Start the proxy: stuntd serve --upstream https://api.openai.com. The upstream is the provider origin only, without the /v1 suffix: your client's own path is appended to it unchanged.

Set OPENAI_API_KEY, and STUNTD_MODEL if you want a model other than gpt-4o-mini.

Run python examples/moderation_client.py to send one typed decision through the proxy.

Four demos of the Jev protocol, each with its own data, client and README:

• snake/ -- a game of Snake, one typed decision per move, trained from a BFS oracle's moves; the launch demo, with the numbers that run measured.
• support/ -- triage a ticket: its category, its urgency and whether a human has to read it, three questions in one request.
• banking/ -- what a customer's message is about and what to do with the payment attached to it.
• devtools/ -- may a coding agent run this shell command, and was the last test failure a flake; the README shows it as a Claude Code PreToolUse hook.

Each demo generates its own labelled rows, imports them into a site, trains a head and then runs a typesafe-sdk client against the daemon. The teacher in all four is a rule or an oracle in the generator that reads the words of the state, so they need no API key. Five of the seven sites of the three text demos hold a holdout agreement of 0.95 or better at the default target of 0.99; each README has the numbers and the cost in GPU seconds.