FirstRead / Technical appendix A1

Why Jev

Yes/no probabilities that rise with real risk, so a cutoff means something.
14 questions about each message, answered in one call.
About 0.5s a check (median), and $0.04 per 1,000 user checks (recorded estimate).

0.992Jev's AUC

Best of the other four: glm-5.3-flash, 0.976.

Model	AUC	Recall at 5%
Jev	0.992	96%
gpt-6-luna	0.933	88%
gemini-3.8-flash	0.973	92%
deepseek-v4.1-flash	0.966	91%
glm-5.3-flash	0.976	93%

Jev gives an answer in between far more often: 23% of its answers are 0 or 1, against 64%–98% for the others. Luna is read from its token probabilities; the flash models state one. BHH test split, 1,750 messages, all answered by every model.

0.50sJev's median check

Fastest of the other four: gpt-6-luna, 1.26s, 2.5× as long.

Seconds per check on the BHH test messages: the dot is the median, the bar the middle 80% of checks (10th to 90th percentile). FirstRead waits 3s for the user check. gemini-3.8-flash takes longer on 41% of messages and deepseek-v4.1-flash on 94%.