Measured Routing Traces for Orbital Mixture-of-Experts Thermal Feasibility: A Protocol Correction and a Domain-Dependence Result.

Two companion studies quantified thermal feasibility for orbital Mixture-of-Experts (MoE) inference using synthetic expert-routing priors: the first showed expert-placement optimization adds essentially nothing in the thermally binding regime 1; the second measured thermal- adaptive precision (INT8 duty-cycling) as the lever that does work 2. Both flagged the synthetic priors as their primary threat to validity. Here we replace them with measured routing: 285,696 WikiText-2 tokens and 67,072 Python (MBPP) tokens captured via forward hooks on all 16 routers of OLMoE-1B-7B. Three results. (1) The placement conclusion survives measured routing. Thermal- seeded local search ties thermally balanced placement (median bandwidth improvement 0.00%) on real traces, and the boundary-classification decision reproduces on natural-language routing. (2) A substrate-protocol correction. Feeding measured (already aux-loss-balanced) expert shares through the model’s load-balance axis re-balances them twice, deflating the binding categories and producing a spurious decision flip; we state the faithful protocol (measured shares enter directly; load-balance axis held at identity) with a worked example, and add a second correction: the radiator-bound tail must be reported as a per-cell rate, not a pooled count. (3) Thermal feasibility is strongly workload-dependent — and reality sits at the harder end of the synthetic envelope. Per-layer load skew is ≈2.8 for natural language but ≈7.4 for code; under the faithful protocol, 50% of natural-language and 75% of code configurations are physically infeasible on balanced-sized radiators, with radiator-bound tails of 50% and 75% respectively (synthetic: 40%). Code routing admits zero feasible placements on a balanced-sized radiator, and its median rescue requirement (3.39×) is roughly double natural language’s (1.95×) — beyond the measured 4-bit energy factor (1.77×). The operator implication: workload mix is a first-order thermal design variable; a code-heavy serving mix needs roughly 1.7–2.7×more energy-reduction or radiator margin than a natural-language mix, which precision software alone does not supply. All analyses were pre-registered before data and reported as the frozen criteria computed them.