless leakage at zero for every depth; committed module analysis/dla_parity_exact_baseline.py).

Live Readiness Gates (block each submission until all pass)

• IQMFakeGarnet dry run of the full 36-circuit window matrix from committed code only (once, before window 1);
• transpiled depth envelope: the executed/interpolated ladder d4→69, d8→129, d10→159, d12→189 with the standing +25 % rule, i.e. d4 ≤ 86, d8 ≤ 161, d10 ≤ 198, d12 ≤ 236, enforced at every submit;
• garnet:mock zero-spend integration submit (once, before window 1);
• pinned-layout calibration check on the day's live snapshot before EVERY window (all three chain edges calibrated), fallback substitution recorded before submission if required;
• explicit owner GO immediately before EVERY window submission.

Budget and Stop Rules

• Grant context: 500 IQM Resonance credits granted 2026-07-22, expiry 2026-11-22 (owner-verified from the dashboard; balances are never available in the client API).
• Expected total: ~2 credits × 10 windows ≈ 20 credits.
• Submit window 1 alone first; abort the campaign if window 1 consumes more than 5 credits on the owner's dashboard reading.
• Hard cap: abort if cumulative campaign spend exceeds 35 credits.
• No submissions beyond the preregistered per-window 36-circuit matrix and the W = 10 target without a fresh preregistration.

Claim Boundary

Blocked regardless of outcome: quantum advantage; coherence protection; coherent-dynamics parity claims; any causal attribution of drift to a specific hardware mechanism (S4 is descriptive); extrapolation beyond the sampled device, layouts, and depths; any modification of frozen submissions under paper/submissions/ (results extend the NEW manuscript submission_007 or its successor only). IQM and IQM Resonance are credited in every resulting output.

Submission Boundary

This preregistration is complete once committed and pushed. QPU execution stays blocked until the readiness gates pass and the owner grants a separate explicit GO for each window.