IQM DLA Backend-Sensitivity Preregistration (Properly Powered)

Date: 2026-07-21

This preregistration commits the design, decision rule, and budget for a properly powered continuation of the 2026-05-13 IQM Garnet DLA-parity replication campaign. It does not submit an IQM job and does not authorise credit spend; execution requires a separate explicit per-submit approval.

Context: IQM granted extended Resonance access on 2026-07-20 (ticket IQMCS-1691) for exactly this continuation, after we disclosed the honest mixed-sign 2026-05-13 outcome. Every publication arising from these runs credits IQM and IQM Resonance for hardware access.

Scientific Question (reframed after the exact classical baseline)

The 2026-07-21 exact statevector baseline (docs/campaigns/dla_parity_exact_baseline_2026-07-21.md) proved that the noiseless XY-Trotter parity leakage is exactly zero at every depth: the promoted ibm_kingston even/odd asymmetry (+10.8 % ± 1.1, Fisher p ≪ 10⁻¹⁶) is structure inside the device-noise floor, not coherent parity-selective dynamics. The live question is therefore:

Is parity-sector-correlated decoherence backend-sensitive? Does IQM Garnet exhibit a stable positive even-over-odd leakage asymmetry of the kind the IBM Heron devices show, or is that noise structure specific to the IBM stack?

The 2026-05-13 Garnet campaign (six pinned layouts, 12,288 shots total, combined Fisher p between 0.05 and 0.57, sign matches 1/3 to 3/3) was too low-powered to answer this. This block concentrates the statistics on one preregistered layout at a power level fixed in advance.

This is the preregistered-power continuation promised in our 2026-07-16 research-credits request to IQM, and a direct extension of the published backend-sensitivity observation (DOI 10.5281/zenodo.20382000).

Hypotheses and Primary Endpoint

• H1 (backend-universal noise structure): Garnet shows a positive pooled even-over-odd relative leakage asymmetry, in the direction of the IBM Phase 2 A+G reference.
• H0 (backend-sensitive noise structure): no stable positive asymmetry on Garnet. The 2026-05-13 data weakly favour H0.

Primary endpoint: pooled across the three depths, one-sided two-proportion test of leak_even > leak_odd at alpha = 0.05. Relative asymmetry metric is the committed (even - odd) / odd from scripts/analyse_iqm_dla_parity.py.

Power (committed before execution): with 12,288 shots per arm pooled and a baseline odd-arm leak near the 2026-05-13 level (~0.42), the minimum detectable relative asymmetry at 90 % power is +0.044 — the magnitude of the weakest IBM Phase 2 A+G per-depth reference effect (d6 +0.0435, d10 +0.0447). Per-depth secondary tests at 4,096 shots per arm resolve d4-like effects (+0.0865 reference against a +0.076 minimum detectable effect).

Layout Rule (selection bias disclosed)

Physical layout [2, 7, 12, 13] — the only 2026-05-13 layout with 3/3 depth-wise sign agreement with IBM. This layout is chosen using the May data, so the May result cannot be reused as evidence; this block is a confirmatory test on entirely new data with the decision rule above. If calibration makes [2, 7, 12, 13] unavailable, fall back to [9, 4, 3, 8] and record the substitution before submission.

Circuit Matrix

Field	Value
Device	IQM Garnet (20 qubits, square lattice) via Resonance
n	4
Circuit family	committed iqm_dla_pinned_n4_d{4,6,10}_{even,odd} builders
States	0011 even reference, 0001 odd reference
Depths	4, 6, 10
Repetitions	4 per state/depth (calibration-drift resolution)
Main shots	1024 per repetition → 4096 per state/depth
Readout states	0011, 0001, 0000, 1111
Readout shots	2048 per state

Circuit count: 2 × 3 × 4 = 24 main + 4 readout = 28 circuits. Shot count: 24,576 main + 8,192 readout = 32,768 shots.

Budget and Stop Rules

• Submit the first repetition block (6 main circuits + 4 readout circuits) alone, then read the actual credit burn from the Resonance dashboard before continuing.
• Abort the campaign if the first block consumes more than one third of the currently visible credit allowance.
• No further IQM submissions beyond this matrix without a fresh preregistration.

Live Readiness Gates

Before any submission:

• dry-run the full matrix against IQMFakeGarnet from committed code only;
• confirm the pinned layout qubits are calibrated and operational on the Resonance dashboard;
• record device name, timestamp, calibration snapshot metadata, transpiled depth and two-qubit gate summaries, shot and circuit counts;
• reject any circuit whose transpiled depth exceeds the 2026-05-13 envelope (d10 transpiled depth 159) by more than 25 %;
• obtain explicit owner approval immediately before submission (per-submit GO).

Analysis Plan

• Primary: pooled one-sided two-proportion test as committed above.
• Secondary: per-depth Wilson intervals, per-depth one-sided tests, per-repetition drift table (the 2026-05-13 partial repeat showed calibration drift matters), sign agreement with the IBM Phase 2 A+G reference, descriptive combined Fisher statistic with dependence caveat.
• Readout handling: four-state exact-state parity-confusion correction; no full confusion-matrix mitigation claim from a four-state calibration.
• All analyses run by committed scripts extending scripts/analyse_iqm_dla_parity.py; raw counts, job identifiers, and SHA-256 hashes published in full.

Decision Rule (frozen)

• Primary test rejects H0 (p < 0.05, positive direction): report Garnet parity-sector noise asymmetry as replicating the IBM direction at preregistered power; publish as backend-shared noise structure.
• Primary test does not reject: report a bounded null — Garnet shows no IBM-magnitude positive asymmetry at 90 % power — and publish it as backend-sensitivity evidence with the same prominence a positive result would receive.
• Either way, no coherent-dynamics claim is available: the exact baseline fixes noiseless leakage at zero, so every observed asymmetry statement is a statement about device noise.

Claim Boundary

Blocked regardless of outcome: quantum advantage; DLA-parity-only causality; monotone scaling; backend-universal protection; any modification of the frozen submissions under paper/submissions/ (results feed a new manuscript only); any claim beyond the sampled device, calibration window, circuit family, layout, and state pair.

Output Artefacts

• data/iqm_backend_sensitivity/iqm_dla_powered_<timestamp>_{plan,executed}.json;
• data/iqm_backend_sensitivity/iqm_dla_powered_analysis_<date>.{json,md};
• raw job identifiers public; access token never committed;
• campaign manifest docs/campaigns/iqm_dla_backend_sensitivity_manifest_<date>.md.

Submission Boundary

This preregistration is complete once committed. QPU execution remains blocked until the live readiness gates pass and the owner grants a separate explicit per-submit GO.