MEDJEV
Research implementation of Jev-style typed decisions over clinical evidence: supported, contradicted or unresolved.
Overview
MEDJEV is an independent, Jev-inspired System One research implementation for typed decisions over clinical evidence and biomedical text, from PAI-CUHK. It treats a clinical record, a proposed statement and an explicit set of candidate meanings as a programmable evidence-relationship query. The default semantic space is supported, contradicted or unresolved; custom candidate sets are accepted but treated as unvalidated unless matching calibration metadata exists. An experimental SLEEPJEV extension applies the pattern to sleep signals.
The design separates responsibilities that medical language systems often collapse into one generated answer: encode the evidence, encode candidate meanings at runtime, and score the relationship, returning explicit probabilities rather than prose. An interactive architecture demo and project site illustrate the pattern with one shared state and parallel Choice, Score and Noul questions. The README is emphatic that it is a research prototype, not a medical device, diagnostic system or source of clinical advice, must not be used with identifiable patient data or in patient care, and is not the official Jev or a reproduction of its weights.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this statement supported, contradicted or unresolved by the clinical evidence? | supported · contradicted · unresolved |
How it works
- Provide clinical text or literature as evidence
- State a proposition and candidate meanings
- Model scores the evidence relationship
- Return probabilities with calibration metadata
- Unvalidated candidate sets are flagged
Strengths
- Explicit evidence-relationship framing
- Calibration metadata gates custom label sets
- Auditable benchmarks
Caveats
- Research only; not for clinical use or patient data
- Custom candidate sets unvalidated
- Not the official Jev
Getting started
See the project site at pai-cuhk.github.io/MEDJEV and the repository documentation.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- research library
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Independent JEV-inspired System One-style typed decision research for clinical evidence, biomedical NLP, calibrated probabilities, and sleep signals.
hellogumbo/awesome-jev
Frequently asked questions
Can I use MEDJEV for patient care?
No, the README says it is not a medical device or source of clinical advice.
What does it decide?
Whether clinical evidence supports, contradicts or leaves unresolved a statement.
Is it TypeSafe's Jev?
No, it is an independent research implementation inspired by the pattern.
Related projects
Python harness stress-testing Jev's confidence, its accuracy as a RAGTruth hallucination judge, and its use as a RAG reranker.
DemoresearchPythonTwo frozen open models answer yes/no per exam option; their agreement drives an auto-release gate and a conformal candidate set.
ServiceresearchPythonEditorial QA tool where Jev judges per fact whether an edit preserved the original, with a confidence gate and human review fallback.
ProjectbetaJavaPython library that replaces hard-coded RAG thresholds with calibrated decisions for stopping, chunking, selection, abstention and caching.
ServicebetaPythonPre-registered study testing Jev's calibration claims on clinical diagnosis cases against Claude Opus 5 and Opus 5.5.
ProjectresearchPythonTLA+-verified consensus kernel around Jev that may escalate but never returns a confident wrong verdict, tested on 1,680 pharmacy cases.
ProjectresearchRust