REPRODUCIBILITY RECORD — SCPN Quantum Control Frontier Campaign 2026¶

Document version: 2.0 Date: 2026-04-26 Author: Miroslav Šotek / SCPN Quantum Control project SPDX: AGPL-3.0-or-later

Purpose¶

This document records the full software and hardware stack used for the Frontier Campaign reproducibility lane. It is a reproducibility checkpoint for restarts of campaign scripts and bounded no-QPU validation output.

Use this record to verify:

package versions used during campaign execution,
environment-level fixes needed for mitiq on Ubuntu Python packaging,
hardware aliasing and backend assumptions in committed runs,
parameter provenance constraints before running synthetic or production experiments.

Repository Access State¶

Repository: https://github.com/anulum/scpn-quantum-control

The repository is public. Reproducibility runners should use the public GitHub repository, tagged release archives, or public PyPI artefacts as their source entry point. Commercial licence grants do not replace public source access for the AGPL-covered repository; they provide a separate proprietary-integration route for users who cannot accept AGPL obligations.

Audience and limits¶

The record is for maintainers, auditors, and reproducibility runners. It is not an external scientific validation statement by itself; it is the first step in an evidence chain for bounded campaign claims.

1. Software Environment¶

Component	Version
Python	3.12.3
Qiskit	2.4.0
qiskit-ibm-runtime	0.46.1
NumPy	2.2.6 (pip)
mitiq	1.0.0
cirq-core	1.6.1
matplotlib	3.10.9 (pip)
scpn-quantum-control	Source tree v1.0.0; public release artefacts may lag until the next tagged package release

Install the exact environment:

pip install qiskit==2.4.0 qiskit-ibm-runtime==0.46.1 numpy==2.2.6 mitiq==1.0.0
pip install -e .    # installs scpn-quantum-control from this repo

mpl_toolkits / mitiq system fix (Linux only)¶

On Ubuntu 24.x, the system python3-matplotlib apt package installs an incompatible mpl_toolkits at /usr/lib/python3/dist-packages/mpl_toolkits/ that is incompatible with pip matplotlib ≥ 3.8. This causes mitiq (which depends on cirq) to fail to import with ModuleNotFoundError: No module named 'matplotlib.tri.triangulation'.

Fix applied 2026-04-26:

# 1. Force-reinstall pip matplotlib to get compatible mpl_toolkits
pip install --upgrade matplotlib --force-reinstall --break-system-packages

# 2. Copy pip-installed mplot3d into system location (overwrite stale version)
sudo mkdir -p /usr/lib/python3/dist-packages/mpl_toolkits/mplot3d
sudo cp -a ~/.local/lib/python3.12/site-packages/mpl_toolkits/mplot3d/. \
           /usr/lib/python3/dist-packages/mpl_toolkits/mplot3d/

# 3. Clear pyc caches
sudo find /usr/lib/python3/dist-packages/mpl_toolkits -name "*.pyc" -delete
sudo find /usr/lib/python3/dist-packages/mpl_toolkits -name "__pycache__" -exec rm -rf {} +

After this fix: import mitiq; from mitiq import zne imports cleanly.

2. Hardware¶

Field	Value
Provider	IBM Quantum (ibm_cloud channel)
Backend alias	`ibm_heron_r2` → resolved to `ibm_fez`
Architecture	IBM Heron r2
Qubit count	156
Native gate set	CZ, RZ, SX, X
Shots per job	min(requested, 4000)

3. Circuit Construction — `StructuredAnsatz.from_kuramoto`¶

File: src/scpn_quantum_control/control/structured_ansatz.py

The Trotterised Kuramoto-XY circuit is built for an N-qubit system as follows:

Initial state: Hadamard on all qubits → uniform phase superposition
Scale coupling: K_scaled = K_nm × coupling_scale (applied once before the loop)
Trotter loop (repeated trotter_depth times, dt = time_step):
Frequency term: RZ(2·ωᵢ·dt, i) for each qubit i
XY coupling: RZZ(2·K_scaled[i,j]·dt, i, j) for all pairs with |K_scaled[i,j]| > 1e-8
FIM feedback (optional): RZ(λ_fim·dt, i) — λ_fim is a concrete float (not a Qiskit Parameter) to prevent name-collision in Qiskit ≥ 2.x

Parameters¶

Parameter	Batches 1–4	Batch 5+
`trotter_depth`	8	8
`time_step`	0.1	0.1
`lambda_fim`	0.0	0.0
`coupling_scale`	1.0 (implicit, no parameter existed)	2.0 (default)

Note on coupling_scale: Added 2026-04-26. Default 2.0 doubles K_nm before circuit construction, pushing the system toward the Kuramoto synchronisation threshold. Stored in ansatz.params for reproducibility. All batches 1–4 used coupling_scale=1.0 (parameter did not exist; K_nm was used unscaled).

