jev-mode
Moves a coding agent's bulk semantic judgments (triage, tagging, routing) to Jev; an A/B reports 78% fewer tokens at similar accuracy.
Overview
jev-mode is a small dependency-free Python tool that cuts a coding agent's token use on bulk semantic judgments, such as which team owns a ticket, whether a file is stale, or which category a row belongs to. Instead of the agent reading every item into context, each item is sent to TypeSafe Jev with a map of typed questions, and answers come back constrained to the supplied options with a confidence.
The author A/B-tested it on 1,000 triage judgments against an in-context control: total tokens fell 77.8%, work-attributable input fell 16x, round trips dropped 65%, and department accuracy was 96.1% versus 93.7%, with Jev's own cost about four cents. The README is careful to call the honest result parity rather than superiority, since the control ranged 93.4-98.2% across runs, and notes the corpus was synthetic. It also shares findings on question design after an initial version lost 5.3 points of accuracy.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of six teams owns this ticket? | — |
| Yes / no | Is this file stale? | yes · no |
| Yes / no | Does this message report a bug? | yes · no |
How it works
- Install and run jev-mode check to verify the key and API
- The agent sends each item with a map of typed questions
- Jev returns constrained answers with confidence
- The agent keeps only compact verdicts in its context
Strengths
- Large token savings reported in a controlled A/B
- No dependencies; Python 3.8+ standard library
- Honest reporting of baseline variance and synthetic data
Caveats
- Benchmark corpus was synthetic; re-measure accuracy on your data
- Not yet on PyPI
- Question design strongly affects accuracy
Getting started
pip install git+https://github.com/ddfeyes/jev-mode, export TYPESAFE_API_KEY, then jev-mode checkAt a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- CLI / library
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
I kept watching coding agents burn context on decisions that aren't hard - triage 400 tickets, tag 600 files, route to one of six teams. jev-mode moves those verdicts to a typed-judgment model. I A/B'd it: 78% fewer tokens, 16x less work-attributable input, accuracy 96.1% vs 93.7%. Python, no deps, MIT.
logicrw/awesome-jev-projectsI kept watching coding agents burn context on decisions that aren't hard - triage 400 tickets, tag 600 files, route to one of six teams. jev-mode moves those verdicts to a typed-judgment model. I A/B'd it: 78% fewer tokens, 16x less work-attributable input, accuracy 96.1% vs 93.7%. Python, no deps, MIT.
hellogumbo/awesome-jev
Frequently asked questions
How much does it save?
The author's A/B on 1,000 judgments reports 77.8% fewer total tokens and 16x less work-attributable input.
Is it more accurate than the agent?
The README says the honest claim is parity; the control varied from 93.4% to 98.2% across runs.
Does it need dependencies?
No. Only Python 3.8+ and a TypeSafe key.
Related projects
Google 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.
AppbetaSupport-ticket triage pipeline where Jev scores every ticket and Python decides whether a general LLM should write a reply.
ProjectprototypePythonTyped, budget-aware language for LLM agents compiling to Python/TypeScript, with a decide{} block for Jev-family decisions.
ProjectprototypeNudge (compiles to Python and TypeScript)Research MVP pairing an LLM-authored, revisable ontology with Jev classification, where low confidence triggers ontology revisions.
ServiceresearchPythonUnofficial TypeScript monorepo of Jev tools: a tool-call guard, a GitHub issue triage Action, a bulk JSONL/CSV sorter and a shared client.
LibrarybetaTypeScript