WHAT NEW IN X11.06 FRI

AI collections from X · 23 collections · 69,408 posts

jevARTICLE

Jev as a voice agent judge: 27 of 34 metrics pass at 96%+

A Codex-written Jev judge covered 34 voice agent metrics, with 27 exceeding 96% on an untuned validation set at about $0.07 of reasoning per metric.

Neethu Mariam Joy
@DrNeethuJoyArchived 2026.09.27Originally posted 2026.09.26

What this article says

作者把 Jev 用作语音 agent 的自动评测裁判,流程是用 codex harness(GPT 5.6 Sol + Terra)为 34 个语音 agent 指标各写一个 Jev 裁判。结果是 27 个指标在从未参与调优的验证集上达到 96% 以上准确率,单个指标的推理成本约 0.07 美元;剩下 7 个指标没有收敛。文章的价值在于给出一个可复用的工作流:让代码模型生成裁判逻辑,再由 Jev 这类 System One 决策模型承担高频、廉价的判定,从而把语音 agent 的评测从人工听录音变成可批量跑的自动化环节。结论是 Jev 适合做这类结构化判定,但并非所有指标都能一次跑通,仍有约两成需要人工介入或换思路。

We used Jev as a judge on 34 voice agent metrics, with a codex harness(GPT 5.6 Sol+Terra) writing each judge. 27 hit 96%+ on a validation set it never tuned against, at about $0.07 of reasoning per metric. 7 didn't converge. Bookmark it for the weeked :) https://x.com/i/article/2103694312626356224

Neethu Mariam Joy @DrNeethuJoyOriginally posted 2026.09.26
Open original on X
x.com/drneethujoy/status/2103697740655956386Archived 2026.09.27 10:11