Update unified Origami Preview model card, wirkung.cls figures, and Family Paper PDF

.gitattributes

@@ -639,3 +639,6 @@ vision/layer-026/self_attn__o_proj.o3f filter=lfs diff=lfs merge=lfs -text
 vision/layer-026/self_attn__q_proj.o3f filter=lfs diff=lfs merge=lfs -text
 vision/layer-026/self_attn__v_proj.o3f filter=lfs diff=lfs merge=lfs -text
 vocab/embed_tokens.o3f filter=lfs diff=lfs merge=lfs -text
+paper/Foss_2026_Origami_HCGM.pdf filter=lfs diff=lfs merge=lfs -text
+assets/architecture.png filter=lfs diff=lfs merge=lfs -text
+assets/mrail.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -2,152 +2,144 @@
 license: other
 license_name: wirkung-model-and-runtime-1.0
 license_link: LICENSE
-library_name: wirkung
-pipeline_tag: text-generation
 base_model:
-- google/gemma-4-31B-it
+- google/gemma-4-31b-it
 base_model_relation: quantized
+library_name: wirkung
+pipeline_tag: text-generation
+num_parameters: 32680000000
 language:
 - en
 - de
-- zh
-- fr
-- es
-- ja
-- ko
 tags:
 - wirkung
 - origami
 - preview
+- training-free
+- lossless-folding
+- constant-memory
+- zero-vram
+- cpu-inference
+- npu
 - hcgm
 - m-rail
-- sgpu
-- hmoe
-- constant-state
-- cpu-inference
-- zero-gpu
+- 31b
 - multimodal
-- gemma4
+- gemma-4
+- vision-language
 ---
 
-![Multimodal symmetric-attention Holographic Causal Graph Machine built for zero-copy disk execution and bounded long-context state.](assets/banner.png)
+![Gemma 4 31B folded without retraining from 62.55 GB to 11.56 GB (5.41x), with a constant state at any context length.](assets/banner.png)
+
+[Paper](paper/Foss_2026_Origami_HCGM.pdf) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Wirkung](https://wirku.ng) · [Base weights](https://huggingface.co/google/gemma-4-31b-it) · [Contact](mailto:contact@wirku.ng) · [License](LICENSE)
 
-[Paper](https://doi.org/10.13140/RG.2.2.28206.68167) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Website](https://wirku.ng) · [Base weights](https://huggingface.co/google/gemma-4-31B-it) · [License](https://huggingface.co/wirkung/Origami-31B/blob/main/LICENSE)
+**`Origami-31B` is a 32.68B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from `google/gemma-4-31b-it` with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 62.55 GB source checkpoint into 11.56 GB (5.41x compression), executes directly from NVMe/SSD storage via OS `mmap` with `0 GB` GPU VRAM and `0 GB` dedicated weight-RAM, and keeps live state strictly constant ($O(1)$) across arbitrary horizons on CPU, NPU, `sGPU`, Metal, and standalone PTX.**
 
-**Origami-31B (Preview)** is a data-free multimodal Holographic Causal Graph Machine (HCGM) compiled purely in closed form from the weights of [`google/gemma-4-31B-it`](https://huggingface.co/google/gemma-4-31B-it) with zero gradient updates, zero fine-tuning steps, and zero calibration samples. Folding interleaved sliding-window layers, symmetric (`K = V`) global attention layers, the multimodal vision tower, and the tied wide-vocabulary readout into self-describing `.o3f` operators, Origami-31B executes straight from NVMe via OS `mmap` under [`wirkung`](https://pypi.org/project/wirkung/) with no GPU requirement and strictly bounded `O(1)` session state.
+`Origami-31B` preserves the full `google/gemma-4-31b-it` architecture (50 sliding-window attention layers, 10 symmetric $K = V$ global-attention layers, a 27-layer SigLIP vision tower, and a tied 262,144-entry softcapped vocabulary) while turning a passive weight checkpoint into a self-contained binary file executed under one cross-silicon contract. It succeeds our first-generation constant-state releases published under [`huggingface.co/tfwnotops`](https://huggingface.co/tfwnotops) by removing `vLLM`, `MLX`, and `Transformers` runtime dependencies altogether. This repository is a **Preview** release: structural byte footprints, fold ratios, and constant-memory invariants below are fixed properties of the compiled artifact, while hardware-dependent throughput figures are omitted from the card and documented in the [Origami Family Paper](paper/Foss_2026_Origami_HCGM.pdf).
 
 ```bash
 pip install -U wirkung && wkng pull wirkung/Origami-31B -o ./Origami-31B
-wkng run ./Origami-31B -p "Explain symmetric K=V Ward carriers." --chat
+wkng run ./Origami-31B -p "Explain why monotone logit softcapping preserves exact Cauchy-Schwarz row pruning." -n 128
 ```
 
