jevbench

JevK5: open-weight alternative to TypeSafe Jev. Typed decisions with probabilities in one forward pass; Apache-2.0 weights and code.

Code and data for evaluating Jev, a System One model, on scientific decisions and how its choices affect downstream results.

Convert any causal LM into a Jev-style typed decision model — no training, no new weights. Measured honestly against JevBench, negative results included.

A decision model: give it a document and a typed question, get calibrated probabilities for every allowed answer. No text generation. Qwen3.5-4B + LoRA, /v1/systemone wire format.

Open-source Jev harness for TypeSafe, OpenJev, LocalJev, Ollama, vLLM, LM Studio, llama.cpp. Typed decisions, calibration, verification, abstention, CI gates.

Local frozen-backbone decision readout: evidence + criterion + options in, a probability per option out, in one forward pass. JevBench public set 0.805 / hard 0.604 with Qwen3.5-4B, no training.