Add Origami-9B Enterprise Preview model card and header figures

.gitattributes

@@ -346,3 +346,7 @@ layers/vision/merger__linear_fc2__weight.o3f filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 vocab/embed_tokens.o3f filter=lfs diff=lfs merge=lfs -text
 vocab/lm_head.o3f filter=lfs diff=lfs merge=lfs -text
+assets/architecture_organs.png filter=lfs diff=lfs merge=lfs -text
+assets/banner.png filter=lfs diff=lfs merge=lfs -text
+assets/mrail_network.png filter=lfs diff=lfs merge=lfs -text
+assets/weights_vs_origami.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,152 @@
+---
+license: other
+license_name: wirkung-model-and-runtime-1.0
+license_link: LICENSE
+library_name: wirkung
+pipeline_tag: text-generation
+base_model:
+- Qwen/Qwen3.5-9B
+base_model_relation: quantized
+language:
+- en
+- de
+- zh
+- fr
+- es
+- ja
+- ko
+tags:
+- wirkung
+- origami
+- preview
+- hcgm
+- m-rail
+- sgpu
+- hmoe
+- constant-state
+- cpu-inference
+- zero-gpu
+- multimodal
+- qwen3.5
+---
+
+![Multimodal hybrid Holographic Causal Graph Machine with disk-mapped execution, constant session state, and native M-Rail support.](assets/banner.png)
+
+[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/Qwen/Qwen3.5-9B) · [License](https://huggingface.co/wirkung/Origami-9B/blob/main/LICENSE)
+
+**Origami-9B (Preview)** is a multimodal hybrid Holographic Causal Graph Machine (HCGM) compiled in closed form from [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) with zero gradient updates and zero fine-tuning steps. Executed directly from NVMe storage via OS `mmap` under the sovereign [`wirkung`](https://pypi.org/project/wirkung/) runtime, Origami-9B requires no GPU VRAM, eliminates PyTorch and CUDA framework overhead, maintains a strictly constant `O(1)` live session state across arbitrary sequence horizons, and produces bit-identical greedy token sequences across server, desktop, and edge CPU architectures.
+
+```bash
+pip install -U wirkung && wkng pull wirkung/Origami-9B -o ./Origami-9B
+wkng run ./Origami-9B -p "Explain constant-state Ward carriers." --chat
+```
+
+## Architectural Overview: Upstream vs. Origami
+
+| Architectural Dimension | Conventional Checkpoint Serving | `Origami-9B` (Wirkung HCGM Preview) | Structural Invariant |
+|---|---|---|---|
+| Compilation & Adaptation | Pretrained dense & hybrid weights | **Closed-form algebraic fold** | Zero gradient or training steps |
+| Execution Substrate | GPU HBM + PyTorch / CUDA / vLLM | **Self-describing `mmap` operator file** | Zero GPU VRAM required |
+| Weight Residency | Deserialized into resident VRAM / RAM | **Demand-paged directly from SSD / NVMe** | Shared read-only OS pages across sessions |
+| Live Conversation State | `O(N)` KV cache growing with context | **Strictly constant `O(1)` carrier state** | Flat memory across arbitrary horizons |
+| Multimodal Backbone | Separate dense vision & text execution | **Unified `.o3f` vision + hybrid graph** | Single coherent operator runtime |
+| Vocabulary Readout | Full dense vocabulary matrix pass | **Certified Cauchy-Schwarz norm gate** | Exact greedy argmax preservation |
+| Cross-Hardware Execution | Backend-dependent floating-point drift | **Bit-identical greedy token trajectories** | Deterministic across ARM64, x86_64 & edge |
+| Fleet Reuse | Stateless per-request recomputation | **Shared `M-Rail` (`X = X_known + R`)** | Associative fleet-wide residual closure |
+
+![Disk-mapped execution and constant session state](assets/weights_vs_origami.png)
+<sub><b>Figure 1.</b> Execution substrate and live session memory trajectory. Conventional serving deserializes full checkpoints into GPU memory and expands KV state linearly with sequence length, whereas Origami-9B executes directly from NVMe via read-only `mmap` and holds live session state strictly constant (`O(1)`) across arbitrary context horizons.</sub>
+
+## Signature HCGM Architecture & Distinguishing Capabilities
+
+- **The File Is the Machine (`mmap` Operator Graph)**: Dense projections across the hybrid language decoder, vision encoder, and vocabulary head fold into self-describing `.o3f` spectral operators stored on disk. At runtime, the OS maps operator pages directly into virtual address space on demand, allowing multiple concurrent sessions to share a single read-only mapping without loading weights into resident RAM or GPU HBM.
+- **Constant `O(1)` Hybrid Carrier State**: Recurrent Gated DeltaNet carriers on linear layers paired with crystallized pinned-sink Ward carriers on full-attention layers bound live conversation memory to a fixed footprint regardless of prompt length, document size, or multi-turn agent duration.
+- **Spectral Gate-Confidence Denoising**: Operating in the randomized Walsh-Hadamard spectral basis, gate-confidence filtering suppresses low-energy pretrained tail noise inside SwiGLU feed-forward blocks without fine-tuning.
