nx_fnet_model_gate.nx

nx_fnet_model_gate.nx -- GATE for MODEL-001: the FIRST end-to-end sovereign SUB-QUADRATIC MODEL. A real sequence classifier with no attention anywhere: tokens -> EMBED (trained) -> FNET token-mix (sub-quadratic) -> relu FFN -> linear -> softmax-CE Trained by AdamW JOINTLY over the embedding table + FFN/head weights. The embedding has no attention matrix and no quadratic cost -- token mixing is the parameter-free Fourier transform. Embeddings are trained without a dedicated op: the tape yields dL/dx at the input leaf, which is SCATTERED back into the embedding rows. Task: classify whether token[0] > token[3] over 8 length-4 sequences (vocab 4). This genuinely needs the FNet to MIX positions 0 and 3 and the relu to compare -- a bag-of-words linear model cannot do it. G_train the model learns: accuracy >= 7/8 AND final loss < first loss. G_repro bit-exact: train twice -> identical accuracy + final loss bits. Evidence -> knowledge/status/fnet_model.log (FNETMODELGATE authored=organ ... verdict=GREEN). license_tier: ORIGINAL