initial deployment: cortex-conv ships pre-trained at 96.8% MNIST

A 34,106-weight convolutional neural network trained with Equilibrium Propagation
(no backprop) on FitzHugh-Nagumo-inspired neurons. Boots pre-trained at 96.8%
test accuracy on MNIST. Toggle to Fashion-MNIST to watch the same network train
from scratch in your browser on WebGPU.

Companion to Kendall 2026 (arXiv:2605.21568). Much smaller variant: 34K params
vs paper's 1,457,674; ~3.2% test error vs paper's 2.8%; runs in browser instead
of Python.

See PAPER_COMPANION.md for the full write-up.

.gitattributes

@@ -1,35 +1,7 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Large data packs and weights via Git LFS — required on HuggingFace Spaces for files > ~10 MB.
+*.json filter=lfs diff=lfs merge=lfs -text
+# Keep README + tests human-readable in the web UI (not LFS-ed).
+README.md -filter -diff -merge text
+index.html -filter -diff -merge text
+tests/* -filter -diff -merge text
+PAPER_COMPANION.md -filter -diff -merge text

DEPLOY.md

@@ -0,0 +1,58 @@
+# Deploying this directory as a HuggingFace Space
+
+This directory is a self-contained static HuggingFace Space ready to push. Total bundle size is ~124 MB (the bulk is the two 60K data packs at 61 MB each); the index.html + scripts + weights together are under 3 MB.
+
+## One-time setup
+
+You need: a HuggingFace account, the `huggingface_hub` Python CLI, and an HF write token.
+
+```bash
+pip install -U huggingface_hub
+huggingface-cli login          # paste your token from https://huggingface.co/settings/tokens
+```
+
+Install Git LFS once (required for the 60 MB data packs):
+
+```bash
+brew install git-lfs           # macOS;  apt-get install git-lfs  on Linux
+git lfs install
+```
+
+## Create the space and push
+
+```bash
+# 1. Create an empty static space named cortex-conv on huggingface.co/spaces/<your-username>/cortex-conv
+huggingface-cli repo create cortex-conv --type space --space-sdk static
+
+# 2. Clone the empty space repo somewhere
+git clone https://huggingface.co/spaces/<your-username>/cortex-conv ~/cortex-conv-space
+
+# 3. Mirror this directory into the clone
+rsync -a --delete --exclude=.git /Users/exobit/developer/fnm-gpu/hf_space/  ~/cortex-conv-space/
+
+# 4. Push (the .gitattributes file already routes the JSON packs to LFS)
+cd ~/cortex-conv-space
+git lfs track "*.json"
+git add .gitattributes README.md index.html PAPER_COMPANION.md tests/ weights/ mnist/ fashion/
+git commit -m "initial: cortex-conv ships pre-trained at 96.8% MNIST"
+git push
+```
+
+The Space will build automatically (a static space is just a file server, no build step). Open `https://huggingface.co/spaces/<your-username>/cortex-conv` and the page should load at 96.8% MNIST test accuracy within ~3 seconds on first visit.
+
+## Verifying the deployed space
+
+Visit the Space URL with a Chrome/Edge/Safari browser that supports WebGPU. Within 3 seconds you should see:
+
+- Test accuracy: **96.8%**
+- Status label: `idle (cortex-conv · webgpu+pretrained · 34,106 params)`
+- Equation header: `CORTEX-CONV (2-conv 16,16, γ=0.1, banked V1 mask, drive-clamp · 34,106 params) · CORTEX NEURON · ADAGO · MNIST`
+
+If WebGPU is unavailable in the visitor's browser, the page falls back to a CPU-only path with the embedded 6K MNIST and runs much slower.
+
+## Notes
+
+- HuggingFace Spaces by default serve from CDN with proper MIME types for `.json` and `.js` — no extra config needed.
+- The `sdk: static` line in `README.md` frontmatter is what tells HF this is a static site (not a Gradio or Streamlit app).
+- Updates: after editing files in `/Users/exobit/developer/fnm-gpu/hf_space/`, repeat the `rsync` + `git add` + `git commit` + `git push` steps. HuggingFace re-deploys automatically on every push.
+- To regenerate the bundled weights snapshot after architecture changes, run `tools/train_cortex_dump.cjs` from the project root (not from `hf_space/`), then copy the resulting `weights/cortex_conv_mnist_R28.json` into `hf_space/weights/`.

fashion/fashion_pack_28_60k.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c5fb58ccc88be6d0f36c49170cecc7b4d17bf0ba45ae7ec2d5df84895944bb2
+size 63846739

mnist/mnist_pack_28_60k.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07ba3802fa3bedde724c219fafb4e7af6ad4924e68468e0b0e4c3ef56be90522
+size 63846739

weights/cortex_conv_mnist_R28.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:721b7fe15f85ad307be698f10b7ea8e1ed093d5a1f84461151dda7994e1ff7fc
+size 719874