Run it yourself

Everything is developed and verified on a Mac; a Linux box with a GPU is used only for full-graph training. The code is Python 3.12 via uv, Postgres in Docker, and a Vite + Three.js web app.

Mac

make install          # uv sync, npm ci
make db-up            # Postgres 16 in Docker
make worker           # collector, labeller, live loop, swarm, ledger, with reload
make web              # the page on http://localhost:5173

You need gmgn-cli on the path and GMGN_API_KEY in .env (see .env.example). The collector polls the trenches feed every 15 seconds and labels tokens an hour later; the fly needs a few hours of that before it is worth training.

make dataset          # join snapshots and labels, split by time and token
make train            # one fly on the nose subgraph, CPU, minutes
make evaluate         # the table from the Results page
make replay           # drive the page from a recorded session, collector off

The connectome

make connectome downloads the MaleCNS tables (1.1 GB, checksums pinned) and builds the graph; make subgraph cuts the nose out of it; make geometry pulls the skeletons the web app draws. All three are deterministic and their outputs carry a manifest with every source URL and SHA-256.

Linux box

Full-graph training, and the eight seeds for the swarm, run on an RTX 5070 Ti over Tailscale. docs/runbook-linux.md in the repository has the driver, CUDA and rsync steps; make remote-train and make remote-swarm do the runs from the Mac.

What is not here

No wallet code. Trading is paper until there are two weeks of a track record on this site, and real execution, tiny sizes, is its own task with its own runbook.