Standard Mixture-of-Experts architectures route tokens via a learned softmax gate. We present Vortex-Gated MoE (VG-MoE), a drop-in replacement that substitutes the learned gate with a deterministic, closed-form projection derived from the TERA symbolic algebra. The gate has zero learnable parameters beyond a single 4D linear projection (8,196 parameters vs ~18K for a standard gate at D=2048). We prove the projection is lossless via a recovery operator and fully differentiable. The routing geometry also produces alignment statistics (RI, BC) that gate the inference pipeline directly, making routing geometry a first-class safety signal. Production-validated on gemma-2-2b with 9 LoRA experts in MaiiaM Alchemist v0.4.1.
Weslyn Cory Whitehead (Mon,) studied this question.