nx_text2motion_gate.nx

nx_text2motion_gate.nx -- TEXT-CONDITIONED neural motion, proven with the honest ML criterion (GENERALIZATION to held-out captions -- a keyword if-else cannot generalize to sentences it never saw). Chain: caption -> LEARNED perceptron encoder -> motion class -> class-conditioned generated tokens -> VQ decode -> rig -> video. T1 the encoder LEARNS: perceptron training mistakes -> 0 (separated the training set) T2 it GENERALIZES: 6/6 HELD-OUT captions (novel word combinations, never trained as sentences) classify correctly -- AND the UNTRAINED (zero) encoder gets ~1/6, proving the LEARNING did it (not the vocab) T3 full chain: a held-out caption -> class -> generated motion tokens -> decoded poses MOVE -> APNG rendered T4 TEXT STEERS: two different captions -> different classes -> different token sequences Emits knowledge/synth_text2motion.png (APNG of the motion generated FROM a sentence). license_tier: ORIGINAL expect_exit: 0