nx_moe.nx

nx_moe.nx -- sovereign MIXTURE-OF-EXPERTS sparse FFN (the last census OPPORTUNITY axis "Mixture of experts (sparse)", momentum 3; Mixtral / Switch-class, Shazeer 2017 / Fedus 2021 / Jiang 2024). Per token: router logits = x @ W_router^T -> softmax -> TOP-K experts -> renormalize the k probs -> out = sum_k p_i * SwiGLU_FFN_expert_i(x). Only k of n_experts FFNs run per token (the sparsity that lets total parameters scale past per-token compute). Experts accumulate in SELECTION ORDER (prob-descending; ties -> lowest index) -- deterministic, documented, and mirrored by the gate's oracle. Composes the EXISTING primitives: nx_f32_matmul_t (weights transposed, the LLM convention), nx_f32_softmax, nx_f32_silu. Weights here are plain F32 buffers; binding real MoE GGUF expert tensors (and the lazy/ quantized paths) is the documented follow-on once an MoE model is on disk. Touch counters per expert make the sparsity MEASURABLE (gated), not asserted. genealogy_id: shazeer_2017_moe + fedus_2021_switch + jiang_2024_mixtral lineage_id: substrate_moe_v1