+- **Certified Cauchy-Schwarz Vocabulary Gate & Cross-ISA Determinism**: Exact row-norm certificates stored alongside the folded output head prune dominated vocabulary candidates prior to full inner-product evaluation while preserving exact greedy argmax tokens identically across Apple Silicon ARM64, x86_64 AVX2 servers, and mobile/edge ARM SoCs.
+
+![HCGM organ and operator architecture](assets/architecture_organs.png)
+<sub><b>Figure 2.</b> Three-stage full-width HCGM operator pipeline of Origami-9B: multimodal vision-and-language operator intake over OS `mmap`, constant-state hybrid sequence processing with spectral gate denoising, and certified Cauchy-Schwarz vocabulary readout with cross-ISA bit-exactness.</sub>
+
+## Shared M-Rail (`X = X_known + R`), sGPU Fleet Economics & HMoE
+
+![Shared M-Rail, sGPU fleet, and HMoE topology](assets/mrail_network.png)
+<sub><b>Figure 3.</b> Fleet topology over the Shared `M-Rail` associative trajectory bus. Recurring trajectories replay directly from verified stations (`R = 0`), while newly encountered residuals (`R != 0`) are solved once by any node and shared immediately across CPU and GPU peers.</sub>
+
+Origami-9B integrates natively with the **Shared `M-Rail`** trajectory bus and bundled **`sGPU`** (`sgpu`, `cpuda`, `rail`) deterministic CPU execution layer:
+
+1. **Warm Up Once, Serve Indefinitely on Disk-Backed CPU Fleets**: Every forward trajectory decomposes into known crystallized structure plus residual (`X = X_known + R`). Organizations can optionally use a single GPU or a brief cloud GPU lease to crystallize recurring workflows, system prompts, and agentic pipelines into `.mrail` tracks, shut GPUs down, and serve production traffic continuously on commodity CPU VMs backed by NVMe storage.
+2. **Collective Associative Residual Closure**: Whenever any node in the fleet computes an unseen residual branch (`R != 0`), it commits verified `ANCHOR` / `DELTA` stations and `OP_COPY` morphisms to the shared `M-Rail` store (`wkng maps merge`), closing that trajectory for every peer node with zero greedy output drift.
+3. **Heavy Mixture of Experts (`HMoE`) Specialist**: Mounted on the synchronous `M-Rail` `cstar` bus alongside `Origami-27B`, `Origami-31B`, `Origami-78B`, and `Origami-H3`, Origami-9B operates as a fast multimodal specialist, tool-calling router, and `DraftRail` span verifier.
+
+## Runtime CLI & Python SDK
+
+The standalone [`wirkung`](https://pypi.org/project/wirkung/) package (`pip install wirkung` or `pip install wkng`, `Python >= 3.10`, `numpy` as the sole Python dependency) provides both CLI tools and an embeddable Python engine:
+
+```bash
+# Verify host SIMD capabilities and model integrity
+wkng doctor
+sgpu doctor
+wkng verify ./Origami-9B
+
+# Launch an OpenAI-compatible local HTTP server or merge fleet M-Rail maps
+wkng serve ./Origami-9B --host 127.0.0.1 --port 8080
+wkng maps merge ./peer_node_maps --into ~/.wkng/maps
+```
+
+```python
+from wirkung import Engine
+
+engine = Engine("./Origami-9B")
+response = engine.generate(
+    "Summarize the architectural role of pinned-sink Ward carriers in constant-state attention.",
+    chat=True,
+)
+print(response)
+```
+
+## Repository Contents
+
+- `bundle.json`, `origami9b_manifest.json`, `config.json`, `generation_config.json`: HCGM bundle schema, operator manifest, and hybrid multimodal architecture metadata.
+- `layers/`: compiled `.o3f` spectral operators across all hybrid decoder blocks and the multimodal vision encoder.
+- `vocab/`: folded vocabulary readout operators with certified Cauchy-Schwarz row-norm bounds.
+- `exact_core.safetensors`: exact normalization parameters, recurrent routing gates, and vision conditioning norms.
+- `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `processor_config.json`, `preprocessor_config.json`, `video_preprocessor_config.json`: standalone tokenizer and multimodal preprocessors.
+- `LICENSE`, `NOTICE`: Wirkung Model and Runtime License Agreement 1.0 and upstream Apache-2.0 attribution notice.
+
+## Enterprise, Custom Models & Research Inquiries
+
+Wirkung partners with enterprises, AI labs, and infrastructure providers on:
+
+- **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.
+
+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)
+
+## 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.
+- **[No GPU, No KV Cache: A Constant-State Language Model on the Neural Engine of a Mac mini](https://doi.org/10.13140/RG.2.2.31562.12484)**. 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.
+
+## Citation
+
+```bibtex
+@misc{foss2026origami9b,
+  author       = {David Tom Foss},
+  title        = {{Origami-9B: Closed-Form Constant-State HCGM Compilation of Qwen3.5-9B}},
+  year         = {2026},
+  howpublished = {Wirkung Inc., \url{https://huggingface.co/wirkung/Origami-9B}},
+  note         = {DOI 10.13140/RG.2.2.28206.68167}
+}
+```
+
+## 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 [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) (Apache License 2.0; see [`NOTICE`](NOTICE)).