UniMind: Calibration-Aware Scheduling and Placement for Quantum-Accelerated AI Middlewars
Yong Xun Tan · Zenodo (CERN European Organization for Nuclear Research) · 2026
UniMind is a user-space middleware for quantum-accelerated AI workloads that scores, places, executes, and refreshes calibration data in a closed loop. This record contains the v7 manuscript in two editions (12-page full and 6-page compact), each as compiled PDF and LaTeX source. The paper is a measurement study of how a quantum runtime should maintain scheduling decisions when hardware calibration changes over time. We instrument a 156-qubit superconducting device (ibm_marrakesh) across three primary daily snapshots (D0-D2, 2026-08-29 to 08-31) plus 34 hourly calibration snapshots (2026-09-01 to 09-02, three backends: marrakesh, kingston, fez). Four findings: (1) calibration validity is temporally bounded and small (day-over-day rank correlation 0.579, top-10 Jaccard 0.25, stale top-3 pins 1.94-2.1x worse across 24-34 h; hourly top-3 margins ~1e-3 vs median |dC| = 0.0127); we formalize the Calibration Validity Horizon T_valid as a lower-bound design primitive. (2) Adaptive refresh (top-10 Jaccard 0.005, ~0.04 ms per scoring pass) dominates static and periodic policies at near-zero cost; a 144-threshold sweep places the default at the Pareto knee. (3) Refresh, not placement selection, dominates end-to-end reliability at the measured scale (+33/37 pp with fresh pins; pinned vs free is a bounded null, 70.0% vs 76.7%, Fisher p = 0.77, MDE ~ 32 pp). (4) Proxy-trained layout models do not transfer: re-fitting the 7-dim utility J(G) on real device labels collapses its size out-of-sample Spearman from 0.818 to 0.13; feature ablation and a connectivity-vs-quality variance crossover at k ~ 100 isolate the mechanism. All experiments are open-source and reproducible from the companion repository; QPU job IDs and calibration snapshots are recorded in its data directory. Every claim traces to a recorded job or snapshot; a failed exploratory result (rank correlation 1.000 -> 0.771) is reported as unreplicated. The LLM-orchestration layer is a separable controlled-parameter reliability model, not a hardware claim.