topic: nofloat

126 modules sharing the nofloat name family (derived from the tree's prefix discipline).

Integer-only machine learning: inference (and the training experiments) expressed without floating point at all — quantized arithmetic end to end, in keeping with the pure-integer core. This is the research track toward models the ecosystem can run and eventually train entirely on its own terms; the f32 family is the pragmatic counterpart for imported weights.

narrated overview -- maintained by the narration lane, module links verified against this wiki.

module	description	lines	funcs
nx_nofloat_adamw32_gate.nx	the REAL win from the R4 scale-up work: a numerically STABLE fixed-point Adam.	78	4
nx_nofloat_adamw_gate.nx	HARD-EVIDENCE gate for the Q16 AdamW optimizer (nfa_adamw), with a MEASURED	130	4
nx_nofloat_adversary.nx	CAP-NF-ADVERSARY: a DURABLE, sovereign adversarial reviewer. Nishi OWNS adversarial	141	9
nx_nofloat_arch.nx	ARCH-CONFIGURABLE inference: read a model's dims from GGUF METADATA instead of	74	5
nx_nofloat_arch_config_gate.nx	proves ARCH-CONFIGURABLE inference: read the model's dims from GGUF	96	4
nx_nofloat_arch_moe_gate.nx	gates the arch-config debt eaten 2026-07-15: EXACT attention scale +	123	6
nx_nofloat_attention_gate.nx	ASSEMBLE the trio into one DETERMINISTIC transformer attention layer.	164	11
nx_nofloat_attn.nx	SOVEREIGN NO-FLOAT ATTENTION (the transformer HEART), integer Q16, deterministic. The hard	80	7
nx_nofloat_attn_gate.nx	HARD-EVIDENCE gate for the ATTENTION-CORE backward (CAP-NF-ATTN-CORE): a single-head	244	13
nx_nofloat_audit.nx	CAP-NF-AUDIT (REBUILT after a 2nd adversarial review called the v1 tautological).	74	9
nx_nofloat_autograd.nx	NO-FLOAT (integer Q16 fixed-point) reverse-mode TENSOR autograd: the genuine	774	38
nx_nofloat_autograd_ddp_gate.nx	DEEPER R4->LLM: bit-reproducible data-parallel training driven by the	170	8
nx_nofloat_autograd_gate.nx	HARD-EVIDENCE gate for the NO-FLOAT (integer Q16) general autograd	221	12
nx_nofloat_batched_forward_gate.nx	THE DECISIVE SPECULATIVE-DECODING FOUNDATION EXPERIMENT.	140	6
nx_nofloat_bench.nx	CAP-NF-BENCH: HONEST measured benchmark of our no-float Q16 stack vs the Unsloth-class	97	11
nx_nofloat_block_gate.nx	HARD-EVIDENCE gate for a COMPLETE pre-norm TRANSFORMER BLOCK backprop + train,	295	11
nx_nofloat_blockfloat_gate.nx	BLOCK-FLOAT quant rung: per-block power-of-2-scaled integer = MXFP8's	118	7
nx_nofloat_blockfloat_gemm_gate.nx	BLOCK-FLOAT weights through the no-float GEMM: the format earns its place	141	9
nx_nofloat_blockfloat_layer_gate.nx	BLOCK-FLOAT weights through a real LINEAR LAYER (W.x + bias -> ReLU),	147	10
nx_nofloat_bpe_gate.nx	R2 (efficiency) from the sovereign-researcher roadmap	84	3
nx_nofloat_bpe_lm_gate.nx	CAPSTONE: the WHOLE sovereign efficient text pipeline wired END-TO-END, pure	149	8