4. Parameter Files¶

Frontier .npy files are a local transport cache, not publication source material. They must be generated from the bridge/provenance contract in scripts/frontier_campaign_2026/generate_params.py.

Default behaviour is fail-closed:

python scripts/frontier_campaign_2026/generate_params.py

The command raises if a bridge is unavailable or does not provide omega. It does not silently invent matrices.

Deterministic smoke-test arrays require explicit opt-in:

python scripts/frontier_campaign_2026/generate_params.py \
  --allow-synthetic \
  --seed 42

Synthetic runs write PARAMETER_PROVENANCE.json with source_mode="synthetic" entries under scripts/frontier_campaign_2026/params/. These files are suitable for interface tests only and are not publication-safe QPU inputs. Use --output-dir <path> only when intentionally writing a separate cache.

Legacy primary, hardware, and sophisticated campaign generators are also fail-closed. They do not create random matrices unless --allow-synthetic is passed, and the resulting PARAMETER_PROVENANCE.json marks every file as source_mode="synthetic". Their shell launchers require an existing provenance file before any hardware script runs; source-backed parameter caches should be produced by the bridge/orchestration layer, not invented by Quantum Control.

5. Error Mitigation — All Batches¶

Batch 1 — 2026-04-26 (T4: 80 jobs, T7: 9 jobs, T1: 1 orphan)¶

Setting	Value
`optimization_level`	1
Dynamical decoupling	None
ZNE	None
`coupling_scale`	1.0 (not yet added)

Batch 2 — 2026-04-26 (T1: 3 jobs, T4: 80 jobs, T7: 9 jobs)¶

Setting	Value
`optimization_level`	3, `seed_transpiler=42`
Dynamical decoupling	Skipped — DD `PassManager` not yet in code
ZNE	None
`coupling_scale`	1.0 (not yet added)

Batch 3 — 2026-04-26 (T1: 3 jobs, T4: 80 jobs, T7: 9 jobs)¶

Setting	Value
`optimization_level`	3, `seed_transpiler=42`
Dynamical decoupling	Skipped — `ALAPScheduleAnalysis` import missing from `_run_blocking` scope
ZNE	None
`coupling_scale`	1.0 (not yet added)

Batch 4 — 2026-04-26 (T1: 3 jobs, T4: 80 jobs, T7: 9 jobs = 92 total)¶

Setting	Value
`optimization_level`	3, `seed_transpiler=42`
Dynamical decoupling	X–X sequence API verified locally — `PassManager([ALAPScheduleAnalysis(durations, target), PadDynamicalDecoupling(durations, [XGate(),XGate()], target)]).run(isa_qc)` — circuit depth 1922 → 1988 in the retained ibm_fez baseline artifact row
ZNE	None (mitiq broken at time of run)
`coupling_scale`	1.0 (not yet added)

DD API note (Qiskit 2.4.0): PadDynamicalDecoupling and ALAPScheduleAnalysis must be chained via PassManager.run() — calling them directly on a circuit raises AttributeError. spacing must be list[float] not a string. Neither pass accepts backend as a positional arg. Three iterations of debugging were required to establish the correct pattern.

Batch 5 — 2026-04-26 (T1: 3 jobs, T4: 80 jobs, T7: 9 jobs = 92 total) ← CURRENT¶

Setting	Value
`optimization_level`	3, `seed_transpiler=42`
Dynamical decoupling	✅ X–X sequence (same verified PassManager pattern as batch 4)
ZNE	✅ Richardson extrapolation, scale factors [1, 2, 3], `fold_global`
`coupling_scale`	✅ 2.0 (default; K_nm doubled before circuit construction)

ZNE implementation (mitiq 1.0.0, verified API):

from mitiq import zne
from mitiq.zne.scaling import fold_global
from mitiq.zne.inference import RichardsonFactory

def _zne_executor(circ):
    counts = sampler.run([circ]).result()[0].data.meas.get_counts()
    return SyncOrderParameter()(counts=counts)["sync_order_z_magnetisation"]  # scalar float

zne_sync_order = zne.execute_with_zne(
    isa_qc,
    _zne_executor,
    factory=RichardsonFactory([1, 2, 3]),
    scale_noise=fold_global,
)

ZNE runs 3 circuits (scale=1,2,3), extrapolates the Z-basis magnetisation proxy to the zero-noise limit. A final scale=1 run collects full counts for DLAParityWitness, IntegratedInformationPhi. Result JSON includes zne_applied=True, zne_scale_factors=[1,2,3], zne_factory="RichardsonFactory".

