Jev 与文本生成模型的取舍
Jev 不生成自由文本。Diogo 用一段对比视频说明并行决策模型与逐字生成模型在任务形态上的差别。

在人工复核错误率上限为1%的条件下,Jev 允许93%的答案跳过人工检查,OpenAI 为66%,大型通用模型虽可达98%,但速度慢约10倍。
The number that matters we think: how many answers can skip a human check? With each model's cut-off set so at most 1 in 100 let through are wrong, Jev lets 93% through and OpenAI's 66%. Big general models reach 98%, but they're 10× slower.