nx_nofloat_corpus_gate.nx

nx_nofloat_corpus_gate.nx -- HARD-EVIDENCE: a CHARACTER-LEVEL language model trains on REAL TEXT, pure integer Q16 (CAP-NF-REALCORPUS). Off the toy cyclic task: the corpus is a real sentence, the vocab is its distinct characters, and the space char recurs before different words -- so the model MUST use causal-attention HISTORY to disambiguate (a genuine context task, not a context-free lookup). Trains ALL weights with AdamW. T1 started untrained (CE > 0) ; T2 CE drops >= 70% (it learned the text) T3 next-char accuracy >= 90% on the corpus (argmax predicts the real next char from history) T4 bit-exact (train twice -> identical integer weights) + prints the model's predicted text vs the target for transparency (you can SEE it learned). Evidence -> knowledge/status/nofloat_corpus.log. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0