DLA Multi-Device Replication Preregistration¶
Date: 2026-05-06
This preregistration prepares a minimal second-backend DLA parity replication. It does not submit an IBM job, reserve backend time, or authorise QPU spend.
Scientific Question¶
Does the promoted n=4 parity-sector/excitation-number correlated leakage
asymmetry replicate on a second Heron-class backend, or is it specific to the
ibm_kingston device/calibration window?
Claim Boundary¶
Supported after successful execution and analysis:
- backend-transfer evidence for the same small
n=4observable; - comparison of sign, magnitude, and per-depth stability versus the promoted
ibm_kingstonPhase 2 A+G result; - no claim beyond the sampled backend, calibration window, circuit family, and initial-state pair.
Blocked even after a positive result:
- broad quantum advantage;
- DLA-parity-only causality;
- monotone scaling;
- backend-universal protection;
- GUESS mitigation validation;
- full readout-matrix mitigation unless a complete basis calibration is also submitted.
Target Backend Rule¶
Use a second Heron-class backend only. Do not rerun on ibm_kingston for this
specific replication item. The actual backend name must be selected from the
live IBM account-visible backend list immediately before readiness checks.
Circuit Matrix¶
| Field | Value |
|---|---|
n |
4 |
| Coupling model | same heterogeneous Kuramoto-XY matrix as Phase 2 A+G |
| States | 0011 even reference, 0001 odd reference |
| Depths | 4, 6, 8, 10, 14, 20 |
| Repetitions | 12 per state/depth |
| Main shots | 4096 per circuit |
| Readout states | 0011, 0001, 0000, 1111 |
| Readout shots | 8192 per circuit |
Circuit count:
- main circuits:
2 states x 6 depths x 12 reps = 144; - readout circuits:
4; - total circuits:
148.
QPU-Time Estimate¶
Expected IBM-reported QPU time: 3-6 minutes if live-transpiled depths remain
comparable to Phase 2 A+G.
Budget ceiling for this preregistered block: 10 IBM-reported QPU minutes.
Abort before submission if the live estimate exceeds the ceiling or if the remaining allocation cannot cover the block plus a 25 % safety margin.
Live Readiness Gates¶
Before submission:
- confirm backend is Heron-class and not
ibm_kingston; - confirm backend is account-visible and operational in the IBM dashboard;
- generate circuits from committed code only;
- live-transpile every circuit on the selected backend;
- reject if max transpiled depth exceeds the Phase 2 A+G envelope by more than 25 %;
- reject if max two-qubit gate count exceeds the Phase 2 A+G envelope by more than 25 %;
- record backend name, timestamp, calibration snapshot metadata, depth summary, two-qubit gate summary, shot count, circuit count, and estimated QPU minutes;
- get explicit approval immediately before submission.
Analysis Plan¶
Primary observable:
- parity leakage per state/depth/repetition.
Primary comparison:
- odd-reference leakage minus even-reference leakage at matched depth.
Promoted summaries:
- per-depth mean asymmetry;
- per-depth Welch/permutation checks where sample size permits;
- Fisher-style combined statistic only as descriptive and with dependence caveat;
- sign agreement with Phase 2 A+G;
- effect-size comparison against Phase 2 A+G.
Readout handling:
- use the four readout states for exact-state parity-confusion correction;
- do not claim full
2^n x 2^nconfusion-matrix mitigation from this four-state calibration.
Falsification Rules¶
The backend-transfer claim is weakened or rejected if:
- the sign reverses for most promoted depths;
- the effect collapses inside readout-only uncertainty;
- the circuit-depth envelope is materially worse than Phase 2 A+G;
- temporal drift or backend calibration metadata makes the comparison uninterpretable.
If the result is null or opposite-sign, publish it as a multi-device boundary, not as a failed experiment.
Output Artefacts¶
Expected paths after approved execution:
data/phase3_multidevice_dla/phase3_multidevice_<backend>_<timestamp>.json;data/phase3_multidevice_dla/phase3_multidevice_summary_<date>.json;data/phase3_multidevice_dla/phase3_multidevice_row_metrics_<date>.csv;docs/campaigns/phase3_multidevice_dla_manifest_<date>.md.
Each artefact must include job ID, backend, circuit metadata, raw counts, SHA256 hashes, depth/gate summaries, and reproduction commands.
Submission Boundary¶
This preregistration is complete. QPU execution remains blocked until backend selection, live readiness artefacts, budget confirmation, and explicit approval are completed in a separate task.