Phase 3 Multi-Device DLA Manifest¶

Date: 2026-05-06

This manifest records the approved second-backend DLA parity replication run on IBM Quantum hardware.

Preregistration¶

Protocol:

docs/campaigns/dla_multidevice_replication_prereg_2026-05-06.md

Execution script:

scripts/phase3_multidevice_dla_ibm.py

Backend and Budget Gate¶

Field	Value
Backend	`ibm_marrakesh`
Backend class	156-qubit Heron-class backend
Excluded backend	`ibm_kingston`
Live readiness timestamp	`2026-05-06T171231Z`
Main job	`ibm-run-63e0a1af74a38c9c`
Readout job	`ibm-run-0f96961442e05a77`
Main circuits	`144`
Readout circuits	`4`
Total circuits	`148`
Main shots	`4096`
Readout shots	`8192`
Estimated QPU minutes	`1.3566666666666667`
Preregistered ceiling	`10` minutes

Live readiness:

depth min/mean/max: 1 / 214.75 / 414
total-gate min/mean/max: 4 / 548.9121621621622 / 1086
ECR min/mean/max: 0 / 0.0 / 0

The max depth 414 passed the preregistered guard derived from the Phase 2 A+G live envelope. No extra QPU job was submitted outside this circuit matrix.

Artefacts¶

Artefact	SHA256
`data/phase3_multidevice_dla/phase3_multidevice_ibm_marrakesh_2026-05-06T171231Z.json`	`cf76b66c69fd6901bc2a9e2c527e978a1c3c1347761d6c16656831a7babfa36a`
`data/phase3_multidevice_dla/phase3_multidevice_summary_2026-05-06.json`	`751965194a1449767c623dcb119c2514be800efcea1a9652d8acf711cd5471eb`
`data/phase3_multidevice_dla/phase3_multidevice_row_metrics_2026-05-06.csv`	`15f5de37592947452e1b076e50d7f74b23854b3ca579eb489f0809db9b143eec`
`data/phase3_multidevice_dla/phase3_multidevice_readout_corrected_summary_2026-05-06.json`	`c43e25d56555f4e1a4593185efddbe51877f638b607c9fa84bcfc07c58ae916f`
`data/phase3_multidevice_dla/phase3_multidevice_readout_corrected_rows_2026-05-06.csv`	`707d9b2767e3d5a635a6919318119653b495b00ca3661247c41fcd3c3756bc33`

Per-Depth Summary¶

Parity leakage is reported as opposite-parity counts divided by total counts. Relative asymmetry is (mean_even - mean_odd) / mean_odd.

Depth	Even leakage	Odd leakage	Relative asymmetry	Interpretation
`4`	`0.082153`	`0.084290`	`-0.025344`	opposite sign
`6`	`0.105184`	`0.101440`	`+0.036903`	same sign, small
`8`	`0.123108`	`0.123271`	`-0.001320`	near zero
`10`	`0.133586`	`0.136983`	`-0.024803`	opposite sign
`14`	`0.163127`	`0.176229`	`-0.074348`	opposite sign
`20`	`0.198771`	`0.205892`	`-0.034585`	opposite sign

Full-Basis Readout-Corrected Summary¶

The matching ibm_marrakesh full-basis readout calibration is documented in docs/campaigns/readout_full_basis_manifest_2026-05-06.md. It uses physical qubits [5,6,7,8] and has assignment-matrix condition number 1.0756988057943142.

Depth	Corrected even leakage	Corrected odd leakage	Corrected relative asymmetry	Interpretation
`4`	`0.043936`	`0.064233`	`-0.315987`	opposite sign
`6`	`0.072900`	`0.080663`	`-0.096237`	opposite sign
`8`	`0.096515`	`0.100098`	`-0.035798`	opposite sign
`10`	`0.110723`	`0.110939`	`-0.001953`	near zero
`14`	`0.143522`	`0.150375`	`-0.045574`	opposite sign
`20`	`0.178952`	`0.188365`	`-0.049971`	opposite sign

The full-basis readout correction does not rescue a positive backend-transfer replication. It strengthens the conservative conclusion that the previously promoted ibm_kingston sign is backend/calibration/layout sensitive.

Claim Boundary¶

The result is best read as a multi-device boundary condition. It does not support a simple backend-stable replication of the ibm_kingston Phase 2 A+G positive leakage-asymmetry pattern.

Supported:

ibm_marrakesh raw-count replication attempt under the preregistered reduced n=4 circuit matrix;
mixed sign and mostly opposite-sign leakage asymmetry across the promoted depths;
evidence that the earlier ibm_kingston effect is backend/calibration/layout sensitive.
full-basis readout-corrected evidence on the matching ibm_marrakesh four-qubit layout showing non-positive relative asymmetry across all promoted depths.

Blocked:

DLA-parity-only causality;
backend-universal protection;
monotone scaling;
broad quantum advantage;
GUESS mitigation validation;
full readout-matrix mitigation from the four readout states.

Reproduction¶

Generate the readiness/submission artefact:

./.venv-linux/bin/python scripts/phase3_multidevice_dla_ibm.py \
  --backend ibm_marrakesh \
  --submit \
  --confirm-budget

The generated summary and row metrics are derived from:

data/phase3_multidevice_dla/phase3_multidevice_ibm_marrakesh_2026-05-06T171231Z.json

No further QPU work is required to recompute the rows above.