jlt-commons/lev is a System 1 decision engine that answers typed questions about state with calibrated probabilities. The latest release listed in the repository is v0.3.4, and the provided information contains no specific details about what changed in this version. lev runs in jolt, which provides Clojure on top of Chez Scheme.
The core structure is that a fast encoder answers first, and only when confidence is low is the request handed to a GGUF thinker. Behind the API sit the ModernBERT-large and mmBERT-base encoders, a GGUF chat model used through llama.cpp, Clef, and TinyDecide. The Laya encoder checkpoints are weights that add a decision head on top of the two encoders. You can specify a model in the request or route to an encoder based on content, and the questions of a single call are bundled into one forward pass. The design calibrates probabilities and allows probability-based reasoning only when needed. In the published comparison, it leads decisively with 97.5% on AG News, while on SST-5 it trails generative models with 27.5%. The numbers for the competing models von-1.0 and localjev are published alongside it.
That said, all performance figures are self-measured on a single 10-core M-series laptop. The encoder checkpoint weights have to be downloaded from elsewhere, so reproducing them requires separate preparation. The published comparison can therefore be treated as a reference for the design, but it is difficult to claim that the same performance would appear in other environments.