What it is
Makes in-database AI labels stable enough to trust.
What it does: Stores, versions and calibrates Jev answers inside Postgres.
Why it belongs in the showcase: Measures the drift and over-confidence most demos ignore.
Caveats: Early phase; labels required.
Evidence
Public Apache-2.0 repo with CI, created 2026-10-19 (Phase 1 in progress). README reports Phase 0 on 2,000 CFPB complaint narratives judged by jev-1.13.0 via pg-jev for about $0.30: 9-way choice ECE 0.089, 100 labels remove 45–85% of calibration error, 4.8% of rows change top answer across 3 identical runs, and re-batching changes about 16.5% of answers. On Kubernetes issues routed to 10 SIGs it agreed with human labels 88.8%.
Caveats
- Phase 1 in progress; version comparison and a release are not built yet
- Needs about 100–200 labels per decision for calibration
- Live runs need your own TypeSafe key
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.