THE JEV GUIDE · CURATED EDITION
Launch news, explanations, demos, and community projects around TypeSafe Jev, organized by topic. Every entry links back to its X source.
Curated selection · Not exhaustive · Updated manually
Sponsors
These products pay for the directory. Want yours up here?
The post discusses the need to set judgment criteria for Jev, questions whether bots could judge by themselves if criteria exist, and notes the benefit of abstraction to reduce changes when criteria change.
Demonstrates Jev automatically selecting and placing furniture in AI-native software, reflecting AI decision-making.
PageIndex announces the Page Search API, claiming it is faster and more accurate than SOTA embedding models, and discusses the Jev decision model's retrieval performance for RAG. Link
The Fulvwa plugin uses account comment history combined with the Jev model to detect AI-generated content. The plugin is under review and will be shared once approved.
A developer built a plugin combining Claude Code with Jev. It suggests and draws diagrams after long responses, supporting 14 diagram types and light/dark themes.
One card in the feed, every sixteen, on every page of it. $50 a week, cancel any week.
A user's test shows that Jev returns mid-range probabilities (0.1–0.9) in about 50% of binary decisions, while Decisions API only does so in 22%, mostly giving decisive values like 0.02 or 0.98.
The author points out that the so-called self-improving playtesting agent is just an LLM writing a game bot, fixing it after errors, essentially a Claude Code wrapper running on a Jev model key, and promotes their own game QA tool.
A user shares combining Jev with Grok 4.7: Grok handles reasoning, Jev handles decisions. Jev does not generate text; it only answers via yes/no, choice, and scoring, so it can only decide from allowed answers.
This visual novel game supports free input in addition to regular choices. Internally, it uses SystemOneModels defined after Jev's debut, which constrain output to numeric values, making them cheaper and faster.
The author shares a small project integrating Jev into a personal budget app for automatic transaction categorization, saying it is more accurate than regex and cheaper than LLMs.
The hard part of a supervised agent is how often to ask the human: too often and they stop reading, too rarely and something slips through. Jev gives a number from 0 to 1: above 0.9 the agent proceeds, below 0.1 it stops, and in between it asks.
Nace AI has released Drex Max v1 weights on Hugging Face: 2.6B? Wait, the post says 260亿参数 (26 billion parameters), bf16 about 52GB. The model does not generate text; given a state and questions, it outputs probabilities for each option in one forward pass, suitable for local routing. The model card compares it with Jev 1.13, scoring 67.52.
