nx_nano_lm.nx

nx_nano_lm.nx -- OUR OWN SOVEREIGN nanoGPT (the bigram/char rung where nanoGPT+makemore start), built from the Nishi rung up. A char-level language model that ACTUALLY TRAINS (integer gradient descent) and GENERATES text back. Model: a V*V logit table W (W[prev][next]); forward = softmax(W[prev]); cross-entropy gradient = probs - onehot (NO log needed -- the softmax+CE gradient is just probs minus the target one-hot). 100% integer Q16 => bit-exact deterministic, $0 on this laptop CPU (no GPU/CUDA/PyTorch). Trains on a tiny cyclic corpus, learns the transitions, and generates the pattern back. T1 P(correct next char) INCREASES across training (it learns the statistics). T2 GENERATION reproduces the learned pattern (from 'a' -> a,b,c,a,b,c). T3 (EXCEED) deterministic bit-identical. T4 argmax(W[prev]) == the correct next char for every char. expect_exit: 0 Sovereign: nx_syscalls.