IQM Garnet FU-3 Per-Size Layout-Transfer Result¶
Execution date: 2026-07-26
This is the result record for the design and decision rules frozen in
iqm_layout_transfer_per_size_prereg_2026-07-22.md. The run used calibration
set c2097be4-1e23-49bc-adaa-8e8c01df6223 and the committed submission
package at 692c915a6ecb272f2083d4007c5d9ec350dc456e.
Execution evidence¶
- IQM Garnet mains job:
019f9c79-1d4b-79e3-a590-aa180819d930(36 circuits at 2,048 shots). - IQM Garnet readout job:
019f9c79-21e0-7eb2-86e1-69c47fd82715(6 circuits at 1,024 shots). - Retrieved matrix: 42/42 labels and 79,872/79,872 shots.
- Provider-transpiled two-qubit depth: 40 for every one of the 36 main circuits; the frozen per-size depth-parity ratio is therefore 1.0.
- Analysis: frozen seed
20260722, 10,000 bootstrap resamples.
The zero-spend garnet:mock provider integration completed the same two-job
split and retrieval before the hardware call. Its synthetic readout counts
produced a non-positive correction denominator, so mock analysis failed
closed as intended; mock counts are not scientific evidence. QPY generated
under Qiskit 2.4.1 emitted a compatibility warning in the isolated Qiskit
2.1.2 environment, but loading, mock execution, live submission, and live
retrieval all completed successfully.
Frozen primary endpoint¶
The corrected-error difference is default − optimised, so a positive value
favours the calibration-aware optimiser and a negative value favours the
default placement.
| Size | Difference | Bootstrap CI95 | Holm-adjusted p | Direction |
|---|---|---|---|---|
| 8 | +0.01868 | [+0.00755, +0.02968] | 0.001200 | optimiser advantage |
| 12 | −0.08728 | [−0.10090, −0.07459] | 0.000600 | optimiser disadvantage |
| 16 | +0.04964 | [+0.04109, +0.05823] | 0.000600 | optimiser advantage |
All three per-size differences reject zero after the frozen Holm correction. This does not mean that the optimiser wins at all sizes: the n=12 effect is significant in the opposite direction and reproduces the motivating n=12 reversal at the powered shot budget.
Secondary endpoints and sensitivity¶
- Pooled default-minus-optimised difference: −0.00632, bootstrap CI90 [−0.01187, −0.00099]. The pooled sign is dominated by the n=12 disadvantage and must not replace the per-size result.
- Cross-size heterogeneity: Cochran Q=292.430 on 2 df,
p=
3.16e-64; the homogeneous-effect model is rejected. - The raw-count sensitivity analysis preserves every primary direction: n=8 +0.01642, n=12 −0.02260, n=16 +0.04378, with all three CI95 intervals excluding zero.
- Default-minus-naive corrected differences are negative at every size (n=8 −0.25877, n=12 −0.20541, n=16 −0.02934), so default placement has lower sampled error than the preregistered naive chain in this window.
Interpretation boundary¶
FU-3 resolves a strongly heterogeneous, size-dependent layout effect on one IQM Garnet calibration window. It supports neither a quantum-advantage claim nor a cross-device generalisation. The result specifically rejects a simple "calibration-aware placement always helps" account: it helps at n=8 and n=16 but is materially worse than default placement at n=12 under the frozen observable and correction model.
Canonical artefacts are under
data/iqm_layout_transfer_per_size/live_2026-07-26/:
live_submission_2026-07-26.json;live_counts_2026-07-26.json;live_analysis_2026-07-26.json.