What it is
Replaces fragile 'return the indices to keep' prompts with numbers.
What it does: Wraps Jev questions in gates with declared failure policies and usage records.
Why it belongs in the showcase: Production-derived gates with their audit numbers.
Caveats: Tune thresholds for your own data.
Evidence
Public MIT repo, created 2026-10-19; peer dependency @typesafe-ai/sdk 0.6 or later. Seven reference gates (relevance, source relevance, candidate, locality, menu page, inbound screen, fact check) came from a production newsletter platform. The README reports audit numbers, for example: a menu-page heuristic was about 59% right over 371 pages, while Jev agreed with the extraction outcome about 87% of the time.
Caveats
- Thresholds come from one platform's audits; re-check them on your data
- Live calls need a TypeSafe key; without one the judge returns no_client
- The inbound screen gate is fail-closed by design
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.