Week 2 test results: JEV by TypeSafe.ai
· Zi Wang · 1 min read
9x faster · 92% agreement · 97% coverage
This week I rebuilt Ask Yeah on JEV from TypeSafe.ai and ran it head to head against Sonnet 5 on live traffic from 2,100 web visitors.
Median latency
JEV ██ 173 ms
Sonnet 5 ████████████████████ 1,613 ms
Same decision as Sonnet 5
█████████░ 92% (1,931 of 2,108)
Answer coverage · 2,183 probes
Before ███░░░░░░░ 27%
After ██████████ 97%
Wrong-shoe answers
Before █████████ 9%
After ▏ 0.4%
I also extended JEV beyond interpreting questions. It now reads 12,132 YouTube reviews and combines its typed judgments with deterministic reviewer counts and citations.