Jev confidence tested on 11,700 decisions for $0.59
The author open-sourced all predictions and charts from a test of Jev on 11,700 decisions, finding that when it says it is 90% sure it is right 81.6% to 99.8% of the time depending on the task.
What this article says
作者对 Typesafe 的 Jev(System One 决策模型)做了一次规模化的置信度校准测试:共 11,700 次决策,总成本仅 0.59 美元。核心发现是模型的自我报告置信度并不总是可靠——当 Jev 声称「90% 确定」时,实际正确率会因任务类型在 81.6% 到 99.8% 之间浮动,也就是说置信度需要按场景分别校准,不能直接当作概率使用。作者公开了全部代码、每条预测记录和图表,供他人复现与进一步分析。结论指向一个实用建议:把 Jev 接入 agent 或交易流程时,应针对自己的任务分布先跑一轮校准,再据此设定阈值或路由规则,而不是照搬默认的置信度门槛。
I tested Jev (@typesafeai) on 11,700 decisions. Cost: $0.59.
When it says "90% sure" it's right 81.6% to 99.8% of the time, depending on what you ask.
Code, every prediction and the charts are open source. Link inside. https://x.com/i/article/2101632070950764545