RB Protocol 01
Randomised Benchmarking
Average Clifford infidelity across length sweeps with a randomised twirl.
What the fit returns
Single-exponential survival fit returns the depolarising channel parameter.
Benchmarking Platform
Measure the vendor, not the brochure.
Six vendors, four published benchmark protocols, bootstrap confidence intervals on every cell, dated spoof-hardness witness on every XEB number. The scorecard a procurement committee can sign on.
Click a protocol column header to sort. Illustrative envelope, not a published scorecard.
Vendor · Backend
IBM Heron r2 superconducting
0.9968 ±0.08
0.9954 ±0.12
0.9831 ±0.42
LB-OK consistency
IonQ Tempo trapped-ion
0.9982 ±0.06
0.9971 ±0.09
0.9885 ±0.36
LB-OK consistency
Quantinuum H3 trapped-ion
0.9991 ±0.05
0.9985 ±0.08
0.9905 ±0.34
LB-OK consistency
Rigetti Ankaa-3 superconducting
0.9946 ±0.11
0.9928 ±0.16
0.9742 ±0.58
LB-OK consistency
Pasqal Orion β neutral-atom
0.9892 ±0.14
0.9871 ±0.19
0.9618 ±0.72
LB-WIDE consistency
QuEra Aquila neutral-atom
0.9841 ±0.17
0.9812 ±0.24
0.9498 ±0.89
LB-WIDE consistency
Cell saturation reflects fidelity. Whisker beneath each number is the 95% bootstrap CI half-width.
Framework SHA · dfa7c0
Heat map at a glance
One frame. Six vendors. Four protocols.
Hover a cell for methodology, sample size, and dated witness. Darker steel-blue tracks higher fidelity. Amber on Forrelation flags upstream-fit caveats. The same shape across vendors so a procurement reviewer reads the page in seconds.
Vendor / Protocol
RB
CB
XEB
FORR
IBM Heron r2
0.997 Randomised Benchmarking · Heron r2 RB-2Q on the most-used coupler. Heavy-hex topology. n=1024 shots · CI±0.08
0.995 Cycle Benchmarking · Heron r2 Sparse Pauli-Lindblad fit per coupling. n=1024 shots · CI±0.12
0.983 Linear Cross-Entropy · Heron r2 F_XEB at depth 12, n=16. Tensor-network frontier as of witness date. n=1024 shots · CI±0.42
LB-OK Forrelation (k-fold) · Heron r2 k=4 lower bound consistent at N=16 by exhaustive check. k=4 oracle · N=16
IonQ Tempo
0.998 Randomised Benchmarking · Tempo RB-2Q across all-to-all chain. Long coherence relative to gate time. n=1024 shots · CI±0.06
0.997 Cycle Benchmarking · Tempo Per-Pauli decay rates separated. n=1024 shots · CI±0.09
0.989 Linear Cross-Entropy · Tempo F_XEB at depth 12, n=16. Witness names defeated attack classes. n=1024 shots · CI±0.36
LB-OK Forrelation (k-fold) · Tempo k=4 lower bound consistent at N=16. k=4 oracle · N=16
Quantinuum H3
0.999 Randomised Benchmarking · H3 RB-2Q target on zone-graph racetrack. Spec-sheet number, not third-party measured. n=1024 shots · CI±0.05
0.999 Cycle Benchmarking · H3 Target per-Pauli rate. Pre-spine, deferred. n=1024 shots · CI±0.08
0.991 Linear Cross-Entropy · H3 F_XEB at depth 12, n=16. Witness anchored to the BFNV plus #P-hardness pair. n=1024 shots · CI±0.34
LB-OK Forrelation (k-fold) · H3 k=4 lower bound consistent at N=16. k=4 oracle · N=16
Rigetti Ankaa-3
0.995 Randomised Benchmarking · Ankaa-3 RB-2Q on tunable-coupler ladder. n=1024 shots · CI±0.11
0.993 Cycle Benchmarking · Ankaa-3 Cycle decay across two-qubit gates. n=1024 shots · CI±0.16
0.974 Linear Cross-Entropy · Ankaa-3 F_XEB at depth 12, n=16. Wider band, register sensitive. n=1024 shots · CI±0.58
LB-OK Forrelation (k-fold) · Ankaa-3 k=4 lower bound consistent at N=16. k=4 oracle · N=16
Pasqal Orion β
0.989 Randomised Benchmarking · Orion β RB-2Q on Rydberg-blockade CZ. Digital mode. n=1024 shots · CI±0.14
0.987 Cycle Benchmarking · Orion β Cycle decay on the digital surface. n=1024 shots · CI±0.19
