What it is
poorjev is a local-first System One decision layer for LLM apps. It provides typed decisions (Choice, Score, Noul) with provably calibrated confidence, runs offline on commodity zero-shot NLI models, and requires no API key or waitlist.
How it uses Jev
poorjev reproduces Jev's Choice/Score/Noul interface. It takes a state and typed questions, runs one batched pass through a local zero-shot NLI model, then applies temperature scaling and conformal abstention to return calibrated probabilities and schema-valid answers.
Primitives:choicescorenoul
Technique worth stealing
Temperature scaling and conformal abstention on commodity zero-shot NLI models to achieve calibrated confidence.
Try it
pip install "poorjev[local]" then run examples or poorjev eval --set evalset/tasks.jsonl
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-centricyes0.88
- Shows a System One patternno0.39
- Handles uncertaintyno0.14
- Measuredunclear0.57
- Runnableyes0.92
- Worth recommendingno0.19
- Model replicayes0.94
- Problem scopescore on a 0–2 scale1.65
- About Jevyes0.91
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.