What it is
Luce is an open recipe for building Jev-style decision models. It scaffolds questions, generates or annotates data with an LLM teacher, trains a LoRA + decision head on Qwen, evaluates calibration, and serves probabilities.
How it uses Jev
Jev is used as a hosted baseline for comparison. The README reports accuracy and ECE of Luce against Jev on identical test items, including where training does not help.
Primitives:choicescorenoul
Technique worth stealing
Train a LoRA + decision head on a small open model to output calibrated probabilities for typed questions.
Try it
pip install "git+https://github.com/scienthoon/luce" then run luce init, synth, train, eval, serve.
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricunclear0.51
- Shows a System One patternno0.35
- Handles uncertaintyno0.05
- Measuredno0.22
- Runnableyes0.62
- Worth recommendingno0.38
- Model replicayes0.96
- Problem scopescore on a 0–2 scale1.10
- About Jevyes0.90
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 21 Sept 2026.