Jev 与 GPT-6 Luna 在 113 个决策上的校准对比
同一批 113 个决策，Jev 准确率不如 Luna，但只保留模型确信的答案后 Jev 升到 98%，Luna 仅 82%，重复 15 次 Jev 答案不变。
THE HIGHER SCORE ISN'T THE PART THAT MAKES AN AGENT SAFE Someone ran the same 113 decisions through Jev and GPT-6 Luna Luna got more right. 67% vs 81%. Ask it to think a little and it hit 99% Then they only kept the answers each model was sure about Jev's accuracy jumped to 98%. Luna barely moved. 82% Same input, 15 times: Jev never changed its answer. Luna changed on 5 of 8 borderline cases That's the split nobody is screenshotting A smarter model is better at picking. A calibrated model is better at refusing If the next step spends money, sends a message, or ships code, the useful output is sometimes nothing Accuracy is a demo metric Knowing when not to act is the product
这条是引用着另一条写的：x.com/rendaichi147380/status/2105615520284455260