nx_llm_ppl_bench_gate.nx

nx_llm_ppl_bench_gate.nx -- THE QUALITY BENCHMARK: perplexity + next-token accuracy of ANY gguf model on a held-out text, via our sovereign arch-config no-float inference. This is the "compare our quality" instrument the operator asked for -- run it on the real Qwen (track a's family), on OUR trained+exported model (track b), or on a PyTorch-fine-tuned gguf; the NUMBER is comparable across all of them. Perplexity is THE standard LLM metric (lower=better). ★NUMERICS: our integer head logit = realdot*65536 (Q24 normed . Q16 weight, >>24), so logit/65536 = the TRUE real logit -> we compute ABSOLUTE, PyTorch-comparable perplexity by doing the logsumexp reduction in f32 (a tiny standard computation) on top of the sovereign integer forward. CE_t = logsumexp_v(logit[t][v]) - logit[t][tgt_{t+1}] (real nats) perplexity = exp(mean_t CE_t) ; accuracy = fraction of positions whose argmax == the true next token. expect_exit: 0 license_tier: ORIGINAL