nx_nofloat_quant_gate.nx

nx_nofloat_quant_gate.nx -- CAP-NF-QUANT: post-training QUANTIZATION for efficient sovereign inference, grounded in the BitNet research (nfs_bitnet158.raw / nfs_intquant.raw: ternary {-1,0,1} and integer-only inference match full precision). Train the near-optimal 3-category grammar LM (Q16), then quantize its weights to INT8 (per-tensor symmetric, scale=maxabs/127) and TERNARY {-1,0,1} (scale=mean|W|), and measure how much held-out quality (CE) survives. T1 INT8 preserves quality: int8 held-out CE ~= full-precision CE (within ~0.15 nats). T2 INT8 stays near-OPTIMAL: int8 CE << uniform ln(8)=2079 (efficient 8-bit inference works). T3 TERNARY carries the signal: ternary CE < uniform (1.58-bit weights still model the grammar; honest: ternary loses more than int8 on a tiny model -- the BitNet "matches at scale" claim is a SCALE result). Pure integer Q16. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0