jevify (Mintzs)
Experimental engine that makes an existing LLM do Jev-like classification and scoring by reusing context and building JSON.
Overview
Jevify (Mintzs) runs classification, yes/no judgments and rubric scoring with an existing pretrained LLM, defaulting to Qwen2.5-1.5B-Instruct with no fine-tuning. It reuses shared context by processing the common input once and reusing its KV cache across question branches, evaluates independent questions in batches, constructs JSON directly from model scores so output is always schema-valid, and offers optional GPU kernels to reduce overhead.
On a local 250-case development benchmark on a 4 GB RTX 3050 laptop GPU, the README reports 90.0% correct and schema-valid at 167 ms median latency, versus 9.2% at 4,359 ms for standard unconstrained JSON generation from the same model (failing mostly on formatting) and 74.8% at 408 ms for an RLCD-tuned Qwen checkpoint's reference implementation; it summarises this as 26.1 times faster than standard JSON generation. The author is careful that these are development-set results with some cases used in prompt tuning, that it is not affiliated with TypeSafe or meant to compete with Jev, that it does not reproduce Jev's architecture or RLCD training, and that its probabilities are uncalibrated.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which label or route applies? (single-decision benchmark cases) | — |
| Yes / no | Yes/no judgments read from model scores | — |
How it works
- Encode shared context once and cache it
- Branch per question and score options
- Batch independent questions
- Assemble schema-valid JSON from scores
Strengths
- 100% schema validity by construction
- README reports 167 ms median on a laptop GPU
- No fine-tuning required
Caveats
- Development-set benchmark, partly used for tuning
- Uncalibrated probabilities
- Experimental
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch / CUDA
- Has code
- yes
Use cases and integrations
How curated lists describe it
An optimized inference engine to turn LLMs into Jev-like machines: optimized for quick, lightweight, and accurate decision-making, classification, and scoring.
hellogumbo/awesome-jev
Frequently asked questions
Is this a Jev replacement?
No, the author says it does not reproduce Jev's architecture or calibration and is not meant to compete.
How accurate was it?
90.0% correct and schema-valid on a 250-case development set, per the README.
Are probabilities calibrated?
No.
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.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchCReads 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