Jev_Ontology
Research MVP pairing an LLM-authored, revisable ontology with Jev classification, where low confidence triggers ontology revisions.
Overview
Ontology + Jev explores combining TypeSafe's Jev with ontologies. Jev classifies items against an ontology; when classification fails or hedges, an LLM revises the ontology; the loop repeats. The author frames this as abduction in Peirce's sense: forming explanatory hypotheses for surprising facts. The project has since pivoted toward classifying images in a university library collection against a revisable taxonomy using the same LLM plus Jev cascade.
The ticket MVP was tested on 78 unique synthetic tickets across five sessions, totalling 190 classification passes and 380 Jev calls, at an estimated $0.0085 in input tokens. One reported example: low-confidence signals exposed a gap in the billing subtree, the LLM added a WrongfulCharge class, and three hedged tickets went from 0.56-0.68 to 1.000 confidence. The README is explicit that these findings are tentative, in-sample and not evidence of general accuracy gains. Offline unit tests and a paid live smoke test are included.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which ontology class does this ticket (or image) belong to? | — |
How it works
- An LLM authors an initial ontology
- Jev classifies items against it with confidence
- Low-confidence items flag possible ontology gaps
- The LLM revises the ontology and items are re-run
Strengths
- Uses confidence as a signal for taxonomy gaps
- Very low Jev cost for the experiments
- Provenance-checked, atomic experiment storage
Caveats
- Work in progress; findings are in-sample and tentative
- Synthetic tickets only for the MVP
- Focus has shifted to an image sub-project
Getting started
python -B -m unittest discover runs the offline suites; python test_jev_api.py is a paid live smoke check.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- Research scripts
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
We built a working MVP that pairs an LLM-authored ontology with Jev's calibrated classification, tested it against the live Jev API on 78 unique tickets across 5 sessions (86 classifications -- Session 4 re-runs Session 2's eight tickets), closed the feedback loop, and ran a 3-iteration convergence experiment. Total cost: $0.0085.
logicrw/awesome-jev-projectsTrying to combine Jev with ontology.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which ontology class an item belongs to, with a confidence that signals possible gaps.
Do revisions improve accuracy?
The README reports confidence changes on authoring examples but says this is not a measured accuracy gain.
What is the current focus?
Classifying library-collection images against a revisable taxonomy.
Related projects
ComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonMoves a coding agent's bulk semantic judgments (triage, tagging, routing) to Jev; an A/B reports 78% fewer tokens at similar accuracy.
ProjectbetaPythonGoogle Sheets custom functions JEV_IF, JEV_PROB, JEV_CHOICE and JEV_SCORE that classify, tag and score cells with Jev.
ProjectbetaJavaScriptMultiplayer GitHub triage dashboard where Jev answers kind, severity, urgency, duplicate and next step per issue.
AppbetaUnofficial TypeScript monorepo of Jev tools: a tool-call guard, a GitHub issue triage Action, a bulk JSONL/CSV sorter and a shared client.
LibrarybetaTypeScriptJag: Jev-inspired Qwen3.5-0.8B model that predicts a visual-grounding box in one forward pass, about 20x faster than its base.
ModelresearchPython