Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem
Guoming Ling, Muen Xue, Zijian Ye
cs.SE
2026-09-25
A census of 2,170 public Jev repos through 22 Sep 2026 finds 1,865 new repos and 43,750 stars in week one. Attribute judgment is in 77% of projects, yet routing takes 41.4% of stars.
Jev is TypeSafe AI's fast decision model, released 15 September 2026. Developers write a question in natural language and get a Choice, a binary Noul judgment, or a Score. It sits between task-specific classifiers and autoregressive LLMs: zero-shot, low latency, cheap enough for routing, filtering, and action selection. The public ecosystem grew quickly. Nobody had counted what those repos actually do, or whether stars follow project volume.
This paper is a census, not a model bake-off.
Collection has three stages. GitHub repository and code search by keywords, API endpoints, and SDK references, deduplicated on repo id. A GPT-6 Luna Max agent keeps a repo only when public code or docs show Jev used for a concrete task; name-drops and empty SDK wrappers are dropped. The same class of agent labels primary domain, decision purposes, and interface mix. A second agent independently reviews inclusion and domain labels, with disagreements resolved against the repo. Cutoff is 22 September 2026: 2,170 verified projects.
Application domain is what the project is for. Decision purpose is the role Jev plays inside it. A project may carry several purposes and several interfaces.
In the first week, 1,865 of the 2,170 repos were newly created and 305 were existing projects adding Jev. Those new repos took 43,750 stars in the same week. Growth continued after the spike.
No domain dominates project count. Content & Expert Tasks is 17.8%, Search & Memory 17.6%. Stars concentrate elsewhere: Routing & Automation takes 41.4% of all stars, Model & Tool Routing alone 32.6%.
On purposes, attribute judgment appears in 77% of projects, scoring or ranking in 52%, action selection in 31%. Among projects with an identified purpose, 69.7% use two or more. Interfaces: Choice 81.0%, Noul 72.2%, Score 45.4%, all three 36.8%. The authors read that as a product requirement: a general decision model should ship categorical selection, binary judgment, and scoring together.
Purpose mix shifts by domain. Action selection is 52% of purpose labels in Simulation & Control and 47% in Interface Agents. Model and tool selection is 27% in Routing & Automation and 6% in Software Engineering. Content filtering is 21% in Search & Memory; outcome judgment is 24% in Safety & Governance. Attribute judgment is still the leading purpose in six of eight categories.
Supply and attention diverge. Routing & Automation has 250 projects at 364 stars on average; Interface Agents 175 at 271. Together they are 19.6% of projects and 63.0% of stars. Content & Expert has 387 projects at 31 stars; Search & Memory 381 at 37. Simulation & Control has 252 projects, almost the same count as Routing, at 8 stars. Stars measure repo fame and can be pulled by a few hits. The paper treats this as uneven visibility, not unmet demand.
Appendix B walks seven repos (browser actions, review triage, tool-history compaction, a tool-call risk gate, per-turn model routing, a game controller, tax-form page typing) to show the same pattern: Jev emits a typed decision, and deterministic code owns execution, thresholds, and fallback.
For people building agent infrastructure, this is an early map: Jev is being used as a swappable decision part, not a chat model. Benchmarks that only score routing and interface agents will miss the high-count, low-star content and retrieval settings. Shipping all three interfaces is not decorative; developers stack them.
This is an observational report. It has no accuracy, latency, or cost numbers.
The snapshot ends on day seven; the mix may change. Only public GitHub is in scope. Inclusion and labels are LLM-agent work with no reported human agreement, so tutorial repos and implicit calls can leak either way. Stars are distorted by marketing and a few hits, and are not usage. Related work already notes Jev's sensitivity to option names; this census does not measure that.