# Neurarch

> Neurarch keeps a neural network as a typed graph: nodes are layers, edges carry tensor shapes. Because the model is structured data, every edit (by a human or an AI agent) is verified by 41 structural checks in under a millisecond, before any GPU time is spent. The same graph exports the PyTorch you would have written, and doubles as a verifiable environment for training and benchmarking architecture-design agents.

Key facts: 182 layer types, 75 prefab blocks, 36 templates, 41 structural checks (6 guardrail gates + 35 advisor rules). In a 264-graph study against real PyTorch (two seeds, torch 2.8), all 96 designs the verifier blocked crashed at forward time and all 80 it passed ran clean. The dense verifier costs 0.27 ms per edit on a 47-layer graph, one laptop core.

## Docs

- [Overview](https://neurarch.com/docs/overview.md): what the typed graph is, what runs on it, and what the product does end to end
- [Quickstart: your first caught bug](https://neurarch.com/docs/quickstart.md): the 60-second path from opening the app to watching a blocking check fire
- [Structural checks](https://neurarch.com/docs/structural-checks.md): the 41 checks, how they are layered, and the measured evidence they track reality
- [The agent](https://neurarch.com/docs/agent.md): how the AI agent edits the graph through guarded structured actions
- [Import and export](https://neurarch.com/docs/import-export.md): paper / PyTorch / HuggingFace / ONNX in, PyTorch / notebooks / deploy bundles out
- [MCP server](https://neurarch.com/docs/mcp.md): point neurarch-mcp at your PyTorch .py and Claude, Cursor or VS Code can read the graph, lint the design offline, and check the blast radius of an edit before making it
- [RL environment and benchmark](https://neurarch.com/docs/rl-env.md): the verifiable environment and public leaderboard for architecture-design agents

## Architectures (machine-readable)

- [Architecture index (JSON)](https://neurarch.com/a/index.json): every shipped architecture with its layer count, parameter count, verifier verdict, and a URL for the graph JSON itself. Fetch a graph, edit it, POST it to /api/v1/check.
- [Architecture index (HTML)](https://neurarch.com/a/): the same set for people. Each entry has a markdown twin at /a/<key>.md.
- [Verify a graph](https://neurarch.com/developer.html): POST /api/v1/check runs the same structural checks the app runs on every edit.

## Reference pages (HTML)

- [The 41 rules, one by one](https://neurarch.com/rules.html): every gate and advisor rule with trigger conditions and citations
- [Caught before the GPU bill](https://neurarch.com/caught-bugs.html): planted-bug gallery plus the 264-graph verifier study
- [Arch-Bench leaderboard](https://neurarch.com/leaderboard.html): which frontier models design the best architectures
- [Environment and SDK](https://neurarch.com/env.html): train your own agent against the environment

## Optional

- [Human-readable docs](https://neurarch.com/docs.html): the same material as the markdown docs, rendered for people
- [Landing page](https://neurarch.com/landing.html): product positioning and pricing
- [neurarch-mcp on GitHub](https://github.com/neurarch-ai/neurarch-mcp): open-source MCP server, MIT