jev-bush
Educational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
Overview
Jev Bush is a CPU-first probabilistic decision engine for DiffusionGemma, a diffusion language model. It takes a state, a natural-language predicate and an explicit candidate set, runs tokenization and templating, an encoder prefill and a single-step diffusion answer canvas, then reads candidate logits and applies a candidate-only softmax to return typed probabilities. It never generates free-form text or parses model-written JSON.
It is a small C11 library with a public header and a generated single-file amalgamation, needing only a C compiler and libm, with optional OpenMP. It supports OpenJev's Noul, Choice and Score decisions, exact schema-prefix reuse, batching, deterministic JSONL output and the public OpenJev evaluation set. The scope is deliberately narrow: only the DiffusionGemma 26B BF16 and NVFP4 text checkpoints, scalar, AVX2, AVX-512 and NEON kernels, an optional CUDA validation backend, and limits such as batch size 16 and 4,096 prompt tokens. Strict modes reproduce canonical results bit for bit. It is described as educational and correctness-first, pointing to OpenJev for production GPU inference.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Boolean predicate over the state | yes · no |
| Choice | Which candidate fits? | — |
| Score | Ordinal level | — |
How it works
- Build a typed request
- Tokenize, template and prefill the encoder
- Run one denoising step on the answer canvas
- Read candidate logits and softmax over candidates
Strengths
- Runs on CPU with only a C compiler
- Bit-exact reproducibility in strict modes
- Embeddable C API
Caveats
- Only supports DiffusionGemma 26B checkpoints
- Educational; not a production server
At a glance
- Type
- Project
- Status
- research
- Language
- C
- Runtime
- library and CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Jev Bush is a CPU-first probabilistic decision engine for bounded DiffusionGemma decisions. Please clap.
hellogumbo/awesome-jev
Frequently asked questions
Is Jev Bush related to TypeSafe's Jev?
It implements OpenJev-style decisions on DiffusionGemma, not TypeSafe's hosted model.
Does it need a GPU?
No, it is CPU-first; CUDA is optional for validation.
Which models are supported?
Only DiffusionGemma 26B BF16 and NVFP4 text checkpoints.
Related projects
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustExperimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
ProjectresearchPythonReads typed decisions from one forward pass of PrismML's 1.75-bit Bonsai 27B under llama.cpp, using the OpenJev/SemIf logit method.
ProjectresearchPythonJev-class decision server on an RWKV-7 1B checkpoint that reads typed answers from a constrained softmax behind /v1/systemone.
ProjectresearchPython