I ran a trained circuit on a real IBM quantum computer — here's the noise

· quantum, ibm-quantum, hardware, error-mitigation, zne, honest-benchmarks


Most of our work runs on a simulator — a normal computer pretending to be a quantum one. This time we sent one trained circuit to a real quantum computer at IBM (ibm_fez), over the internet, for free. Real hardware is noisy: the delicate qubits get bumped, so answers come out fuzzy. The question is how fuzzy, and how much of it we can clean up — honestly accounting for what the cleanup costs.

Three columns, one honest story

We never train on a real QPU (it burns free minutes for zero learning). We train on the simulator, then run inference on hardware. The native-Qiskit circuit provably matched the PennyLane model to ~1e-15 before it flew. Backend: ibm_fez, transpiled depth 63, 12 two-qubit gates.

Results

measurement⟨Z₀⟩ (sample 0)what it is
simulator (exact)+0.724perfect, noise-free reference
raw hardware+0.486real QPU, noise drags it toward 0
ZNE-mitigated+0.870error-mitigated, back near the clean value

Figures

sim vs raw vs mitigated bar chart

Noise, then cleanup (on-brand, generated): the simulator’s clean ⟨Z₀⟩, the real hardware’s noise-dragged value, and where zero-noise extrapolation pulls it back — bought with 3× the shots.

Honest caveats

This validates that a sim-trained model runs on real quantum silicon and that error mitigation lifts the raw signal — it is not a quantum-advantage claim (D1).

Reproduce it

Job id(s) d97qoc52su3c739if17g. Every number traces to summary.json / equivalence.json (D7). experiments/20260709-141149_p3-ibm-fez.