Jevmlx is a library that enables MLX models on Apple Silicon to make typed, schema-valid JSON decisions in a single batched pass by prefilling context once and broadcasting the KV cache across schema fields. It guarantees valid JSON output by assembling values from allowed choices rather than generating them token-by-token, and provides a Python API, HTTP server, and honest confidence scores for each field.
A high-performance LLM serving scheduler written in Rust achieves 191k tokens/second on a single vCPU using continuous batching, paged KV cache allocation, and lock-free scheduling. The repository includes formal SMT Z3 mathematical verification, an independent anti-cheat auditor, and a reproducible benchmark harness under strict hardware isolation.