What it is
Turns "pick a threshold" into a measured step.
What it does: Fits per-question thresholds from labelled examples and shows how much you can automate.
Why it belongs in the showcase: Practical companion to any confidence-gated design.
Caveats: Synthetic demo data; needs your labels.
Evidence
Public MIT stdlib-only Python tool created 2026-10-18 with an explainer on GitHub Pages. README reports, on 120 synthetic shell commands with jev-1.13.0 (2026-10-18), that a destructive-command gate at 0.85 handled 89.2% of commands itself at precision 1.000, while the fitted 0.60 handled 98.3% at the same precision. Fitting 120 commands took 25 s and $0.002. `--demo` replays a recorded run with no key.
Caveats
- Results come from synthetic data; fit on your own labels
- Calibration is a property of a sample, not a guarantee on one answer
- Live mode sends each item's text to TypeSafe
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.