nx_nofloat_census.nx	CAP-NF-CENSUS: a LIVING, liar-killed, negative-controlled census of our S-class	87	5
nx_nofloat_corpus_gate.nx	HARD-EVIDENCE: a CHARACTER-LEVEL language model trains on REAL TEXT, pure	210	8
nx_nofloat_deepstack_gate.nx	HARD-EVIDENCE gate that the no-float transformer scales in DEPTH: a	215	10
nx_nofloat_distill_gate.nx	R3 (cheap quality) from the sovereign-researcher roadmap	91	4
nx_nofloat_exceed_gate.nx	the UNIQUE Nishi exceed axis: NO-FLOAT determinism.	68	3
nx_nofloat_gemm.nx	canonical sovereign INTEGER SIMD GEMM (vpmaddwd / __i16x16_madd):	41	4
nx_nofloat_gemm_gate.nx	the NO-FLOAT compute lever: integer SIMD GEMM (vpmaddwd), the deterministic	103	3
nx_nofloat_generalize_gate.nx	HELD-OUT RULE GENERALIZATION in pure integer Q16 (CAP-NF-GENERALIZE). The	115	9
nx_nofloat_generate_gate.nx	AUTOREGRESSIVE TEXT GENERATION from a trained no-float LM, integer Q16	88	8
nx_nofloat_grammar_gate.nx	CAP-NF-GRAMMAR: the AFFORDABLE "landing" step the sovereign researcher	150	11
nx_nofloat_hw_research_fetch.nx	SOVEREIGN researcher: GROUND the HARDWARE-RUNG-UP exceed in REAL fetched	71	5
nx_nofloat_hwladder.nx	CAP-NF-HWLADDER: the HARDWARE-RUNG-UP ladder, silicon -> AI, each rung grounded on a	62	5
nx_nofloat_incontext_gate.nx	IN-CONTEXT COPYING (CAP-NF-INCONTEXT): the basis of in-context learning.	136	8
nx_nofloat_induction_gate.nx	CONTENT-BASED INDUCTION (CAP-NF-INDUCTION), REBUILT after the 3rd adversarial	134	6
nx_nofloat_k3block.nx	LOGICALLY-INTEGRATED K3-style transformer BLOCK in no-float (operator 2026-07-19	68	4
nx_nofloat_k3block_gate.nx	gate: the K3 primitives LOGICALLY INTEGRATE into one deterministic block.	75	2
nx_nofloat_k3interleave.nx	THE K3 BLOCK STRUCTURE, ASSEMBLED (operator 2026-07-19 "logically integrated,	90	2
nx_nofloat_k3interleave_gate.nx	gate: the 3:1 KDA:MLA interleaved K3 block structure, assembled from the	123	2
nx_nofloat_k3stack.nx	MULTI-LAYER K3 stack with ATTENTION RESIDUALS in no-float (operator 2026-07-19	97	5
nx_nofloat_k3stack_gate.nx	gate: N-layer K3 stack with Attention Residuals, deterministic AT DEPTH.	81	2
nx_nofloat_kda.nx	KIMI DELTA ATTENTION (KDA) in sovereign no-float = the K3 namesake attention	62	4
nx_nofloat_kda_gate.nx	gate: Kimi Delta Attention, the K3-distinctive behaviors (not just "it runs").	147	2
nx_nofloat_kvcache_gate.nx	CAP-NF-KVCACHE: a KV-CACHE for O(T) autoregressive decode (vs O(T^2)	145	12
nx_nofloat_kvcache_int8_gate.nx	CAP-NF-KVCACHE-INT8: a QUANTIZED KV-cache (store cached K/V at int8 =	129	15
nx_nofloat_ladder.nx	CAP-NF-LADDER: the CONSOLIDATION capstone. Grounds the whole no-float DeepMind ladder	69	4
nx_nofloat_landing_gate.nx	THE MEASURED LANDING (roadmap R4 increment; finding 7 = perplexity is the	119	8
