What it is
Jev decisions with an external self-improvement memory loop.
What it does: Runs lead-routing and related business demos where Jev chooses under a written policy, then compares policy-only, retrieval memory and SenseLab arms with rebuildable HTML reports.
Why it belongs in the showcase: Shows that Jev alone is not enough when tacit rules drift, and publishes negative and positive arms side by side with recomputable rows.
Caveats: Demo seeds; dual API keys for live mode; dollar figures are internal scoring.
Evidence
Public MIT repo with runner, demos, tests and results-sample. On the committed 120-lead routing run, policy-only Jev made 59 wrong decisions ($142,900 lost, 51% accuracy) versus 7 wrong ($12,900, 94%) with a top-8 retrieval memory and 6 wrong ($10,400, 95%) with SenseLab; the judge graded 119 of 120 decisions for $0.78. Multi-agent and 400-lead noise variants are also committed. CI rebuilds reports with --replay. The README explains Jev's input-token limit and why memory must live outside the prompt.
Caveats
- 0★ at intake
- Sample numbers are from author-committed demo seeds, not an independent field study
- Needs TypeSafe and SenseLab keys for live runs; replay mode needs none
- Business dollars are scored by the demo's own cost model
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.