jev-brain
⚠️ Why not a model
A gateway that reads option-letter logprobs from an unmodified chat model served by vLLM. The weights are unchanged; no decision training of its own.
Technical specs
- Base LLMs
- Gemma-4-31B (Decision Index run), any vLLM-served chat model
- Decision types
- choice, score, noul
- Features
- training-free, logit-readout, vllm, typesafe-api, single-gpu, chat-and-decisions
- License
- MIT
Links
jev-brain (MIT, created 27 Sep 2026) runs one vLLM process for normal OpenAI-style chat and adds a small gateway with a Jev-compatible POST /v1/systemone endpoint over the same loaded model. Each typed question becomes a multiple-choice prompt that stops at Answer: (; vLLM generates one token and the gateway reads the log-probabilities of the option letters.
Board result. On Decision Index 0.4 (20 Oct), “jev-brain · Gemma-4-31B” has Full 65.29 (#15), above Jev 1.14 (62.22). That is a stock Gemma-4-31B read through prompts and logits, which is a useful baseline for how much fine-tuned decision models add.
Why it is here, not in the catalog. The project trains no weights; chat keeps working because the model is untouched. Same class as AnyJev and simple-jev.
