Headline results

The latest frozen benchmark, physical-hardware evidence, and precise evidence labels.

Usable means the verified error is inside the published threshold.

Native-drift maintenance on a public-cloud QPU: 96 qubits of ibm_kingston, 50 of 50 cycles over nine hours, decision SUPPORTED under a rule fixed before the run (protocol sha256 ab2168ea…, independently reproduced from the raw records). On the 34 qubits the monitor found drifting, usable output rose 25.7 points (9.6% to 35.4% in spec); the contract residual held at 0.27 times the measured drift with a two-cycle lock; maintenance downtime was zero across every job; and an end-to-end AI pipeline through the maintained stage gained 34.0 points of numeric validity. The aggregate hypothesis stayed inside the loose fresh-floor tolerance; the registered drifting-subset branch carries the claim, and we say so.

The same run measured the feasibility wall it replaces: one 20-point sequential calibration sweep at 96 qubits costs 22 QPU-minutes, 2.2 times an entire monthly public-cloud allocation, against 97 seconds of total probe spend for the whole nine-hour maintained run.

Latest frozen QSC-Bench confirmation: retained-residual and fixed-window members completed 210 of 210 registered simulator runs. In the primary scale ladder, both completed 30 of 30 seeds at every width through 65,536 controls with a flat acquisition exponent; the retained median was approximately twelve acquisitions.

Development extension: both confirmed five of five runs at one million controls. This is development evidence, not a primary confirmation tier and not an independent community replication.

Earlier physical hardware: a Rigetti Cepheus closed loop under software-commanded disturbance and static or ablation diagnostics on IQM and IBM hardware. Those earlier archives did not establish native-drift compensation on a QPU; the ibm_kingston campaign above now does, on its stated drifting-subset branch.

Independent manuscript reference: a separate controlled-simulation ladder reaches one hundred million channels at one seed on the top rung. That range, plant, and implementation do not authorize a production switching rule at the hosted 4,096-control ceiling.