Quantum Dot Live Twin

Semiconductor quantum dot labs read out device state via dispersive/RF measurements running far faster than the classification loops meant to track them. When a twin falls behind, drift goes unnoticed until fridge time is burned re-characterizing. This project builds a live digital twin that stays synchronized with a quantum dot device's changing state under exactly that pressure — a passive mirror, not a tuning/control system. Frames come from a real QArray-simulated device trajectory (Rust-backed for speed). Measured on a 2000-frame run on AMD MI300X (PyTorch/ROCm): serial CPU baseline hits 0.222s worst-case lag; GPU-batched inference cuts that to 0.020s, an ~11x reduction in worst-case staleness. A rule-based triage agent then holds that same 0.020s worst-case while cutting compute further — FULL 245 / CHEAP 51 / SKIP 0 tier decisions from real queue depth, staleness, and drift signals, a decision genuinely consumed by the control loop, not narrated after the fact. An LLM supervisor (Fireworks AI) is wired in to tune those triage thresholds in the background, reasoning over backlog trends without blocking the real-time hot path. It did not produce measurable threshold updates during our testing window, so the deployed system runs on the validated static thresholds above — flagged here rather than overclaimed. The same stream is also exposed as a real QCoDeS Instrument, so a lab's existing control stack could subscribe to it with no new integration work. Results ship in an interactive control console. Known limitations — synchronous micro-batching, untrained classifier weights — are documented, not hidden.