Jev's first week shows adoption spread wide while attention clusters on routing

Jev is a decision model with three typed interfaces—Choice for selecting among options, Noul for binary judgments, and Score for rating inputs against predefined levels—designed to sit between the latency of full LLM generation and the rigidity of supervised classifiers. This paper maps its first week of public life: 1,865 new GitHub repositories and 43,750 stars in seven days after a September 15, 2026 release, plus 305 existing repositories that integrated it.

The analysis covers 2,170 verified public projects, annotated by GPT-6 Luna Max agents with a second-agent review pass for disagreements. The domain spread is genuinely wide—the two largest categories, Content & Expert Tasks and Search & Memory, hold only 17.8% and 17.6% of projects respectively. No single use case dominates supply. Attention is another matter entirely: Routing & Automation pulls 41.4% of all stars despite representing a fraction of total projects, and Model & Tool Routing alone accounts for 32.6%. Routing & Automation and Simulation & Control have nearly identical project counts (250 and 252), yet average 364 and 8 stars per project respectively. Stars measure repository visibility, not unmet need, and the paper is careful to say so.

The usage patterns are the more practically useful finding. Attribute judgment appears in 77% of projects; scoring or ranking in 52%; action selection in 31%. Nearly 70% of projects with an identified purpose use Jev for two or more purposes simultaneously, and 36.8% deploy all three interfaces together. The model is not being used as a single-purpose classifier—it is being wired into workflows that need several different decision types at once.

The appendix case studies make the integration pattern concrete. A browser agent uses Choice to pick the next DOM operation and delegates text generation to a separate small model. A code-review triage tool chains Noul risk signals into Choice evidence selection and then Score severity estimation. A context-compaction system uses two Noul keep-probabilities to decide whether each tool call and its result should survive into the next agent turn—no summarization, just retention decisions. A tool-call risk gate uses Noul and Score to flag potentially destructive operations before execution, with the surrounding policy deciding whether to warn or block.

The paper's own limitations are stated plainly: this is a snapshot of public GitHub repositories only, taken one week after release. Private and commercial deployments are invisible. The ecosystem will shift.

What the data actually argues is that Jev functions as a reusable judgment layer whose role is determined by the surrounding workflow rather than by the model itself. The design implication the authors draw—that general-purpose decision models should support all three output types and be evaluated across the full range of application contexts, not just the high-visibility routing cases—follows directly from the mismatch between where projects are and where attention lands.