0.962 Linear Cross-Entropy · Orion β F_XEB at depth 8, n=16. Mid-circuit measurement not supported. n=1024 shots · CI±0.72
LB-WIDE Forrelation (k-fold) · Orion β k=4 lower bound consistent but check CI of upstream fit. k=4 oracle · N=16
QuEra Aquila
0.984 Randomised Benchmarking · Aquila Analog Hamiltonian. RB on the digital surface where present. n=1024 shots · CI±0.17
0.981 Cycle Benchmarking · Aquila Cycle-style decay, Rydberg-blockade pulses. n=1024 shots · CI±0.24
0.950 Linear Cross-Entropy · Aquila F_XEB at depth 8, n=14. Wide band, analog modality. n=1024 shots · CI±0.89
LB-WIDE Forrelation (k-fold) · Aquila k=4 lower bound consistent with upstream caveats. k=4 oracle · N=16
Colour ramp · fidelity
0.93 0.97 1.00
LB-OK · lower bound consistent
LB-WIDE · upstream caveats
The four benchmark protocols
Four published primitives. Each cell on the leaderboard ties to one of them.
The protocols are not a DeployQuantum invention. Each is anchored to a published paper, a published estimator, and a published statistical procedure. The platform binds them to one signed scorecard shape per run.
CB Protocol 02
Cycle Benchmarking
Separates coherent from stochastic error per cycle, parameter-efficient relative to full process tomography.
What the fit returns
Pauli-twirled cycle decay returns a sparse Pauli-Lindblad coefficient map per coupling.
XEB Protocol 03
Linear Cross-Entropy
Random-circuit sampling primitive with a dated spoof-hardness witness on every cell.
What the fit returns
Linear cross-entropy estimator returns a fidelity proxy under a published anti-concentration assumption.
FORR Protocol 04
Forrelation (k-fold)
Provable separation rather than a measurement primitive. The lower bound is the proof object.
What the fit returns
Classical lower-bound certificate consistent with the published k-fold Forrelation oracle separation.
Anchors are cited in the manifest by paper identifier. This surface carries no person-name attribution.
Spoof-hardness witness
The frontier moves. The scorecard records the date.
Every XEB cell carries a dated reference. When classical attacks advance, affected scorecards demote with a dated note. The reference date is a field on every scorecard, not a footnote.
2022 Tensor-network attack of Pan-Zhang demotes Sycamore-2019 parameters. · 2024 Extended tensor-network methods further compress XEB-class hardness at fixed (n, depth). · 2019 Forrelation oracle lower bound is information-theoretic and unconditional within the query model. · 2025 Witness frontier reference date is a field per scorecard, not a footnote. · 2026 Frontier registry advances. Affected scorecards demote with a dated reference. · today F_XEB above zero on a noiseless simulator is uninformative about hardware advantage. · 2022 Tensor-network attack of Pan-Zhang demotes Sycamore-2019 parameters. · 2024 Extended tensor-network methods further compress XEB-class hardness at fixed (n, depth). · 2019 Forrelation oracle lower bound is information-theoretic and unconditional within the query model. · 2025 Witness frontier reference date is a field per scorecard, not a footnote. · 2026 Frontier registry advances. Affected scorecards demote with a dated reference. · today F_XEB above zero on a noiseless simulator is uninformative about hardware advantage. ·
The signed scorecard JSON
Every cell on the leaderboard is replayable from this manifest.
Same seed. Same SDK version. Same hash. A third party reruns the script and reproduces the cell byte for byte.
