Headline results

Evidence that the maintenance layer reaches its floor, plus the limits of that evidence.

Live hardware: 107-channel closed-loop regulation on Rigetti Cepheus-1-108Q converged in twelve parallel 150-shot acquisitions in both runs. At the same budget, finite differences could not complete its first update.

Controlled validation: on IQM Emerald, five of five pre-registered hypotheses passed under the committed test plan.

Third-vendor validation: on IBM Heron at 156 qubits, the runtime reduced RMS error by 45.8% under a frozen budget. As the problem grew, its advantage over the tested comparison methods widened from 1.7× to 2.5×.

Independent reference implementation: the published equations held a coupled, drifting scale ladder flat from one hundred thousand to one hundred million channels, with one parallel acquisition per cycle and linear digital work. The in-class diagonal Broyden comparison also survived every rung and reached a lower floor, as committed.

Analog compute stack: a pre-registered depth-eight simulated stack with 8.4 million resident weights retained 78.6% top-1 agreement under elementwise drift, versus 65.7% free-running and 70.1% with scheduled recalibration. A 16,384-channel structured case retained 75.4%, versus 61.3% and 66.9%.

Scope: the maintenance layer and its scaling floor are demonstrated. A native-drift photonic or in-memory hardware pilot and an end-to-end workload executed across both digital and analog substrates remain undemonstrated.