What it is
A measured cheap-first cascade.
What it does: Escalates only low-confidence items and reports cost, latency and accuracy per stage.
Why it belongs in the showcase: Checks the usual cascade assumption instead of taking it on faith.
Caveats: Synthetic benchmark; small sample.
Evidence
Public MIT repo (0★ at intake) created 2026-10-18 with a GitHub Pages explainer and an offline --demo. README reports a 2026-10-18 run on 120 synthetic tickets with jev-1.13.0, then claude CLI sonnet on the 40 escalated: stage two 137 of 160 answers vs stage one's 131. It estimates stage one at about $0.029 per 1,000 tickets at $0.042 per million input tokens.
Caveats
- Synthetic data, small n; re-measure on your own traffic
- Stage-two cost depends on the model you plug in
- Sends ticket text to TypeSafe and to the stage-two model in live mode
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.