jev-survey
Preprint survey of Jev-style typed decision models: their research roots, adaptation costs and what current evidence supports.
This repository is mostly documentation or write-up rather than runnable code.
Overview
A Survey on Jev is a preprint (version 0.5, October 2026) by researchers at Nokia that reviews typed decision models, also called Jev-style or System One models. These take a state and a call-time question and return a probability distribution over a closed answer set instead of generating text. TypeSafe offers this interface in Jev, and several open implementations reproduce it.
The survey traces the ideas to two research areas: routing research, which supplies the goal of a decision component that adapts without retraining, and mechanisms such as zero-shot classifiers, safety guards and label-probability readout. It then examines adaptation costs for new candidate models (not free, since routing still needs candidate-specific evidence), new decision types (specifiable without retraining) and new deployment constraints like target error rates (still need labelled data). Evidence is mixed: valid outputs and large cost savings versus LLM judges, but lower accuracy than frontier LLMs on hard tasks, variable calibration and context sensitivity. It recommends testing against a label-probability readout and a cheap generative cascade under the same acceptance rule.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Studies how typed decision models adapt to new models, new decisions and new constraints, and what evidence supports their use | — |
How it works
- Define typed decision models and their interface
- Trace prior work in routing and zero-shot mechanisms
- Analyse adaptation costs for three kinds of change
- Review empirical evidence and recommend baseline comparisons
Strengths
- Balanced review of strengths and weaknesses
- Concrete adoption advice: compare against two simpler baselines
Caveats
- Preprint, not peer reviewed
- Survey of others' evidence rather than new experiments
At a glance
- Type
- Model
- Status
- research
- Has code
- no
- Has tests
- no
- Stars
- 0
Use cases and integrations
How curated lists describe it
A Survey on Jev: Prior Research, Adaptation Costs, and Evidence of System One Models (preprint, v0.5)
wh000wh000/awesome-jev-live
Frequently asked questions
What does the survey conclude?
Typed models give valid outputs and big savings versus LLM judges but trail frontier LLMs on hard tasks and vary in calibration.
What does it recommend before adopting Jev?
Testing against a label-probability readout on an open model and a cheap generative cascade, using the same acceptance rule as deployment.
Where can I read it?
A PDF and paper page are linked from the repository on GitHub Pages.
Related projects
Korean study notes and TypeScript mock-ups of Jev concepts, with a local Laya versus GPT-5 comparison and a calibration toolkit.
ProjectresearchTypeScriptLearning project that builds choose-an-answer decision models from scratch and from Qwen, with downloadable 0.8B and 4B releases.
LibraryresearchPythonInteractive web workbench for exploring and benchmarking Jev's Choice, Noul and Score via OpenRouter, with threshold transfer curves.
ProjectprototypeTypeScript- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython 33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonJoke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPython