Nishi vs the Field -- Sovereign Model Lifecycle (Modelwright)

Sovereign Model Lifecycle (Modelwright) — mechanically measured, liar-killed, sovereign.

Capability	Nishi	PyTorch	HuggingFace	llama.cpp	Megatron-LM	Notes
Reverse-mode autodiff training substrate	Yes	Best	Yes	No	Yes	Tensor reverse-mode autograd; the first sovereign LM trained from scratch beat the analytic midpoint bit-exact (role scorecard MODELWRIGHT, 2026-06-10); PyTorch autograd is the bar
Transformer forward (100 percent integer no-float)	Yes	Yes	Yes	Yes	Yes	The core forward every model has, done with zero floating point
BPE tokenizer training	Yes	Yes	Yes	Yes	Yes	Own byte-pair tokenizer + vocab training
Integer K-quant weights	Yes	Yes	Yes	Best	No	Q4-K dot kernels; llama.cpp K-quants are the CPU bar; TRT/Megatron are GPU-float
SIMD integer GEMM	Yes	Best	Yes	Best	Yes	AVX2 integer GEMM (7x pooled); the field's GEMM is float/cuBLAS
Open-weights GGUF loading	Yes	Best	Best	Best	Yes	Real Qwen2.5-0.5B loads and greedy-matches the float reference 4-for-4 (2026-07-09); llama.cpp is the GGUF home
Served chat API (OpenAI-shaped)	Yes	Yes	Best	Best	No	Sovereign seat on :11434 serving real Qwen; vLLM/llama.cpp lead production serving
Attention and layernorm forward	Yes	Yes	Yes	Yes	Yes	Score-matrix attention + softmax; the transformer core
Large-scale pretraining (billions of params, trillions of tokens)	No	Best	Yes	Yes	Part	Megatron/PyTorch FSDP pretrain frontier models; ours is a toy LM
GPU / CUDA training	No	Best	No	Yes	Part	CUDA/ROCm training is the bar; Nishi trains CPU-integer by design today
Distributed tensor-parallel training	No	Yes	No	No	Best	Megatron-LM 3D parallelism is the bar; Nishi single-process
Instruction fine-tuning (SFT) pipeline	No	Yes	Part	No	Yes	HuggingFace TRL SFTTrainer is the bar; no SFT loop here
RLHF / DPO preference alignment	No	Yes	Part	No	Yes	TRL DPO/PPO alignment is the bar; the models run un-aligned
LoRA / parameter-efficient fine-tuning	No	Yes	Part	Yes	Yes	PEFT/LoRA adapters are the bar; absent here
Standardized eval harness (MMLU / HumanEval)	No	Yes	Best	Yes	Yes	lm-eval-harness is the field standard; ours is qabench (referee census) not a model-eval harness
Flash-attention / fused training kernels	No	Best	Yes	Yes	Part	FlashAttention fused kernels are the bar; ours is a plain score matrix
Mixture-of-experts training	No	Yes	Yes	Yes	Best	Sparse MoE training (Megatron/DeepSpeed); Nishi dense-only
Competitive train and inference speed	No	Best	Yes	Best	Best	Measured 2.4 tok/s CPU decode = 8-25x behind llama.cpp-class; GPU stacks train orders faster
Model hub and distribution	No	Yes	Best	Yes	No	HuggingFace Hub is the bar; our models live on the NAS, not a public hub
100 percent integer no-float train and inference (bit-exact reproducible by construction)	Best	No	No	No	No	A 100 percent integer LLM greedy-matches the float reference 4-for-4 and reproduces bit-identically on any CPU; the entire field is float and nondeterministic across runs/hardware
Sovereign whole training-and-serving loop (own language, compiler, kernels; zero PyTorch/CUDA/BLAS)	Best	No	No	No	No	Train + quantize + serve entirely on our own stack bits-up; PyTorch/Megatron sit on CUDA + cuBLAS, llama.cpp on foreign C++ toolchain -- our brain runs on OUR substrate, the literal own-the-loop

Honest verdict. The Modelwright owns Nishi's own brain -- and the honest headline is that it is REAL but TOY-scale. It trains from scratch (a reverse-mode autodiff substrate; the first sovereign LM beat the analytic midpoint bit-exact, 2026-06-10), it runs a 100 percent integer no-float transformer that greedy-matches the float reference 4-for-4 (2026-07-09), it loads real open weights (Qwen2.5-0.5B via GGUF), tokenizes with its own BPE, quantizes to integer K-quants, and serves an OpenAI-shaped chat API -- the whole loop is ours, our brain not Claude's. But at production scale it is not close: no large-scale pretraining, no GPU or CUDA training, no distributed tensor-parallelism (Megatron), no instruction fine-tuning or RLHF/DPO alignment (TRL), no LoRA, no standardized eval harness, and measured CPU decode is 8-25x behind llama.cpp-class. The two genuine exceeds are structural, not scale: 100 percent integer training and inference is bit-exact reproducible BY CONSTRUCTION where the entire field is float and nondeterministic, and the whole stack is ours bits-up -- own language, compiler, kernels, zero PyTorch/CUDA/BLAS beneath. The climb is the deepest rung in the ledger: train a real model at real scale on our own substrate, then alignment and an eval harness (which the referee shares).

Generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/modelwright.matrix — every Nishi cell verified against real organ source on disk. Zero JS, zero trackers.