-## Architectural Overview: Upstream vs. Origami
+## Source checkpoint against Origami-31B
 
-| Architectural Dimension | Conventional Checkpoint Serving | `Origami-31B` (Wirkung HCGM Preview) | Structural Invariant |
-|---|---|---|---|
-| Compilation & Calibration | Pretrained multimodal checkpoint | **Data-free closed-form algebraic fold** | Zero training & zero calibration corpus |
-| Execution Substrate | High-VRAM GPU + PyTorch / vLLM | **Self-describing `mmap` operator file** | Zero GPU VRAM required |
-| Weight Residency | Deserialized into resident VRAM / RAM | **Demand-paged directly from SSD / NVMe** | Instant cold startup without RAM preload |
-| Global Attention State | Linearly expanding `K = V` global cache | **Symmetric (`K = V`) pinned-sink Ward state** | Constant `O(1)` long-context memory |
-| Local Attention State | Sliding-window tensor ring buffers | **Bounded fixed-horizon local carriers** | Strictly bounded total session state |
-| Tied Wide Vocabulary | Full dense tied-embedding projection | **Certified Cauchy-Schwarz norm gate** | Exact greedy argmax preservation |
-| Speculative Verification | Autoregressive token-by-token passes | **Single-pass `DraftRail` span verification** | Deterministic across CPU server ISAs |
-| Fleet Reuse | Isolated per-request recomputation | **Shared `M-Rail` (`X = X_known + R`)** | Associative fleet-wide residual closure |
+| Property | Source (`google/gemma-4-31b-it`) | `Origami-31B` (HCGM Preview) |
+|:--|--:|--:|
+| Parameters | 32.68B | **32.68B** (identical graph topology) |
+| Retraining / fine-tuning steps | Pretrained baseline | **0** (closed-form compilation) |
+| Checkpoint footprint on disk | 62.55 GB BF16 | **11.56 GB** (**5.41x** smaller; 601 `.o3f` text, vision, and vocabulary files + `exact_core.safetensors`) |
+| Required GPU VRAM | Multi-tens of GB HBM | **0 GB** (runs on plain CPU / NPU / `sGPU` or optional GPU) |
+| Dedicated weight-RAM | Pinned in host / device RAM | **0 GB** (read-only OS `mmap` directly from SSD/NVMe) |
+| Live session / horizon state | Linearly growing $O(T)$ global KV cache across 10 full-attention layers | **Constant $O(1)$ bounded sliding-window + symmetric $K = V$ carrier at any length** |
+| Operator & readout path | Dense per-step matrix scan | Certified Cauchy-Schwarz gate under $\kappa \tanh(z/\kappa)$ logit softcapping (median 2.0 of 262,144 rows, 0 drift) |
+| Runtime dependencies | PyTorch / `vLLM` / `MLX` / CUDA | **`wirkung` + `numpy`** (standalone C11, NPU, `sGPU`, Metal, PTX) |
 
-![Disk-mapped execution and constant session state](assets/weights_vs_origami.png)
-<sub><b>Figure 1.</b> Execution substrate and live session memory trajectory. Origami-31B maps compiled `.o3f` operators directly from NVMe without GPU VRAM and replaces linearly expanding global caches with a strictly constant `O(1)` carrier state.</sub>
+![Physical byte anatomy and sequence-horizon state scaling of Origami-31B](assets/architecture.png)
 
-## Signature HCGM Architecture & Distinguishing Capabilities
+Figure 1. (a) Physical byte scale contrasting the 62.55 GB BF16 `google/gemma-4-31b-it` checkpoint against the 11.56 GB `Origami-31B` `.o3f` bundle mapped directly from SSD (`0 GB` GPU VRAM, `0 GB` dedicated weight-RAM). (b) Live conversation state across 4k to 1M tokens: linear $O(T)$ KV growth against the bounded $O(1)$ `Origami-31B` carrier.
 
