openJev Verdict 实测：修三个推理 bug 后 JevBench 74.9

Hemant @heman10x原推发布于 2026.09.22

在 X 上打开原推

People asking why Jev is blowing up when it is "just a transformer encoder" are missing what developers dealt with for the last three years. Teams spent dollars calling 70B causal decoders for simple routing. You waited two seconds for tokens to stream, hoped the JSON parser would not crash on a markdown block, and dealt with prompt order bias where permuting the choices flipped the decision up to 27% of the time. A bidirectional encoder evaluates all candidates simultaneously in one forward pass. Flip rates stay below 5%, inference takes 25ms on local CPU, and serialization delays disappear. TypeSafe AI took a bidirectional encoder, calibrated its output probabilities, and exposed it through typed primitives (Choice, Score, Noul). I tested how far that architecture goes with openJev Verdict, an open-source version of TypeSafe AI's Jev on a 151M base model. Without retraining, fixing three deterministic inference bugs brought it to 74.9 on JevBench (#2 globally, trailing closed Jev by 0.5 points and outscoring 4B decoders). The reaction is not about architectural novelty. It is relief that teams can finally stop using chat models as glorified if-else statements.