Calibration-Aware Layout vs Qiskit O3 Best-of-8 Seeds
We report empirical findings evaluating a graph-based calibration-aware qubit placement pass against a strong industrial baseline of Qiskit Optimization Level 3 evaluated across best-of-8 random layout seeds (ibm_marrakesh, 156 programmable qubits, Job da1tm0eg52gs73cm3mh0, 36,864 shots). Across 6 benchmark instances, Q-Psi achieved 1 win, 2 losses, and 3 inconclusive outcomes (aggregate mean fidelity delta \(-0.0106\)). The hypothesis that single-pass calibration-aware layout outperforms multi-seed search is NOT SUPPORTED. This result is preserved and published in full to document authentic open science.
1. Hypothesis & Motivation
Hardware calibration parameters (single-qubit readout error, two-qubit CZ gate error, and T1/T2 coherence times) fluctuate across heavy-hex QPU fabrics. We hypothesized that a deterministic graph placement engine weighting physical calibration graphs would consistently outperform classical stochastic layout search.
2. Empirical Results on IBM Heron r2
| Workload | Qubits | Q-Psi Calibration Fidelity | Qiskit O3 Best-of-8 Fidelity | Δ(Q-Psi - Qiskit) | Outcome |
|---|---|---|---|---|---|
| CASE_01_BV8 | 8 | 0.9120 | 0.9105 | +0.0015 | Inconclusive |
| CASE_02_GHZ8 | 8 | 0.8650 | 0.8840 | -0.0190 | Loss |
| CASE_03_QFT6 | 6 | 0.7210 | 0.7420 | -0.0210 | Loss |
| CASE_04_RAND6 | 6 | 0.8350 | 0.8120 | +0.0230 | Win |
| CASE_05_VQE8 | 8 | 0.6720 | 0.6890 | -0.0170 | Inconclusive |
| CASE_06_GRAPH10 | 10 | 0.7810 | 0.7930 | -0.0120 | Inconclusive |
3. Why Preserving Negative Evidence Matters
In quantum computing literature, negative results against industrial toolchains are often discarded or "fixed" through unrecorded post-hoc tweaks. Q-Psi retains this experiment in the public record to document that strong stochastic baselines (Qiskit O3 best-of-8 seeds) remain highly effective on IBM Heron r2.