liqiiiii/jeva is a repository for a System 1 (system one) decision model built by researchers at the National University of Singapore. Given a question with a specified state and type, it returns per-candidate probabilities, the probability that a correct answer exists, and an expected evaluation score. The public model is Jeva-4B, and the repository focuses on explaining the System 1 model from first principles.
The model uses a language encoder to represent the state and each candidate independently, while a shallow decision stack examines the representations of all candidates and updates only the state. This structure makes permutation symmetry possible, so results are preserved even when candidate order is changed, and enables open-ended processing that scores candidate sets without predefined answer labels. It also presents coverage estimation that distinguishes the case of no correct answer from competition among candidates, along with a cache of all candidate combinations and reuse of individual candidate vectors. The published deployment environment is PyTorch 2.9.1 and CUDA 12.8, and kernels can be compiled on the first request. The license is Apache-2.0. The repository's primary language is Python with 85 stars.
That said, the repository has no tests and only one evaluation document and a single evaluation module, so its technical depth is hard to verify. What the repository backs up therefore extends only to explanations of the structure and design principles; the information given offers no basis for judging actual performance or suitability for practical use.