Jev Kit — FlorianRiquelme

A client answers whether Jev can be called, while the bench runs labeled JSONL through your real questions and returns accuracy, calibration, and threshold coverage.

Yes. The sample recipe run shows needs_human at 46.7 percent while other questions score high, which is exactly the compound judgment the docs flag as weak.

The digest reports a Gateway context of 32,000 tokens against a direct-API 64k, and this kit talks to the Gateway path only.

Jev cannot write text, explain, do arithmetic, count reliably, compare dates, or serve as a security boundary against injected content.

pnpm add @friqdev/jev-kit ai, then set AI_GATEWAY_API_KEY, which the AI SDK reads at call time and which stays server-side only. createJev() gives you evaluate(state, questions) returning answers plus costUsd and latencyMs. Node 22 or newer, MIT licensed, and the versions it was checked against are [email protected] (experimental_evaluate needs 7.0.105 or newer), @ai-sdk/[email protected] and @typesafe-ai/[email protected].

The PR finding triage recipe against its 15 fixtures, from a Hetzner EU server on 2026-09-20 against typesafe-ai/jev: p50 latency 809 ms and p95 1172 ms, total $0.000601 with a mean of $0.0000401 per call on 954.6 mean input tokens. Latency is measured from wherever the bench ran, not TypeSafe's US-West figures, and a separate three-call probe on the same setup read 504 ms, 315 ms and 272 ms.

Recipes/ holds real question sets from real projects with the source context left in a comment at the top of each file, to be read as documentation of question style rather than an API the package guarantees. pr-finding-triage re-scores a code-review finding's realness and severity from the finding text alone, because the reviewing LLM's self-reported confidence is not calibrated. source-triage routes an incoming source before ingestion. capture-classify comes from a German family inbox app.