nx_nofloat_layernorm_gate.nx	the THIRD no-float transformer primitive: integer rsqrt -> LayerNorm.	101	7
nx_nofloat_linattn.nx	SOVEREIGN no-float LINEAR ATTENTION recurrent kernel = the foundation of Kimi	52	4
nx_nofloat_linattn_gate.nx	gate for the no-float linear-attention recurrent kernel (KDA foundation).	100	2
nx_nofloat_llm.nx	CANONICAL sovereign no-float LLM library (ONE source of truth; no copy-paste debt).	1340	80
nx_nofloat_lm_gate.nx	THE CAPSTONE: a tiny TRANSFORMER LANGUAGE MODEL trains END-TO-END on next-token	295	12
nx_nofloat_mla.nx	MLA (Multi-head Latent Attention, DeepSeek's marquee KV-compression) in no-float	129	5
nx_nofloat_mla_gate.nx	MLA no-float SERVE (the census's #1 remaining July-2026 gap, 2026-07-15).	139	4
nx_nofloat_mlp.nx	SOVEREIGN NO-FLOAT MLP + INTEGER BACKPROP (scale the trainer linear->multi-layer, toward an	112	11
nx_nofloat_mlp4.nx	CLOSE THE GAP from nx_nofloat_mlp (the 2-2-1 hit the predict-~0.5 local min). A 2-4-1 ReLU	112	11
nx_nofloat_mm_i8_batch_gate.nx	STEP 1 of the speculative serve integration (operator-approved build):	129	11
nx_nofloat_moe.nx	no-float Mixture-of-Experts FFN (2026-07-15, MoE rung 1: the layer MECHANISM).	122	2
nx_nofloat_moe_block_gate.nx	gates the MoE transformer BLOCK (MoE rung 2, 2026-07-15): per-token	268	7
nx_nofloat_moe_gate.nx	gates the no-float MoE FFN layer (MoE rung 1, 2026-07-15). The lib does	285	6
nx_nofloat_moe_real_gate.nx	MoE rung 3 (2026-07-15): REAL MoE WEIGHTS, LAZILY. Loads a real sparse-MoE	311	6
nx_nofloat_multihead_gate.nx	HARD-EVIDENCE gate for MULTI-HEAD attention in pure integer Q16. The defining	252	9
nx_nofloat_multistep_gate.nx	WHY no-float wins the MULTI-STEP-LOGIC arena (operator thesis 2026-06-18).	62	5
nx_nofloat_muon.nx	the FULL Per-Head Muon optimizer LOOP in sovereign no-float (K3 F405; operator	50	4
nx_nofloat_muon_gate.nx	gate: the FULL Per-Head Muon optimizer loop (momentum + NS-orthogonalize +	151	4
nx_nofloat_muon_ns.nx	SOVEREIGN no-float Newton-Schulz orthogonalization = the core of the Muon	113	7
nx_nofloat_muon_ns_gate.nx	gate for integer Newton-Schulz orthogonalization (Muon core).	83	2
nx_nofloat_mxfp4.nx	SOVEREIGN no-float MXFP4 dequant (K3 F406 gap; operator 2026-07-19 "beyond SOTA").	50	3
nx_nofloat_mxfp4_gate.nx	gate for the no-float MXFP4 dequant (authored ON nx_gate_verdict).	59	1
nx_nofloat_mxfp8.nx	SOVEREIGN no-float MXFP8 dequant (K3 F406; operator 2026-07-19 "state of the art").	75	5
nx_nofloat_mxfp8_gate.nx	gate: MXFP8 (OCP E4M3 + E5M2) dequant to Q16, EXACT known values.	71	1
nx_nofloat_olmoe.nx	the faithful OLMoE transformer LAYER in deterministic no-float (MoE rung 4,	185	2
nx_nofloat_olmoe_forward_gate.nx	MoE rung 5 (2026-07-15): a COMPLETE OLMoE forward -- 16 stacked	294	6
nx_nofloat_olmoe_gate.nx	MoE rung 4 (2026-07-15): a FAITHFUL OLMoE transformer LAYER on REAL blk.0	221	7