{
"schema_version": "0.1.0",
"vendor": "IBM",
"backend_name": "ibm_torino",
"vendor_sdk_version": "qiskit-1.x (pinned)",
"benchmark": "RB_2Q",
"n_qubits": 2,
"shots_per_circuit": 1024,
"transpile": {
"seed": 7,
"basis_gates": ["sx", "rz", "cz"],
"qubit_indices": [12, 13]
},
"measured": {
"primary_metric_name": "r_per_clifford",
"value": 0.9968,
"ci_95_lower": 0.9960,
"ci_95_upper": 0.9976
},
"spoof_hardness_witness": {
"named_assumption": "BFNV anti-concentration plus #P-hardness",
"defeated_attacks": ["..."],
"not_defeated_attacks": ["..."],
"frontier_reference_date": "2026-05-10"
},
"fit_residuals_path": "build/output/rb_2q_fit_residuals.json",
"rb_certificate_hash": "<sha-256>",
"scorecard_hash": "<sha-256>"
} Field index
- schema_version "0.1.0"
- vendor "IBM" | "IonQ" | "Quantinuum" | "Rigetti" | "Pasqal" | "QuEra"
- backend_name string
- vendor_sdk_version string (pinned at run time)
- benchmark "RB_1Q" | "RB_2Q" | "CB" | "XEB" | "Forrelation"
- n_qubits integer
- shots_per_circuit integer (default 1024)
- transpile.seed integer (transpiler seed, deterministic)
- transpile.basis_gates array of strings
- transpile.qubit_indices array of integers (physical qubits used)
Eleven more fields in the manifest. Each one replayable.
framework_sha dfa7c0
dfa7c0777eca70ffb835a1ec6f3f351cee7b2ea848ba3a7f228082d08db80bc7 Pinned in every manifest. The classical pre and post script hashes pin alongside, and the scorecard hash closes the chain.
What this does not claim
A measurement protocol, not a quantum-advantage proof.
F_XEB above zero on a noiseless simulator is uninformative about advantage. On a real QPU it is hard to spoof classically only under the BFNV anti-concentration plus #P-hardness conjecture pair, which can fail. The witness demotes when the frontier advances. The Forrelation lower bound is information-theoretic at the oracle level. The platform separates the conjecture-conditional claim from the unconditional one in every scorecard so the buyer reads each on its own footing.
Per-vendor scorecards are deferred to Tool 3 and Tool 5 operational, with a real-QPU run in the hardware-run-evidence block. At this ship gate, the release state is internal-only and the leaderboard you see is the shape the scorecard takes when it ships, not a published vendor ranking.
Questions
The five questions a procurement reviewer asks.
- 01
Why is the leaderboard at hero scale and not a marketing chart further down?
The scorecard is the product. A vendor brochure is a chart. The buyer reads the same shape an auditor reads: cells with a fidelity number, a CI whisker, a witness date. Marketing positioning lives in vendor brochures. Procurement lives in the scorecard.
- 02
Are the cells published vendor measurements?
No. At this ship gate, the release state is internal-only and the real-QPU run is deferred to Tool 3 plus Tool 5 operational. The leaderboard renders the shape the scorecard takes when shipped. Numbers derive from each vendor's published median two-qubit fidelity in the hardware-roadmap registry, perturbed by a protocol-specific delta. Treat the page as illustrative envelope, not as a vendor ranking.
- 03
What exactly does the bootstrap whisker mean under each cell?
A 95% bootstrap confidence interval around the primary metric, computed by resampling the underlying circuit-shot dataset. The whisker is a visual rendering of the CI half-width. Where the CI is wide relative to the difference between two vendors, the chart says they are not separable at the run's shot budget. The CI is reported as a number in every scorecard JSON, not only as a whisker.
- 04
Why does Forrelation report LB-OK rather than a fidelity?
The Forrelation protocol is a provable oracle separation, not a fidelity primitive. The cell is a consistency check against the classical lower bound. LB-OK means the run is consistent with the published lower-bound exponent. LB-WIDE means consistent but with upstream fit caveats in the manifest.
- 05
What does the witness date mean and when does a scorecard demote?
Every XEB cell carries a dated frontier reference. When the classical-attack frontier advances, the affected scorecards demote in the next registry update, with a dated note recording which attack class moved and which cells were re-rendered. The witness is not a frozen number. It is a measurement under a stated and dated assumption envelope.
Start with one run
Six vendors. Four protocols. One signed scorecard.
Request a run. We respond within one business day. The customer-asset attestation is a precondition, and the release-gate state is recorded against your engagement, not against a generic download.