Jev Delivers 94ms Field Photo Judgments

Jev’s appeal is becoming clear: it turns narrow visual checks into cheap, typed decisions instead of expensive vision-language conversations. The results are promising, but still early evidence rather than production proof.

• –Jev caught 32 of 35 unreadable photos and falsely flagged only 2 of 105 readable ones.
• –Sol and Opus caught all unreadable photos but produced five times as many false positives.
• –One Jev request answered both readability questions, improving throughput and lowering cost.
• –The evaluation used training data, not a held-out test set, so real-world accuracy may differ.
• –Jev’s image support remains experimental, making human escalation and confidence thresholds essential.