Question 1

Are those per-app accuracy figures independent results or the project's own calibration?

Accepted Answer

The project's own, on samples it assembled. The denominators do not line up with the sample counts: jev-proof reports 68 of 68 gated judgments correct from a 60-sample set, and jev-fit 115 of 115 from 60 samples, so the counts are per-judgment rather than per-sample. jev-fidelity reports 91 of 92 over 110 facts and jev-rental 0.910 gated against 0.854 raw. Nothing here is an independent replay.

Question 2

When the model is unreachable, does it fail open or closed?

Accepted Answer

Toward caution. An unreachable model sends the item to human review or escalation, never to auto-pass, and a check with no retrieved evidence fails closed because there is nothing to judge. The shared NoulGate makes the middle explicit as a three-way split: high-confidence pass, confident fail, and everything else to human review. That gate is the piece to read if you plan to reuse the kernel.

Question 3

Is this one tool or four separate ones I have to pick between?

Accepted Answer

Four apps over one shared kernel. jev-proof asks whether a sponsored video delivered the brief, jev-fit asks where a candidate falls short of the job requirements, jev-fidelity asks whether an edit preserved the information in the original, and jev-rental asks what must be confirmed with a listing agent. jev-kit supplies the JevClient interface with typed answers, the gate, and a sanitizer for text that came from outside.

Question 4

Is prompt injection actually handled, or just mentioned in a bullet?

Accepted Answer

Measured, at least on synthetic cases. jev-kit ships ExternalStringSanitizer for externally sourced strings, and each app reports injection cases in its calibration: 0 of 10 flips for jev-proof and jev-rental, 0 of 20 for jev-fit and jev-fidelity. Read those as a floor on the project's own test set rather than proof, and note every app keeps thresholds, routing matrices, and veto rules in plain Java.

Question 5

What four tools does the suite ship?

Accepted Answer

jev-proof checks whether a sponsored video delivered the brief, jev-fit shows where a candidate falls short of job requirements, jev-fidelity checks whether an edit preserved the original's information, and jev-rental lists what must be confirmed with a listing agent before viewing.

Question 6

What is in the shared jev-kit kernel?

Accepted Answer

A JevClient interface with typed Score, Choice and Noul answers plus a TypeSafe HTTP client that is 429/5xx/timeout-aware and a keyword-fixture mock for offline work, a three-way NoulGate (confident pass, confident fail, else human review), and an ExternalStringSanitizer for prompt-injection defense.

Question 7

What design rules do the apps follow?

Accepted Answer

Jev judges and code decides, with thresholds and routing in unit-tested code; the system degrades toward caution, so an unreachable model goes to review rather than auto-pass; evals refuse mock data and hard-fail without an API key; and no evidence means the check fails closed.

Jev Suite

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community