GitHub Next ships LocalJev and documents plainly that its probabilities are self-reported, not read from logits
githubnext/localjev appeared on 2026-09-18 (195 stars in a day), a Bun/TypeScript bridge exposing a local Jev-compatible POST /v1/systemone endpoint backed by diffusiongemma-26B-A4B-it-4bit over an OpenAI-compatible Chat Completions API. It translates state and typed questions into a classification prompt, asks for a JSON probability vector, validates and retries malformed output, normalizes, and returns Jev-shaped choices, expected scores and entropy-based confidence. The README is unusually honest about the limit: unlike razorback16/openjev, which reads probabilities through unmerged vLLM extensions (diffusion_seed_canvas, diffusion_read_only, token logprobs), LocalJev's numbers are generated by the model, so 'evaluate their calibration on your own workload before relying on them for consequential decisions.'
↳ Follow the thread