Docs

Neurarch documentation

The pre-check for a change to an open-weight model, before it takes a GPU or serves traffic: what changed, what the job needs, whether the parallel layout is legal and balanced, and whether the serving config will start. Same code behind every entry point below.

Start here

CLI quickstartnpx neurarch: memory, layout, serve-check, data fit, past jobs. No account. Admission hooksSkyPilot policy from PyPI, SLURM cli_filter, and a Claude Code hook. PR bot and ActionThe plan on every model change pull request; a blocker fails the check. Change feedsRenovate preset, OSV advisories and badges for upstream model changes. CycloneDX propertiesThe neurarch:* properties --cyclonedx writes into a BOM, and what each means. APIPOST a graph with data, layout or serve; same calls self-hosted. Trust and data handlingWhat each of these sends, and what it does not.

Reference

API reference: every endpoint and field.
MCP server: the same checks as tools inside Claude Code, Cursor and other MCP clients.
Memory calibration: every measured run behind the training-peak estimate.
Model changes and model pages: the tracker and per-model sizing.
The checks: the structural rule catalogue.
Changelog: what shipped, from merged pull requests.

Older docs

Kept at their addresses for existing users and links. They describe the design app and the research work that came before the pre-check.

Design app documentation (canvas, agent, exports, training) and its feature pages.
Markdown docs for agents: overview, quickstart, agent, structural checks, import and export, MCP.
Research: the RL environment, environment page, Arch-Bench leaderboard.