6. Cross-Batch Results — T4 `SyncOrderParameter` (80 jobs each)¶

Batch	`coupling_scale`	Mitigation	sync mean	sync std	Δ vs B1
1	1.0	opt_level=1, no DD, no ZNE	0.1409	0.0241	baseline
2	1.0	opt_level=3, no DD, no ZNE	0.0867	0.0071	−38%
3	1.0	opt_level=3, no DD, no ZNE	0.0867	0.0071	−38% (reproducible)
4	1.0	opt_level=3 + DD X–X, no ZNE	0.0066	0.0037	−95%
5	2.0	opt_level=3 + DD X–X + ZNE	pending	pending	—

Key finding batch 4: DD reduces spurious sync_order by 95% relative to batch 1. Without DD, idle qubit dephasing and crosstalk create false correlations that inflate the magnetisation signal. With DD, sync_order approaches the physical noise floor (~0.007), consistent with a disordered Kuramoto system below the synchronisation threshold at these coupling strengths.

Batch 5 prediction: coupling_scale=2.0 doubles all RZZ angles; combined with ZNE Richardson extrapolation, sync_order should increase above batch 4 if the system is genuinely pushed toward the phase-coherent regime (Kuramoto critical point ~K_c ≈ 2σ_ω/π).

7. Observables — Scientific Classification¶

Class	Computes from real counts	Description
`SyncOrderParameter`	✅ YES	Z-basis magnetisation proxy from bitstring marginals; emits legacy `sync_order`, explicit `sync_order_z_magnetisation`, and `is_xy_kuramoto_order_parameter = 0.0`
`DLAParityWitness`	✅ YES	Odd/even Hamming-weight parity asymmetry
`IntegratedInformationPhi`	❌ NO	No production IIT/causal-state implementation is wired; entropy is available only as an explicitly labelled diagnostic
`QuantumFisherInformation`	✅ YES, when Hamiltonian inputs are supplied	Routes explicit coupling matrix and natural frequencies through the spectral QFI engine; the sync-order/DLA estimate is available only as an explicitly labelled diagnostic proxy
`ThermodynamicWitness`	✅ YES, when work protocol data are supplied	Requires explicit work samples or a calibrated work value in joules; refuses default/synthetic work
`LogicalSyncWitness`	✅ YES, when DLA-protected counts/probabilities are supplied	Routes counts or probabilities through the DLA-protected QEC witness; scalar fidelity is available only as an explicitly labelled finite unit-interval diagnostic proxy

Publication guidance: Only ✅ observables are attributable to real QPU measurements. ⚠️ PROXY observables must be labelled as model estimates and must not be reported under production observable keys.

8. Reproducing Results¶

# 1. Set credentials
export SCPN_IBM_TOKEN="<token>"
export SCPN_IBM_CRN="<instance-crn>"  # optional when the account has a default instance

# 2. Generate source-backed parameter files
python3 scripts/frontier_campaign_2026/generate_params.py

# Optional smoke-test cache only; not publication-safe
python3 scripts/frontier_campaign_2026/generate_params.py \
  --allow-synthetic \
  --seed 42

# 3. Run campaign (batch 5 settings: coupling_scale=2.0, DD, ZNE)
python3 scripts/frontier_campaign_2026/run_credible_tests.py

# 4. Retrieve results
python3 scripts/retrieve_all_jobs.py

Independent job verification:

import os
from qiskit_ibm_runtime import QiskitRuntimeService
service_kwargs = {"channel": "ibm_cloud", "token": os.environ["SCPN_IBM_TOKEN"]}
if os.environ.get("SCPN_IBM_CRN"):
    service_kwargs["instance"] = os.environ["SCPN_IBM_CRN"]
service = QiskitRuntimeService(**service_kwargs)
job = service.job("<job_id_from_result_json>")
print(job.result())

9. Known Limitations¶

Shot cap: min(shots, 4000) — higher shot requests are silently capped.
N=160 skipped: IBM Heron r2 has 156 qubits; T1 N=160 point is always skipped.
Integrated information: IntegratedInformationPhi does not report Φ from output counts. Near-uniform entropy over 12–20 qubits at 4000 shots is available only as a labelled entropy diagnostic and is not a meaningful IIT measurement.
T2 hardware run not executed: scpn_neurocore.bridge.load_live_stream now exposes a replayable artifact contract for the live-loop script, but the historical T2 IBM hardware run remains unexecuted in this record.
ZNE job_id: When ZNE succeeds, job_id is set to "zne_mitigated" — individual scale-factor job IDs are managed internally by mitiq and not exposed in result JSON.
ZNE cost: Each ZNE run submits 3 circuits (scale=1,2,3) plus 1 final run = 4× IBM job cost vs unmitigated. T4 ZNE = 80 × 4 = 320 IBM jobs.
Batch 1 baseline: Run without DD or ZNE; inflated sync_order (0.14) is dominated by noise artefacts, not physical synchronisation.
Retired injectors: Legacy local campaign injector modules now fail at import time. Hardware campaigns must use AsyncHardwareRunner and source-backed or explicitly labelled smoke-test artifacts.