Real HardwareApril 2026
Hardware Validation
We ran real circuits on 5 quantum devices from 3 vendors, spanning both superconducting and trapped-ion architectures. Here's how our predictions held up against actual hardware.
0.96
Pearson r
0.92
R²
0.096
RMSE
15
Pairs
Predicted vs. Observed Fidelity
IBM (uncalibrated)
IQM (calibrated)
IonQ (calibrated, trapped-ion)
Perfect prediction
All 15 Circuit-Device Pairs
| Circuit | Device | Vendor | Predicted | Observed | Δ | Calibration |
|---|---|---|---|---|---|---|
| 4Q GHZ | ibm_fez | IBM | 0.7311 | 0.8738 | +0.143 | None |
| 4Q GHZ | ibm_kingston | IBM | 0.8172 | 0.9675 | +0.150 | None |
| 4Q GHZ | ibm_marrakesh | IBM | 0.7356 | 0.9404 | +0.205 | None |
| 4Q VQE Ansatz | ibm_fez | IBM | 0.2848 | 0.2849 | +0.000 | None |
| 4Q VQE Ansatz | ibm_kingston | IBM | 0.4594 | 0.2783 | -0.181 | None |
| 4Q VQE Ansatz | ibm_marrakesh | IBM | 0.2824 | 0.2742 | -0.008 | None |
| 4Q Deep Ladder | ibm_fez | IBM | 0.2596 | 0.2778 | +0.018 | None |
| 4Q Deep Ladder | ibm_kingston | IBM | 0.3962 | 0.4780 | +0.082 | None |
| 4Q Deep Ladder | ibm_marrakesh | IBM | 0.2345 | 0.2908 | +0.056 | None |
| 4Q GHZ | iqm_garnet | IQM | 0.8658 | 0.8870 | +0.021 | Adaptive |
| 4Q VQE Ansatz | iqm_garnet | IQM | 0.5290 | 0.5300 | +0.001 | Adaptive |
| 4Q Deep Ladder | iqm_garnet | IQM | 0.3792 | 0.3740 | -0.005 | Adaptive |
| 4Q GHZ | ionq_forte | IonQ | 0.8667 | 0.9530 | +0.086 | Adaptive |
| 4Q VQE Ansatz | ionq_forte | IonQ | 0.5320 | 0.5620 | +0.030 | Adaptive |
| 4Q Deep Ladder | ionq_forte | IonQ | 0.3827 | 0.3620 | -0.021 | Adaptive |
Same-Vendor Accuracy
On IBM hardware (3 devices, 9 pairs), predictions achieve r = 0.96 with no device-specific tuning. The physics model works directly from public Metriq benchmark data.
Superconducting Cross-Vendor
On IQM Garnet, stale benchmark data caused over-prediction. One-shot adaptive calibration recovered accuracy to Δ = 0.001 on out-of-sample circuits — a 94% RMSE reduction.
Trapped-Ion Cross-Vendor
On IonQ Forte (trapped-ion), the same calibration protocol generalized: RMSE dropped 82% and out-of-sample VQE matched within 3 percentage points. Architecture-agnostic.
Data from real hardware experiments on IBM Quantum Platform and AWS Braket, April 2026. IonQ Forte accessed via AWS Braket Hybrid Jobs (priority queue).