nx_nofloat_mlp.nx

nx_nofloat_mlp.nx -- SOVEREIGN NO-FLOAT MLP + INTEGER BACKPROP (scale the trainer linear->multi-layer, toward an LLM). A 2-2-1 ReLU network in integer Q16 with full backprop (output layer + ReLU gradient + hidden layer), trained on XOR -- the canonical task a LINEAR model provably CANNOT solve, so success proves multi-layer integer backprop genuinely works. All integer => bit-exact deterministic training (the moat vs float/CUDA). T1 forward with the analytic ReLU-XOR weights outputs EXACT XOR [0,1,1,0] -- architecture + nonlinearity proven. T2 integer backprop from a distinct init REDUCES loss (the gradient mechanism works in fixed-point). T3 (EXCEED) DETERMINISM: two training runs -> BIT-IDENTICAL params. T4 substantial learning (final << initial). expect_exit: 0 Sovereign: nx_syscalls.