Context
Superconducting-processor calibration involves interdependent measurements. Conventional scripts struggle with anomalous signals, while experts must continuously assess results and choose the next operation.
Work completed
The project organized expert knowledge into decision-tree Skills containing parameterized measurement commands, quantitative acceptance criteria and audit records. Human-in-the-loop workflow distillation and validated trajectories supported model fine-tuning.
Team-member contributions include workflow packaging, quality gates, audit logs, recovery and trajectory collection, retaining raw arrays, images, fitted values and accept/retry/skip/rollback decisions.
Reported results
On a 112-qubit processor, the system autonomously calibrated 108 qubits in 4.7 hours. A comparison with expert manual calibration on a 16-qubit subset showed agreement on 14 qubits.
These results apply to the experimental conditions in the paper, not to arbitrary hardware or industry tasks. This was an external research collaboration involving the team.