nx_nofloat_prefix_cache_gate.nx	proves the PREFIX-KV-CACHE speed optimization is BIT-EXACT + measures the	122	3
nx_nofloat_propose_chat_gate.nx	PROPOSE->VERIFY rung 3 (2026-07-15): a BETTER PROPOSER on the SAME	205	1
nx_nofloat_propose_solve_gate.nx	PROPOSE->VERIFY rung 2 (2026-07-15): SOLVE-FORMS. Rung 1 proved the	209	1
nx_nofloat_propose_verify_gate.nx	THE PROPOSE->VERIFY LOOP, rung 1 (2026-07-15). The sovereign no-float	179	1
nx_nofloat_prose_gate.nx	CAP-NF-PROSE: the no-float char-LM trains on REAL ENGLISH PROSE (words + spaces,	146	6
nx_nofloat_q6k_gate.nx	sovereign Q6_K -> Q16 integer dequant (unlocks Qwen embeddings + LM head).	155	6
nx_nofloat_quant_gate.nx	CAP-NF-QUANT: post-training QUANTIZATION for efficient sovereign inference,	163	13
nx_nofloat_qwen_bias_probe.nx	recon: does Qwen2.5 carry Q/K/V projection biases (a known Qwen2 detail	46	5
nx_nofloat_qwen_config_gate.nx	rung 2a: read the REAL Qwen2.5-0.5B config from GGUF metadata.	94	6
nx_nofloat_qwen_decode_gate.nx	TOKENIZER decode: token id -> text, from the real GGUF vocab.	91	5
nx_nofloat_qwen_diag_gate.nx	FAITHFULNESS LOCALIZER (memory-safe: ONE no-float forward, lazy per-layer load).	130	5
nx_nofloat_qwen_dqprobe_gate.nx	MEASUREMENT PROBE (not a milestone gate): isolate how much of a decode step	99	4
nx_nofloat_qwen_encode_gate.nx	BPE ENCODE: text -> token ids, sovereign, from the real GGUF merges+vocab.	131	7
nx_nofloat_qwen_encode_prep_gate.nx	ENCODE foundation: the BPE merges + the reverse (string -> id) lookup.	88	6
nx_nofloat_qwen_fastgen_gate.nx	KV-cached generation with a DEQUANT-ONCE weight+head cache.	159	4
nx_nofloat_qwen_fastgen_i32_gate.nx	Stage 1 of the decode-speed arc: the dequant-once cache narrowed	156	4
nx_nofloat_qwen_fastgen_i8_gate.nx	Stage 2a of the decode-speed arc: W8A8 SIMD decode (__i16x16_madd,	161	4
nx_nofloat_qwen_fidelity_gate.nx	FIDELITY: the library's no-float Q16 matmul vs IEEE f32 on a REAL Qwen	83	5
nx_nofloat_qwen_forward_gate.nx	CAPSTONE: a COMPLETE sovereign no-float Qwen2.5-0.5B forward.	118	3
nx_nofloat_qwen_gen_gate.nx	GREEDY MULTI-TOKEN GENERATION through the sovereign no-float Qwen2.5-0.5B.	121	4
nx_nofloat_qwen_kvgen_gate.nx	KV-CACHED greedy generation through the sovereign no-float Qwen2.5-0.5B.	166	5
nx_nofloat_qwen_rope_h2h_gate.nx	ARCH H2H: the library's no-float Q16 RoPE vs the reference f32 RoPE.	78	4
nx_nofloat_qwen_stack_gate.nx	STACK real Qwen layers (lazy per-layer load), deterministic no-float.	83	3
nx_nofloat_qwen_text_gate.nx	GRAND FINALE: text -> text through the sovereign no-float Qwen2.5-0.5B.	122	4
nx_nofloat_qwen_tok_gate.nx	TOKENIZER CHECK (fast, no forward): BPE-encode a multi-word prompt and print the	52	5
nx_nofloat_qwen_wload_gate.nx	exercises the canonical library's by-name Q16 loader over the real model's	81	4