-- **Calibration-Free Closed-Form Compilation**: Every decoder projection, vision encoder block, and tied embedding operator is compiled directly from the pretrained weight matrices via closed-form spectral subspace decomposition without touching calibration text, synthetic prompts, or gradient updates.
-- **The File Is the Machine (`mmap` Out-of-Core Execution)**: Stored as immutable, self-describing `.o3f` operators on SSD/NVMe, Origami-31B starts immediately via OS virtual-memory mapping and runs cleanly on both CPU-only server nodes and compact unified-memory machines where dense weights exceed physical RAM.
-- **Interleaved Local & Symmetric (`K = V`) Global Constant State**: Bounded local sliding-window carriers pair with symmetric key-value global Ward carriers so that global document context remains accessible inside a fixed `O(1)` memory envelope over arbitrarily long sessions.
-- **Certified Tied Vocabulary Gate & `DraftRail` Amortization**: Row-norm certificates over the tied vocabulary head prune dominated token candidates without altering greedy argmax outputs, while `DraftRail` verifies speculative multi-token drafts in a single amortized pass.
+## Zero-retraining compilation and constant-state architecture
 
-![HCGM organ and operator architecture](assets/architecture_organs.png)
-<sub><b>Figure 2.</b> Three-stage full-width HCGM topology of Origami-31B: data-free multimodal vision-and-text `.o3f` intake via OS `mmap`, interleaved local and symmetric (`K = V`) global constant-state core, and certified tied-vocabulary gating with `DraftRail` span verification.</sub>
+Every Origami model is produced by the same four-part compilation pipeline without gradient descent:
 
-## Shared M-Rail (`X = X_known + R`), sGPU Fleet Economics & HMoE
+1. **The file is the machine (disk-resident `mmap` execution).** An HCGM model is an immutable binary artifact on disk. The operating system maps operator pages read-only on demand from SSD or NVMe storage, requiring neither GPU VRAM nor weight-deserialization buffers in host RAM. Under our `v3-exakt` dyadic integer reduction contract, CPU (`AVX2` / `NEON`), general NPUs, embedded `sGPU` (`cpu_cuda`), Apple Metal, and driver-API PTX evaluate identical canonical reductions rather than drifting floating-point kernel approximations.
+2. **Closed-form multimodal folding (`.o3f`, 5.41x without retraining).** Compiled directly from the pure BF16 62.55 GB `google/gemma-4-31b-it` repository (61.40 GB active text and vision weights) with zero gradient steps, `Origami-31B` packs all 60 decoder layers, the 27-layer SigLIP vision encoder, and the tied 262,144-row head into 11.56 GB across 601 `.o3f` operator files via Rademacher-scrambled `FWHT_128` bitplanes and rank-$r$ projective crystals ($U_r V_r^\top$).
+3. **Constant $O(1)$ symmetric $K = V$ carrier synthesized via `Wirkung`.** Through `Wirkung`, our internal exact structural-law and closed-form operator synthesis engine, the 50 local sliding-window layers pair with a constant-capacity pinned-sink carrier on the 10 global symmetric $K = V$ attention layers, holding live memory strictly independent of sequence length $T$ while preserving exact argmax ordering across monotonically softcapped $\psi(z) = 30 \tanh(z/30)$ vocabulary logits.
+4. **Certified readout and online trajectory decomposition ($X = X_{\mathrm{known}} + R$).** Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure $X_{\mathrm{known}}$ (cosine trajectory banks, multi-order suffix tables, and rank-1 Sherman-Morrison-Woodbury updates) plus novel residual $R$, verifying multiple candidate steps in one weight pass with zero greedy output drift.
 
