nx_nofloat_bench.nx

nx_nofloat_bench.nx -- CAP-NF-BENCH: HONEST measured benchmark of our no-float Q16 stack vs the Unsloth-class incumbents (efficient LLM training/inference: Unsloth, QLoRA, GPTQ, AWQ, LLM.int8/bitsandbytes, llama.cpp). Competitor claims are GROUNDED in the sovereign-researcher corpus (knowledge/fetched/eff_*.raw, 12/12 live): Unsloth 2-5x faster / 50-80% less VRAM (GPU+PyTorch+Triton); QLoRA 4-bit / 65B on 48GB GPU; GPTQ 3-4 bit "negligible" loss (GPU); LLM.int8 = 8-bit; float arithmetic is NON-ASSOCIATIVE (eff_float.raw). We report a per-axis EXCEED/PARITY/BEHIND scorecard, each axis tied to a REAL organ result (liar-kill), and we MEASURE the headline exceed (determinism) concretely. NO-WAVE: we EXCEED on the design-goal axes and are honestly BEHIND on scale/throughput/ecosystem (a different design point) -- not a floor-claim. T1 MEASURED determinism exceed: our integer reduction is ORDER-INVARIANT (fwd==rev, diff=0) while a float-sim reduction is ORDER-DEPENDENT (diff>0) -> bit-exact reproducibility the float incumbents cannot guarantee. T2 HONESTY (anti-wave): the scorecard marks >=2 axes BEHIND (we do NOT claim exceed everywhere). T3 GROUNDED: every EXCEED axis is backed by a real on-disk organ (determinism/quant/sovereign-base exist). expect_exit: 0 Sovereign: nx_syscalls only (this benchmark itself uses ZERO float ops -- that IS axis 2).