A three-study series established the components of thermal-adaptive precision control for orbital Mixture-of-Experts (MoE) inference: expert placement offers no thermal headroom 1; INT8 duty-cycling is the working lever, measured accuracy-free at a 1.55–1.58°ø energy reduction 2; the linear duty-cycle mixture is a valid planning model and the precision-swap break-even dwell is ≈12 min on this hardware (the latter from the same measurement campaign); and measured routing shows thermal feasibility is workload-dependent 3. Here we close the loop. We build a controller that reads a (modeled) time-varying radiator capacity over a 90-minute low-Earth-orbit cycle and a measured serving-power signal, and selects expert precision in real time—subject to a measured-power safety backstop and the 12-minute minimum-dwell constraint—to keep dissipation under the envelope while maximizing time at full precision. We run it live on Grace–Blackwell hardware for two full orbits per configuration, against static baselines. The mechanism is demonstrated. The controller records zero sustained capacity violations across both orbits in both configurations, precisely where a static 16-bit baseline violates in the sunlit binding phase (two violations each). With a BF16↔Q8 pairing the rescue is accuracy-free; with a BF16↔Q4 pairing the controller meters a real quality cost (perplexity 9.08, near full-Q4) precisely because the binding envelope forces ≈89% of tokens to low precision— demonstrating honest trade management rather than a free lunch. We also report a criterion the controller failed: it spent only 9–11% of each orbit at full precision against a pre-registered target of 35%. The failure is instructive and we did not patch it: the controller’s measured-power backstop is more conservative than the open-loop arithmetic that set the target—closed-loop BF16 draws ≈17 W against an eclipse capacity of ≈17.7 W, leaving 3–5% real headroom where the open-loop derivation assumed 15%. The controller is correctly more cautious than the design math, in the safe direction. This open-loop/closed-loop conservatism gap, quantified on measured power, is the paper’s main systems finding.
Vincent Zhang (Tue,) studied this question.