-![Shared M-Rail, sGPU fleet, and HMoE topology](assets/mrail_network.png)
-<sub><b>Figure 3.</b> Shared `M-Rail` fleet architecture. Known spans replay from associative stations (`R = 0`), while newly solved residuals (`R != 0`) propagate across CPU servers and `HMoE` peers without greedy drift.</sub>
+![Constant-memory read geometry and folded operator execution in Origami-31B](assets/mechanism.png)
 
-Origami-31B operates natively over the **Shared `M-Rail`** trajectory fabric and bundled **`sGPU`** (`sgpu`, `cpuda`, `rail`) deterministic CPU execution engine:
+Figure 2. (a) Constant-memory attention read over token coordinate $j$: pinned sink $s_0$, bounded symmetric $K = V$ far-history carrier, and exact sliding-window frontier. (b) Single-pass `.o3f` contraction with 4-class Walsh-Hadamard bitplanes, rank-$r$ projective crystal ($U_r V_r^\top$), and softcap-invariant Cauchy-Schwarz pruning.
 
-1. **Warm Up Once, Serve Indefinitely on Disk-Backed CPU Fleets**: Under `X = X_known + R`, enterprises can warm up `.mrail` trajectory maps using a single owned GPU or temporary burst lease, shut GPUs down once recurring document and agentic workflows are crystallized, and run production workloads permanently on NVMe-backed CPU VMs.
-2. **Collective Associative Residual Closure**: Whenever any node in the fleet evaluates an unseen residual branch (`R != 0`), it commits verified `ANCHOR` / `DELTA` stations and `OP_COPY` morphisms to Shared `M-Rail` (`wkng maps merge`), resolving that branch fleet-wide with zero greedy output drift.
-3. **Heavy Mixture of Experts (`HMoE`) Specialist**: On the synchronous `M-Rail` `cstar` bus, Origami-31B serves as a multimodal, multilingual symmetric-attention specialist alongside `Origami-9B`, `Origami-27B`, `Origami-78B`, and `Origami-H3`.
+## Shared trajectory assimilation (M-Rail), GPU-less server fleets and multi-model routing (HMoE)
 
-## Runtime CLI & Python SDK
+Beyond single-node execution, `Origami-31B` attaches natively to **Shared `M-Rail`** (`.mrail` v3), a content-addressed associative trajectory manifold shared across processes or across an enterprise network:
 
-Install the standalone native runtime from PyPI (`Python >= 3.10`, `numpy` as the sole Python dependency; zero PyTorch, zero CUDA toolkit, zero vLLM, zero `transformers`):
+- **Warm up once on one GPU, run permanently on disk-only CPU/NPU servers.** Because HCGM models require only SSD storage and megabytes of live state, an organization can run a single GPU (or a short temporary cloud GPU lease) until its recurring domain workflows are crystallized into a shared `.mrail` file, then shut the GPU down and serve frontier models across commodity CPU/NPU VMs and CPU-only servers. Known trajectories execute via `Spur` copy morphisms or resume in $O(1)$ space from `Station` (`ANCHOR` + `DELTA`) snapshots; whenever any server encounters an unseen task ($R \neq 0$), computing and committing that residual once makes the solution immediately available to every other node on the rail.
+- **Heavy Mixture of Experts (`HMoE`).** Because idle disk-mapped models consume zero weight-RAM, `M-Rail` routes across complete compiled Origami models (`Origami-9B`, `Origami-27B`, `Origami-31B`, `Origami-78B`, `Origami-H3`), zero-recompile static-graph NPU specialist pools, and sub-2 KB exact algebraic group organs on a synchronous station bus (`cstar`) under deterministic check gates.
 
