IQM Garnet Layout-Transfer Per-Size Powered Preregistration (FU-3)

Date: 2026-07-22

This preregistration commits the design, decision rules, and budget for the per-size powered follow-up to the executed layout-transfer campaign (iqm_layout_transfer_square_lattice_prereg_2026-07-21.md, executed 2026-07-21 with Amendment 1). It is committed and pushed BEFORE any of its own hardware data exist. Execution requires a separate explicit owner GO.

Motivating (already-collected) evidence

The 2026-07-21 layout-transfer campaign FAILED its frozen all-three-sizes primary (wins_all_sizes_optimised_vs_default = false): the pooled default-minus-optimised corrected-error difference was positive (+0.0345, bootstrap CI90 [+0.026, +0.043] — the optimiser helps on average) but the per-size profile was non-uniform: n = 8 (0.147 vs 0.253) and n = 16 (0.292 vs 0.314) favoured the optimiser while n = 12 REVERSED (optimised error 0.209 vs default 0.185) at 2,048 shots per arm. The default arm beat the naive arm at every size (+0.0989 pooled). Two open questions:

Q1: Is the n = 12 reversal a real per-size disadvantage of the calibration-aware optimiser against the default transpiler placement, or shot noise at the executed per-arm budget? Q2: Is the optimised-vs-default effect profile heterogeneous across sizes beyond shot noise?

The 2026-07-21 data are design input here and are never re-used as confirmatory evidence for this campaign's endpoints.

Circuit Matrix

Field	Value
Device	IQM Garnet (20 qubits, square lattice) via Resonance
Circuit family	committed iqm_layout_transfer_benchmark builders (TROTTER_DEPTH 5, transpiler seed 20260721, IQM basis r/cz) — identical to the executed campaign
Sizes	8, 12, 16
Arms	optimised (best_chain_region → optimise_kuramoto_layout), default (transpiler placement), naive (Amendment 1: lexicographically smallest connected chain) — all three recomputed on the submission day's calibration snapshot and recorded BEFORE submission
Repetitions	4 per arm/size (execution-order replicates)
Main shots	2048 per repetition → 8192 per arm/size (4× the executed campaign)
Readout	per size: all-zeros / all-ones over the union of the arms' measured qubits at 1024 shots (exact tensored per-qubit correction, unchanged)
Observable	corrected order-parameter error E = |R̂_corrected − 0.5| per arm/size, fail-closed denominator — identical to the executed campaign

Circuit count: 3 × 3 × 4 = 36 main + 3 × 2 = 6 readout = 42 circuits. Shot count: 73,728 main + 6,144 readout = 79,872 shots.

Batching disclosure (frozen): the whole matrix submits in one pass — mains batched into one job (or the minimum number the batch limit allows) and readout into one more. All arms of every size ride in the SAME batch, so window-level drift (the FU-1 finding, quantified by FU-W) cancels in the per-size ARM DIFFERENCES; the campaign makes no cross-window claim.

Endpoints and Decision Rules (frozen)

Per size n, pool the 4 repetitions per arm into the corrected errors E_opt(n), E_def(n), E_naive(n) and form the primary difference D(n) = E_def(n) − E_opt(n) (positive = optimiser advantage, the executed campaign's sign convention). All intervals and tests use the committed bootstrap machinery (10,000 resamples, seed 20260722).

• Primary (per-size resolution): two-sided bootstrap test of D(n) ≠ 0 at every size, Holm–Bonferroni adjusted across the three sizes, α = 0.05. Power: the executed pooled bootstrap CI90 gives a per-size SE ≈ 0.0090 at 2,048 shots per arm, hence ≈ 0.0045 at 8,192; the minimum detectable per-size |D| is ≈ 0.017 at 90 % power under the worst Holm threshold. The executed n = 12 reversal (−0.024) and the executed n = 8 advantage (+0.106) both exceed this.
• Secondary S1 (n = 12 anomaly): sign and bootstrap CI95 of D(12), reported regardless of the primary outcome — the direct resolution of the executed reversal.
• Secondary S2 (pooled replication): pooled D across sizes with bootstrap CI90, compared descriptively against the executed +0.0345 [+0.026, +0.043].
• Secondary S3 (heterogeneity): inverse-variance Cochran's Q across the three per-size D(n) (bootstrap variances), α = 0.05 — the formal test of Q2.
• Secondary S4 (naive reference): per-size E_def(n) − E_naive(n) with bootstrap CI95, descriptively against the executed uniform default-over-naive result.
• Secondary S5 (correction sensitivity): raw vs corrected D(n) side by side (the correction is exact for the observable; divergence flags a readout-model violation and is reported, never silently fixed).
• Depth-parity gate (multiplicative, max ≤ min·(1 + tolerance) per size across arms, the committed depth_parity_gate) enforced at dry run AND at submit exactly as executed; a violation blocks submission (no post-hoc arm handicap).

Every outcome branch is publishable: a Holm-surviving negative D(12) establishes a real per-size optimiser disadvantage; a positive or null D(12) bounds the executed reversal as shot noise; S3 settles whether layout transfer is uniform or size-modulated.

Live Readiness Gates (block submission until all pass)

• IQMFakeGarnet dry run of all 42 circuits from committed code only, with all arms' layouts computed from the day's calibration snapshot and the per-size depth-parity gate green;
• garnet:mock zero-spend integration submit;
• calibration check on the day's live snapshot for all arms' chains (every chain edge calibrated), layouts recorded before submission;
• explicit owner GO immediately before submission.

Budget and Stop Rules

• Grant context: 500 IQM Resonance credits (expiry 2026-11-22).
• Expected cost: ~2–3 jobs ≈ 2–3 credits (single-pass batching).
• Abort if the submission consumes more than 6 credits on the owner's dashboard reading.
• No submissions beyond this 42-circuit matrix without a fresh preregistration; a second-window replication (if ever wanted) is a NEW preregistration, not an extension.

Claim Boundary

Blocked regardless of outcome: quantum advantage; coherence protection; claims about non-sampled sizes, devices, or calibration windows; 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.