Live hardware
Demonstrated.
Rigetti. 107 channels. 12 acquisitions.
IQM. Five of five registered hypotheses.
IBM. 45.8% lower RMS error. 156 qubits.
Research & reproducibility
Evidence for the
hybrid architecture.
Hardware, simulation, and architectural projection kept separate.
THE CLAIMS LEDGERhardware · simulation · projection
Controlled simulation
Scale tested.
10⁸ channels. Flat to four decimals.
Photonics. WDM, pulse shaping, and meshes.
Boundaries. Misses stay public.
Classes are not performance claims. Test the interface.
Open compatibilityScale ladder
10⁵ to 10⁸ channels.
Flat to four decimals.
Pre-registered controlled simulation.
10⁵flat
10⁶flat
10⁷flat
10⁸flat
Experiment archive
Results across
the application map.
cycles to lock
256 to 4,096 simulated wavelength channels, with 0.041–0.058 dB tail error.
Strehl ratio
Locked in six cycles at 128, 640, and 4,096 simulated control pixels.
unitary fidelity
Measured at 4,096 simulated phase controls; 0.9992 at 256 controls.
normalized result
Per-cell sensing reached the 0.95 operating threshold in six cycles.
globally coupled variables
Twenty of twenty fixed and changing-response trials cleared the verified RMS threshold in six acquisitions.
median final error
Controlled response-matching simulation under a fixed six-acquisition budget.
channels tested
Finite differences, SPSA, and CMA-ES matched no control under the same one-read budget. Simulated.
measurements to usable
Zero setup reads at n = 1,024, with every comparison setup and diagnostic read charged to the same budget. Controlled simulation.
channels held flat
From 10⁵ to 10⁸ in the pre-registered controlled-simulation ladder.
channels by D = 6
No alternative cleared the deadline. Live-confirmed through 4,096 channels.
resident weights held
Depth eight. 78.6% top-1 agreement versus 65.7% free-running and 70.1% scheduled recalibration. Controlled simulation.
monitored columns
75.4% agreement versus 61.3% free-running and 66.9% scheduled recalibration. Controlled simulation.
Misses beside passes
Predictions that did not survive testing.
More cycles did not always win. Delayed acquisition diverged. Light commissioning was insufficient.
Public work
Reproduce it yourself.
Plants, baselines, seeds, thresholds, and results.
01Reproducible Headline Results
Published reference implementation, pre-registered scale ladder to one hundred million channels, analog compute-stack campaign, per-seed records, and figure source data.
02Rigetti Cepheus hardware107-channel physical-hardware runs, job identifiers, billing metadata, registration, and finite-difference control.
03IBM Heron hardwareFrozen twelve-acquisition comparisons through 32 channels and the 156-qubit hold result.
04QUBO information boundaryVector-feedback, ranking-only, measurement-priced greedy, and blind comparisons behind the information-bound result.
05Combinatorial reproductionCommitted planted-instance experiments, baselines, budgets, and retained raw result records.
06Grid Expansion PlanningMeasurement-priced grid optimization from IEEE cases to a mega-scale ladder.
07Fraud-model drift adaptationFrozen-budget recovery under controlled model drift with pre-registered hypotheses.
08WDM gain flatness256 to 4,096 wavelength channels, six-to-seven-read lock, and a serial-coverage comparison.
09Photonic mesh fidelityAn explicit complex mesh unitary, coupled phase drift, and 4,096 controlled phases.
10Pulse-shaper controlTransform-quality hold across 128 to 4,096 spectral channels with a matched scalar-feedback arm.
11QPU shot budgetA fixed fifteen-acquisition budget, binomial shot noise, and 107 to 4,096 controlled channels.
12GHZ fidelityStatevector-verified evaluation with pre-registered and amended records preserved.
13THz beam tracking1,024 to 4,096 elements, moving-beam hold, and per-element versus scalar instrumentation.