Jev\_Onco\_Statistical\_Hierarchy
Rust research workbench where Jev gives a cancer-of-unknown-primary origin distribution over 22 classes, gated locally for review.
Overview
Jev Onco Statistical Hierarchy (JOSH) is a Rust workbench with a terminal UI and CLI for cancer-of-unknown-primary research, at version 0.8.0. It imports molecular, expression and clinical evidence, asks the hosted Jev model (jev-1.13.0) for an origin distribution, and then checks, gates, charts and exports the answer locally.
Each request (molecular-origin-v6) sends one Choice over 22 primary-site classes plus unresolved outcomes, two gating Nouls on whether the evidence is sufficient and whether it conflicts, and five boundary Nouls, all reading the same evidence independently. Local gates use the leading score, its margin, sufficiency and conflict to mark results abstained or review_required. An optional Shapley analysis re-runs the questions with feature groups masked. Live calls accept only declared synthetic data. The study question is inspired by OncoNPC, but JOSH contains none of its code, weights or data.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | primary_site: which of 22 origin classes (or unresolved) fits this evidence? | — |
| Yes / no | evidence_sufficient: is the evidence sufficient to call an origin? | yes · no |
| Yes / no | conflicting_evidence: does the evidence conflict? | yes · no |
How it works
- Import and validate sample evidence in Rust
- Send one request with a Choice and gating and boundary Nouls
- Validate model, answer keys and probability mass
- Apply local gates to abstain or require review
- Rank, chart and export; optionally compute Shapley attributions
Strengths
- Explicit abstention and review gates
- Offline prepare commands show requests before sending
- Saved runs keep their prompt version
- Modular crates and pinned toolchain
Caveats
- Research tool only; live calls accept synthetic data
- Relies on a hosted model for inference
- Not a clinical diagnostic
Getting started
./scripts/start (requires Cargo, Rust 1.98.1 and a C compiler), then open a bundled sample and run analysis in the TUI.At a glance
- Type
- Project
- Status
- research
- Language
- Rust
- Runtime
- TUI, CLI and loopback API
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
New prepares use \`molecular-origin-v6\`. The Choice, the two gated checks, and five boundary checks share the evidence and cannot read each other's answers. Boundary answers are reported with the result. Gates still use the leading score, its margin, sufficiency, and conflict. The parent of the leading class is a local lookup on the taxonomy. A request may use Jev's window: 64k tokens for the call, and 32k for the evidence plus the longest question.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide?
A distribution over 22 primary-site classes plus whether evidence is sufficient or conflicting.
Is this for clinical use?
No. It is a research workbench and live calls accept only declared synthetic data.
Does it include OncoNPC?
No. It is inspired by OncoNPC but contains none of its source, weights or data.
Related projects
Jev-inspired PyTorch research framework that encodes an overnight sleep recording once and answers many Choice, Noul and Score questions.
ServiceresearchPythonExperiment screening ADHD review abstracts as include or exclude with Jev, scored against Cohen et al. 2006 abstract triage labels.
AppresearchPythonNextflow plugin exposing Jev noul, choice and score judgments as pipeline functions you can gate on with ordinary operators.
PluginbetaGroovyResearch pipeline that scores NYC 311 complaint descriptions with Jev and maps severity and safety signals on a city grid.
ServiceresearchPythonMines timestamped driving scenes from vehicle logs by asking Jev lateral and longitudinal manoeuvre questions per moment.
ProjectbetaPythonFrench library ranking ADEME emission factors for an activity label, splitting code rules from Jev's semantic match.
ProjectprototypeJavaScript