nx_mt_r1_gate.nx

nx_mt_r1_gate.nx -- GATE for MT-R1: SEQUENCES. Proves, by RUNNING, that the sovereign MT core (nx_mt_core) translates multi-token SEQUENCES with a SHARED, position-independent learned word-map -- and that the map GENERALIZES to HELD-OUT sentences it was never trained on. This is the seq2seq plumbing + the no-overfit test harness the rest of the arc rides on; R0 proved the single-token atom, R2 will add cross-token CONTEXT (reordering) via fnet_mix. Dictionary (EN->ES), target indices are a PERMUTATION (not identity): tgt_map=[2,0,5,1,3,4]. 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 TRAIN = 4 sentences x 3 tokens, arranged so EVERY word occurs EXACTLY TWICE (m=2) -> with lr=1/8 the GD factor (1-2*lr*m)=0.5 is stable+monotone for all weights. HELD-OUT = 3 sentences whose token TRIPLES never appear in training (novel combinations of the same vocab). FOUR GATES: A LEARNS (train set): per-token acc == all-train-tokens, sum-loss < 1/1000, loss decreased. B GENERALIZES (held-out): per-token acc == all-held-tokens on sentences NEVER trained -> the shared map is compositional/position-independent, not a per-sentence memorization. C BIT-EXACT: retrain from zero-init -> identical W bits. D UNTRAINED FAILS (liar-kill): zero-epoch model scores < all-held-tokens on held-out. genealogy_id: rumelhart_1986_backprop (realized_in nx_autograd) lineage_id: sovereign_neural_mt_r1_sequences_v1 license_tier: ORIGINAL