从发布线程进入 Jev 技术博客
Diogo 在发布线程末尾给出官方技术博客与早期访问入口,适合继续核对模型原理和产品说明。
该基准包含 900 条真实 Jev 决策,由 Opus 5 评分,一致率 85.8/100,分歧几乎都出现在 Jev 报告低置信度的地方,任何人都可对自己的模型复现。
Open weight builders will need a public yardstick. One went out today: 900 real Jev decisions graded by Opus 5, 85.8 of 100 agree, and the misses sit almost entirely where Jev reported low confidence. Anyone can rerun it against their own model: