Set up

Describe your setup once.

You get back a brief.json and two commands. Nothing on this page is sent anywhere: the file is written by your browser, and everything after it runs on your machine. No checkpoints, no data, no account.

Project nameTaskModel or architectureFrameworkCurrent checkpointCandidate checkpoint Optional, can come laterWhere the cases areHow one case is storedGround truthOne line on the layout Optional, and it lands in the adapter next to the code that reads itRefinement iterationsDevice

Your own evaluator and metric Recommended

Path to your evaluation scriptPath to your model codeMetric idUnitWhat this review has to decide Optional; it goes in the receipt

Naming your evaluation script is the single most useful thing on this page. It becomes the reference path: before any run reports a number, the adapter has to reproduce what your own evaluator computes, and a disagreement beyond 0.001 relative stops the run. Without it every receipt says the agreement was never established.

Then, on your machine

Two commands, then a ladder.

pip install rabbit-brain
rb onboard --brief brief.json

What comes back

A configured rb.toml, an INTEGRATION.md with the ladder, and rb_adapter.py: the shape of an adapter for your model, with the parts only you can write marked one by one. It is not a working adapter, because nothing can guess your loader, your valid mask or your metric formula.

The ladder

rb doctor for the environment, rb verify-hook to prove the recorder fires once per iteration, rb verify-adapter to prove the numbers match your own evaluator, then rb run --baseline <current checkpoint> --candidate <candidate checkpoint> --limit 5 before the real review. Each step says what it proves and what to do when it fails.

If you would rather have help, or your model turns out to need an adapter we should write, the setup request is still there. It is not a prerequisite for any of this.