Shell-Coherent Expert Specialization: Alignment-Posture Assignment in Mixture-of-Experts Language Models

We present Shell-Coherent Expert Specialization (SCES), a declarative approach in which MoE experts are assigned to alignment postures (shells) rather than knowledge domains. The VortexGate directs each prompt to the expert whose shell matches the prompt's TERA geometry. A calibration cue system makes alignment posture explicit at inference time. Evaluation via HarmonicCoherenceSuite measures yoked-rate (fraction of tokens accepted on first sample) and vortex tightness. An ablation compares shell-based vs domain-based vs emergent specialization on the same base model and corpus. SCES is the first MoE specialization approach where the assignment axis is alignment posture (a continuous geometric property) rather than knowledge domain (a discrete categorical property).