-```bash
-# Inspect host SIMD threads, sGPU driver status, and operator bundle integrity
-wkng doctor
-sgpu doctor
-wkng verify ./Origami-31B
+![Shared M-Rail GPU-less server fleet topology and multi-session memory scaling for Origami-31B](assets/mrail.png)
+
+Figure 3. (a) Shared `M-Rail` (`.mrail` v3) fleet architecture: optional temporary GPU warmup feeds a permanent zero-GPU CPU/NPU/`sGPU` server fleet where any novel residual $+R$ solved once commits back to all nodes alongside `HMoE` routing on `cstar`. (b) Total resident RAM/VRAM across $N = 1..32$ concurrent 32k-token sessions: pinned weights plus $N \cdot \mathrm{KV}$ against shared read-only `mmap` plus $N \cdot O(1)$ constant state.
+
+## How to run
 
-# Launch an OpenAI-compatible HTTP server or merge shared M-Rail maps
+Requires Python 3.10+ and the `wirkung` runtime package (`pip install -U wirkung`), which installs the `wkng`, `wirkung`, `sgpu`, and `rail` CLI entrypoints.
+
+```bash
+pip install -U wirkung
+wkng pull wirkung/Origami-31B -o ./Origami-31B
+wkng verify ./Origami-31B --full
+wkng run ./Origami-31B --chat
 wkng serve ./Origami-31B --host 127.0.0.1 --port 8080
-wkng maps merge ./peer_node_maps --into ~/.wkng/maps
 ```
 
 ```python
 from wirkung import Engine
 
-engine = Engine("./Origami-31B")
-response = engine.generate(
-    "Compare calibration-free algebraic folding against post-training quantization.",
-    chat=True,
-)
-print(response)
+eng = Engine("./Origami-31B")
+print(eng.generate("Compare linear O(T) KV caches against fixed-capacity symmetric K=V carriers:", max_new_tokens=128))
 ```
 
-## Repository Contents
-
-- `bundle.json`, `config.json`, `generation_config.json`, `processor_config.json`: HCGM bundle manifest, multimodal processor definitions, and symmetric-attention architecture metadata.
-- `layers/`: compiled `.o3f` spectral operators across all interleaved sliding-window and symmetric `K = V` global decoder layers.
-- `vision/`: compiled `.o3f` operators for the multimodal vision encoder and cross-modal projector.
-- `vocab/`: folded tied-vocabulary readout operators with certified Cauchy-Schwarz row-norm bounds.
-- `exact_core.safetensors`: exact normalization parameters and attention carrier anchors.
-- `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`: standalone wide-vocabulary tokenizer and chat template.
-- `LICENSE`, `NOTICE`: Wirkung Model and Runtime License Agreement 1.0 and upstream Gemma attribution notice.
+## Files
 
-## Enterprise, Custom Models & Research Inquiries
+- `layers/*.o3f`, `vision/*.o3f`, `vocab/*.o3f`: 601 folded HCGM text, vision, and vocabulary operator files (11.56 GB total bundle)
+- `exact_core.safetensors`: exact norms and structural scalars
+- `bundle.json`: shard inventory and SHA-256 digests verified by `wkng verify`
+- `config.json`, `generation_config.json`, `processor_config.json`, `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`: multimodal and tokenizer configuration
+- `paper/Foss_2026_Origami_HCGM.pdf`: the Origami Model Family technical report
+- `assets/`: `banner.png`, `architecture.png`, `mechanism.png`, `mrail.png`
+- `LICENSE`, `NOTICE`: Wirkung Model and Runtime License 1.0 and upstream attribution
 
-Wirkung partners with enterprises, AI labs, and infrastructure providers on:
+## Inquiries and collaboration
 
-- **Custom Model Compilations & Preview Access**: Closed-form HCGM folding of proprietary, domain-specific, or frontier checkpoints into zero-GPU, constant-state operator artifacts.
-- **Exact Formula & Algorithm Synthesis (`Wirkung`)**: Symbolic and spectral operator discovery replacing dense compute bottlenecks with certified closed-form kernels.
-- **Fleet & Inference-Provider Partnerships**: Production deployments of Shared `M-Rail`, disk-backed `sGPU` CPU cloud fabrics, and multi-specialist `HMoE` routing.
-- **Research & Academic Collaboration**: Joint investigations into constant-state sequence modeling, deterministic cross-ISA inference, and reusable computation.
+We welcome inquiries regarding custom zero-retraining HCGM model compilation, closed-form formula and exact algorithm synthesis via `Wirkung`, Shared `M-Rail` / zero-GPU enterprise fleet deployments, and research collaboration:
 
-Reach us directly:
-- **Email:** [`contact@wirku.ng`](mailto:contact@wirku.ng)
-- **X:** [`https://x.com/FossDT`](https://x.com/FossDT)
-- **Website:** [`https://wirku.ng`](https://wirku.ng)
+- **Email:** [contact@wirku.ng](mailto:contact@wirku.ng)
+- **X (Twitter):** [https://x.com/FossDT](https://x.com/FossDT)
+- **Web:** [https://wirku.ng](https://wirku.ng)
 
 ## Papers
 
-- **[The Holographic Causal Graph Machine: Compiling a Pretrained Transformer into a Constant-State Model without Training](https://doi.org/10.13140/RG.2.2.28206.68167)**. David Tom Foss, 2026.
-- **[Crystallized Attention: Fixed-State Carriers and Exact Index Correction for the Full-Attention Heads of a 27B Hybrid Language Model](https://doi.org/10.13140/RG.2.2.21299.18725)**. David Tom Foss, 2026.
-- **[Compute Once: Constant-State Language Models Make Answers Reusable and Agents Cheap](https://doi.org/10.13140/RG.2.2.18140.35202)**. David Tom Foss, 2026.
-- **[Scarcity as Method. Attention, Replaced: Exact Forms, Constant State, and Reusable Inference](https://doi.org/10.13140/RG.2.2.34347.40489)**. David Tom Foss, 2026.
+- **[The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines](paper/Foss_2026_Origami_HCGM.pdf)**. David Tom Foss, 2026. Included directly in this repository at `paper/Foss_2026_Origami_HCGM.pdf`.
+- **[Holographic Causal Graph Machines: Source-Free Constant-State Execution of Pretrained Transformers](https://doi.org/10.5281/zenodo.19040842)**. David Tom Foss, 2026.
+- **[Crystallized Attention: Constant-Memory Exact Store and Exponential-Mode Carriers](https://doi.org/10.5281/zenodo.19058017)**. David Tom Foss, 2026.
+- **[Intelligence Is Not Made of Scarcity: Constant-Memory Causal Execution and Associative M-Rail](https://doi.org/10.13140/RG.2.2.28823.43689)**. David Tom Foss, 2026.
 
 ## Citation
 
