nx_mt_r0_gate.nx

nx_mt_r0_gate.nx -- GATE for MT-R0: the first rung of the SOVEREIGN NEURAL MACHINE TRANSLATION arc (GEN-WL-WL-NN-TRANSLATION). Proves, by RUNNING, that a sovereign neural net LEARNS to translate -- from random/zero init via the team's own scalar-tape autograd, NOT a hardcoded table. The trainable atom of neural MT = a learned bilingual map: one-hot source word -> linear layer W (the learned parameters) -> output vector -> argmax -> target word. Composes ONLY shipped+gated organs (nx_autograd TRAIN-R1 GREEN + nx_f32 tower). No attention/weights loaded; this is the BACKWARD/learning half proven on a translation task in miniature, the foundation the rest of the arc (sequences -> FNet-mix context -> BPE vocab -> corpus -> the /video /translate endpoint) climbs on. Demo dictionary (English -> Spanish), TARGET INDICES ARE A PERMUTATION (not identity) so a pass cannot come from a trivial copy: src EN: 0 hello 1 family 2 love 3 good 4 day 5 water tgt ES: 0 familia 1 bueno 2 hola 3 dia 4 agua 5 amor (vocab in a DIFFERENT order) map: hello->hola(2) family->familia(0) love->amor(5) good->bueno(1) day->dia(3) water->agua(4) => tgt = [2,0,5,1,3,4] (a real permutation the net must LEARN) FOUR GATES: A LEARNS: after training, argmax == tgt for ALL 6 pairs (acc 6/6) AND mean loss < 1/1000 AND loss decreased. The exact statement of "the net learned to translate". B UNTRAINED FAILS (liar-kill): the zero-epoch (W=0) model scores < 6/6 -> the LEARNING did it, not the harness/architecture. C BIT-EXACT REPRODUCIBLE: train twice from zero-init; final W bits identical (determinism exceed-axis carried into ML, like TRAIN-R1 gate C). D DATA-DRIVEN (anti-hardcode liar-kill): train against a DIFFERENT permutation tgt2; the net learns tgt2 (acc_vs_tgt2 == 6/6) AND no longer matches the original (acc_vs_tgt < 6/6). Kills any "the answer was baked in" cheat -- it learns whatever data it is given. genealogy_id: linnainmaa_1970_reverse_mode_ad + rumelhart_1986_backprop (realized_in nx_autograd) lineage_id: sovereign_neural_mt_r0_learned_wordmap_v1 license_tier: ORIGINAL