Setup help
Getting it running on your model.
Rabbit Brain needs per-case numbers for both checkpoints and somewhere to run. Most teams set that up themselves, and the page below writes the configuration for you. We also do it as fixed-price work with defined scope and an acceptance test.
Describe your setup once.
Answer nine questions about your model and data. You get a brief.json, an rb.toml, an integration guide, and the shape of an adapter for your model with the parts only you can write marked out. Nothing is sent anywhere, and it costs nothing.
Paid setup
Two engagements, approved separately. The first delivers a working gate. The second is research and only makes sense after the first.
Release review setup
We connect your evaluation outputs, produce the first comparison on your cases, agree the limits, and hand over a command your CI runs on every candidate.
- We connect your existing evaluation outputs, writing the results adapter or the model adapter for a supported family
- A baseline and candidate comparison on your own cases, with the receipt
- Saved checks configured to limits you agree, not to our defaults
- A repeatable command and a CI configuration you own and can run without us
- Scope
- one model family, one case set, one lower-is-better metric with a unit
- Delivery
- ten working days from receiving inputs and an agreed start
- Accepted when the command runs in your CI, reproduces the report, and the checks fail on a case you nominate and pass on one you do not
- Training, retraining, monitoring, certification and on-call support are out of scope
Trajectory scorer study
Measures whether the trajectory signal adds anything on your model, against the checks you already run. It may conclude that it does not.
- We fit a trajectory scorer to your model and compare it against the checks you already run
- Measured at equal retained coverage, with the runtime overhead recorded
- You get the scorer, the protocol and the report, and can re-run it on the next checkpoint
- This is research. Fitting a scorer is not evidence that it helps your model
- A result of "this does not improve on your current checks" is a result, and the work is still billed
- Approved separately, after the release review setup, never bundled with it
Prices exclude tax. The form charges nothing: scope, inputs and a start date are agreed in writing first. If the acceptance test does not pass, you do not pay for the second half.
Other options
The tool is free and complete. Read the docs or look at a real comparison first.