jevfire

What jevfire does

jevfire is a Jev-inspired toolkit for assigning typed variables from finite choice sets on CUDA-backed LLMs. Instead of asking a generative model to write constrained JSON token by token, it scores verified single-token labels through vLLM, maps winners onto your schema, and assembles the JSON in application code. Independent fields that share the same instruction and context can reuse the engine prefix cache, which is why multi-field workloads often land far below a generate-and-parse baseline on the same weights. The project ships browser demos (including a Super Mario World 1-1 recreation on WebGPU), a visual learn guide, race/driving experiments, and reproducible benchmarks with published fixtures. It is useful when you already trust an off-the-shelf instruct model and need structural guarantees that the model cannot invent fields or out-of-set values — factual correctness of the chosen allowed value is still your problem to evaluate. Named as a tribute to JEV / RLCD research, jevfire is an independent CUDA implementation, not the hosted TypeSafe Jev API.

Where it fits in the Jev stack

• Game-agent action layers that need many enum decisions per frame
• Batch structured extraction where every field has a closed option set
• Latency benchmarks comparing parallel scoring vs grammar-constrained generation

Browse more in Benchmarks & research, or return to the Jev AI tools directory.