A neural net you can read

A page layout computed by a hand-written neural network (6,693 neurons in 201 layers, built with handwired). You breed the layout, then a five-part interpretability lesson takes the network apart: reading neurons, seams as neurons flipping, knockouts vs. gradients, activation patching, and probes. Every neuron was written by a person, so each method is checked against an answer key.

bun install
bun start          # dev server for site/index.html
bun test ./tests   # circuit vs. oracle, lens evaluator, breeding, and the lessons' claims
bun run build      # static site in dist/ (relative paths, works under a sub-path)

Layout

• vendor/handwired/: handwired's runtime and generative-fields experiment, unchanged (MIT, see VENDOR.md).
• src/layout/: our fork of the layout circuit (a minimum tile width in px, and neuron labels), its float64 oracle, the genome schema, and helpers to evaluate a genome.
• src/lens/: the network view of the circuit (layers, labels, an evaluator that can override any neuron) and the experiments the lessons run.
• src/breed/: mutation, crossover, the 7-weight taste model, and share links.
• site/: the page.
• SPEC.md: what was planned, and what changed while building it.

Deploys to GitHub Pages through .github/workflows/pages.yml on a push to main.