Our mission is to make post-training accessible to everyone.
Frontier labs post-train their models in environments they build in house, on data nobody else has. Everyone else runs a general model and pays for it, on tasks where a smaller model trained on the right data would do as well or better. The recipes are published. What is missing is the environment: an executable copy of the world your agent works in, tasks that came from real use, and a way to check that a task was done.
That is the piece we are building first. Kullback is an open-source harness that turns the traces your agent already produced into an environment: the data your runs touched, one function per tool that behaves the way the real tool was observed to, the rules your agent was given compiled into checks, a simulated user who knows what the real user knew, and a verifier per task that looks at what changed in the world, not at how the model talked about it. Before anything else happens, the rebuild is checked by replaying recorded runs it has not seen.
With that environment you can run any model through your own tasks and see which ones it clears, route each task to the cheapest model that clears it, and post-train a smaller model on it, because every run comes with a verdict you can trust.
The claim we are out to earn, and have not made yet: a 2B parameter model post-trained in an environment Kullback built performs on the real tasks. The numbers get published either way.
Kullback is Apache-2.0. The design, the decision log, the research and the list of what is not built yet are in the repository. Contributions are welcome; CONTRIBUTING.md says how.