 ```bibtex
-@misc{foss2026origami31b,
-  author       = {David Tom Foss},
-  title        = {{Origami-31B: Data-Free Constant-State Multimodal
-                   HCGM Compilation of Gemma-4-31B}},
-  year         = {2026},
-  howpublished = {Wirkung Inc., \url{https://huggingface.co/wirkung/Origami-31B}},
-  note         = {DOI 10.13140/RG.2.2.28206.68167}
+@techreport{foss2026origami,
+  title       = {The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines},
+  author      = {Foss, David Tom},
+  institution = {Wirkung Inc.},
+  year        = {2026},
+  month       = nov,
+  url         = {https://wirku.ng}
 }
 ```
 
 ## License
 
-Licensed under the **Wirkung Model and Runtime License Agreement 1.0** ([`LICENSE`](LICENSE)). Free for non-commercial research, academic evaluation, and personal use in compiled Object form; commercial or production deployment of Origami models, `wkng`, or `sgpu` requires a commercial license from Wirkung Inc. ([`contact@wirku.ng`](mailto:contact@wirku.ng)). Compiled in closed form without training from [`google/gemma-4-31B-it`](https://huggingface.co/google/gemma-4-31B-it) (see [`NOTICE`](NOTICE)).
+Released under the [Wirkung Model and Runtime License 1.0](LICENSE). Compiled in closed form without training steps from Gemma Terms of Use / upstream license (`google/gemma-4-31b-it`); full attribution and closed-form transformation notices are recorded in [`NOTICE`](NOTICE).

assets/mrail_network.png → paper/Foss_2026_Origami_HCGM.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e086b00ea58e35a5d496fec7b28cc9f8b2febe9fc9e3d1a2a82f4daffca1e185
-size 226190
+oid sha256:72360f42140569826578d8c872b993a5be6ec1b43015d2536b82ec7e33ff9868
+size 1716287