One directory per tutorial. Each opens with a README that says what you will see, then a notebook that runs the whole pass cell by cell, then the script form. Configs and helper code sit in subdirectories. The recipe examples under recipes/ are the canonical run configs for each method; the tutorials are the walkthroughs.
| Tutorial | What you see | Needs | Open |
|---|---|---|---|
| Evolve your harness | Three coding tasks go through Reef, one fails, the served model proposes a change to its own harness, the gate runs both versions, and the winner publishes as a new harness version. | An OpenAI compatible model endpoint (a local ollama is enough), no GPU; the pi coding agent for the pi variant, nothing extra for the native variant. | notebook, or ./run.sh in the directory |
| Reefine | A person asks reef-pi for a capability in plain words; the service proposer writes the change, the gate scores it, an evolved extension waits for a promote, and the next session runs the new tree. Two demos (a bug fix flow, a research loop) and a measurement of which requests won the gate. | An OpenAI compatible model endpoint (a local ollama is enough), no GPU; the pi coding agent. | ./run.sh bugfix in the directory |
The Tinker smoke covers remote weight training with four short rollouts and one update. For full training experiments, Evolve your model and the recipe examples under recipes/ are the path.