nx_mt_core.nx

nx_mt_core.nx -- reusable CORE of the sovereign neural MT arc (GEN-WL-WL-NN-TRANSLATION). A learned word-map W (T x S), SHARED across all sequence positions, trained by the team's scalar-tape autograd (MSE to one-hot targets, full-batch gradient descent, zero-init). Applied per-position it translates token SEQUENCES of any length. R0 (nx_mt_r0_gate) proved the single-token atom inline; this factors that into a reusable organ and generalizes it to multi-token sequences with weight-sharing. Because W is position-independent, the map GENERALIZES to HELD-OUT sequences (combinations never trained). CONTEXT / reordering (an output token depending on MORE THAN its own input token) is R2, via fnet_mix -- not here. Shared-weight gradient note: a param W[j,s] is one tape leaf referenced by EVERY (sentence, position) token whose source word is s, so backward accumulates m contributions -> the GD update factor is (1 - 2*lr*m) where m = that word's occurrence count. Callers must pass an lr with 2*lr*m_max < 2 (convergence) -- lr is a PARAMETER, never hardcoded (no magic numbers). genealogy_id: rumelhart_1986_backprop (realized_in nx_autograd) lineage_id: sovereign_neural_mt_core_shared_wordmap_v1 license_tier: ORIGINAL