nx_mt_r3_gate.nx

nx_mt_r3_gate.nx -- GATE for MT-R3: REAL TEXT via the production BPE tokenizer (nx_bpe). The MT arc moves off hand-coded integer toy indices: input/output are now actual TEXT STRINGS, tokenized to subword IDs by the team's shipped Byte-Pair-Encoding organ and detokenized back. The learned word-map (proven in R0/R1/R2, reused from nx_mt_core) maps EN token IDs -> ES token IDs in between. This is the text-interface bookend the rest of the arc needs; it does NOT yet do open-vocabulary sentence MT (that's R4: real corpus / trained weights). Sovereign self-contained vocab (no external files; nx_bpe builds vocab in-code): EN (with merges, so encoding is real BPE): bytes g o s e a t ; merges (e,a)->ea, (s,ea)->sea, (t,ea)->tea, (g,o)->go. Words: "sea" "tea" "go" each encode to ONE token; "ea" is a SHARED subword of sea/tea. Out-of-vocab "seat" decomposes to known subwords [sea, t]. ES (decode side): whole-word tokens "mar" "te" "ir". translation map (learned via nx_mt_core): sea->mar, tea->te, go->ir. FOUR GATES: A BPE ROUND-TRIP: decode(encode(x)) == x byte-exact for "sea","tea","go". B SUBWORD DECOMPOSITION (the BPE value): encode("seat") == [sea, t] (OOV word -> known subwords) and round-trips byte-exact -- a hand-coded index table cannot do this. C END-TO-END TEXT TRANSLATE: translate("sea")=="mar", translate("tea")=="te", translate("go")=="ir" (real string in, real string out; 3/3). D UNTRAINED FAILS (liar-kill): the zero-epoch map mistranslates (< 3/3) -> the LEARNED map does the translating, not the tokenizer. genealogy_id: sennrich_haddow_birch_2016_bpe (realized_in nx_bpe) + rumelhart_1986_backprop lineage_id: sovereign_neural_mt_r3_bpe_realtext_v1 license_tier: ORIGINAL