澄清 Jev 与 SalesRLAgent 架构差异
有人误把 2025 年 Nandakishor M. 的开源销售转化模型当成 Jev，作者澄清两者架构完全不同。
Fake news as usually The 2025 work by Nandakishor M. is real and open-sourced, but it is not the same architecture TypeSafe launched as Jev. Nanda’s March 2025 paper and Hugging Face model (SalesRLAgent) predict sales-conversion probability turn-by-turn. It takes a sales conversation, embeds it, then runs PPO to output a 0–1 conversion score. That is a task-specific classifier, not a general model you can ask arbitrary typed questions. Jev is a general System One model: you give it unstructured state plus typed questions (Choice, Score, Noul) and it returns calibrated probability distributions in one parallel pass, with no text generation. TypeSafe also has a new training method, RLCD.