Madmartigan Global Native-Bridge: A 96-Active-Qubit Structured-Output Benchmark on NISQ Hardware

This record contains the finalized Madmartigan Native-Bridge benchmark paper and supporting documentation for a global 96-active-qubit structured-output benchmark executed on IBM superconducting hardware.

The benchmark extends the prior tiled Madmartigan scale-up work by testing whether the original seed-1337 16-qubit Madmartigan structured-output band could survive a stricter native-bridged condition: a single global 156-qubit hardware execution with 96 active qubits, six simultaneous 16-qubit regions, and native cross-region bridge operations inserted between tile regions.

The final benchmark used the T6 rank-2 physical tile layout on IBM Marrakesh, preserved the exact original 16-qubit Madmartigan tile kernel, used 4096 shots per hardware execution, and used no quantum error correction or post-selection. The native-bridge layer selected the available active-topology cross-region bridge edges, corresponding to physical pairs (97,87) and (86,85). The raw exact-kernel native-bridge circuit transpiled to depth 654 with 1241 CZ gates, 2653 SX gates, 2496 RZ gates, 156 measurements, and 72 barriers.

Across five global hardware executions, the raw bridged Madmartigan circuit produced repeatable Madmartigan-reference structured-output preservation, with mean F_XEB = 1.106965, mean HOG = 0.652515, mean Shannon entropy = 11.886623 bits per 16-qubit tile, and mean IPR = 3661.616. All 30 tile-level observations across five runs and six tile projections produced positive F_XEB, and all 30 HOG values exceeded 0.55.

The benchmark includes a three-control ladder: an exact-scaled generic RCS native-bridge control, a phase-scrambled native-bridge control, and a partial-entanglement native-bridge ablation. The generic RCS and phase-scrambled controls preserved their own-reference structure but failed to reproduce the Madmartigan-reference band. The partial-entanglement ablation produced a separate high-attractor own-reference mode while retaining only weaker Madmartigan-reference overlap.

The paper includes statistical control-separation analysis over tile-level observations, including t-intervals, Welch tests, Mann–Whitney tests, Cohen’s d, and bootstrap confidence intervals. The defensible claim is not full 96-qubit statevector fidelity, full tomography, fault-tolerant computation, or universal quantum advantage. The claim is narrower: repeatable, statistically separated, reference-specific structured-output preservation under global native-bridged NISQ execution.

This upload is intended as a technical benchmark record and reproducibility-indexed evidence package. Full executable circuit artifacts, QPY/QASM3 files, and source-level reproduction materials may be withheld from public release where necessary to preserve IP integrity and may be made available through protected review channels.

Media coverage / news release: The Quantum Insider, “Quantum Midi Posse Reports 96-Qubit Structured-Output Benchmark on IBM Hardware,” May 24, 2026.

https://thequantuminsider.com/2026/05/24/quantum-midi-posse-reports-96-qubit-structured-output-benchmark-on-ibm-hardware/