Semantic pub/sub experiment using TypeSafe AI's Jev System One model. Routes messages by meaning—one batched request evaluates all subscribers in parallel. Includes measured stability, batch degradation, and wording sensitivity tests.
该仓库以 TypeSafe AI 的 Jev / System One 模型为核心构建语义发布/订阅系统，真实调用 Jev API（TYPESAFE_API_KEY），提供实测成本（$0.000074/次、$0.064/全部实验）和延迟数据，并通过 M1/M2/M3 三项实验验证了 Jev 的稳定性、批量并行能力和措辞敏感性，属于实质性的集成测试与深入研究。
TypeSafeAITypeSafeAI/jev-harnessOfficialTypeSafeAI community repository implementing a Jev-based proposal-review workflow: an LLM proposes, Jev answers four narrow questions, code decides and leaves receipts; includes offline tests, routing experiments and a Next.js demo.
ekzhangekzhang/openjev-sglangJev-compatible API endpoint based on open models (prefill-only), implementing TypeSafe Jev/System One evaluation endpoints with Qwen3.6-35B-A3B on SGLang.
keltokhykeltokhy/jgrepjgrep is a CLI semantic grep that calls TypeSafe's Jev/System One decision model to judge lines or diff hunks by description, returning probabilities instead of generated text.