nx_nofloat_rope_gate.nx	HARD-EVIDENCE gate for RoPE (rotary position embedding) backward + RoPE composed	186	9
nx_nofloat_sample_gate.nx	TEMPERATURE SAMPLING decode in pure integer Q16 (CAP-NF-SAMPLE). Real LLMs	113	8
nx_nofloat_scale2_gate.nx	R4 DEPTH test: does a 2-BLOCK no-float LM reach the richer-grammar floor where	108	9
nx_nofloat_scale_batch_gate.nx	R4 retry via the ONE untested fundamental lever: MINI-BATCH GRADIENT	116	7
nx_nofloat_scale_ffn_gate.nx	R4 FINAL test: the COMPLETE transformer block (attention + FFN) on the richer	122	7
nx_nofloat_scale_gate.nx	HONEST documented attempt mapping the frontier. On the richer	182	9
nx_nofloat_scale_research_fetch.nx	SOVEREIGN researcher: GROUND the "how do we AFFORDABLY + EFFICIENTLY	88	5
nx_nofloat_scale_wide_gate.nx	R4 the genuine SCALE test: real WIDTH scale-up. Prior R4 tests stayed at	121	7
nx_nofloat_serialize_gate.nx	CAP-NF-SERIALIZE: sovereign model PERSISTENCE (save/load trained weights).	84	7
nx_nofloat_serve.nx	the DEPLOYABLE no-float LLM daemon (ops shell around the gated pure core	153	6
nx_nofloat_serve_core.nx	the PURE CORE of the no-float LLM serve organ (2026-07-10): model session +	1244	38
nx_nofloat_serve_gate.nx	gates the no-float LLM serve CORE in-process (pure-core+gate+daemon idiom:	157	7
nx_nofloat_serve_live_gate.nx	LIVE socket proof for the no-float LLM daemon: connects to	95	7
nx_nofloat_simd_dot_probe_gate.nx	DECISIVE probe for Stage 2 (SIMD decode). Before rebuilding the whole	138	7
nx_nofloat_softmax_gate.nx	a SYSTEM for the HARD no-float primitive: fixed-point exp + softmax.	107	6
nx_nofloat_speed_bench.nx	honest tok/s baseline for the no-float Qwen serve (the inference-speed arc's	64	6
nx_nofloat_tok.nx	canonical sovereign BPE tokenizer library for the GGUF vocab (ONE source of truth).	182	10
nx_nofloat_tokdec.nx	byte-level-BPE piece decoder (GPT-2/Qwen convention): GGUF vocab pieces are UTF-8	53	2
nx_nofloat_train.nx	SOVEREIGN NO-FLOAT TRAINING CORE (ledger #1, the biggest gap). Integer Q16 fixed-point	83	7
nx_nofloat_train_exceed.nx	the MEASURED S-class exceed for no-float TRAINING. Extends the established	123	5
nx_nofloat_transformer.nx	SOVEREIGN NO-FLOAT TRANSFORMER BLOCK (the AI culmination), pure integer Q16,	118	13
nx_nofloat_vs_float_gate.nx	CROSS-VALIDATION: do we need the float equivalent? Answer: it already EXISTS	120	9
nx_nofloat_w4_gemm_gate.nx	THE DECISIVE W4A16 EXPERIMENT (de-risks the inference-speed arc before any	161	7
nx_nofloat_weightconv_gate.nx	the BRIDGE: load REAL BF16 model weights -> no-float Q16 integer.	100	4
nx_nofloat_weighttie_gate.nx	CAP-NF-WEIGHTTIE: tied input-embedding / output-projection (GPT-style	110	8
nx_nofloat_xformer_gate.nx	HARD-EVIDENCE gate for the TRANSFORMER-SUBLAYER backward ops added to the	220	11