It is a Korean AI companion platform with 20 distinct avatars, persistent per-user memories, relationship progression, multimodal chat, voice and image experiences, safety controls, and scenario-based QA.

It is a Korean AI companion platform with 20 distinct avatars, persistent per-user memories, relationship progression, multimodal chat, voice and image experiences, safety controls, and scenario-based QA. I’d like to evaluate whether Jev can replace expensive prompt-and-parse evaluation steps with typed judgments for: • conversation safety and jailbreak detection • persona, memory, and location consistency • affection-score calibration • QA failure classification and release gating We plan to start with offline QA and shadow evaluation before using any result in production decisions. I’d appreciate feedback on how best to structure these judgments with Jev.