Video unavailable. Open the source below.
NewJev Align
An open-source CLI helps developers tune Jev criteria using good and bad examples.
Added to Jevfast
How it uses Jev
Jev evaluates configured dataset rows with a typed AI Function and reports where its answers are uncertain. The CLI samples ambiguous and audit rows for human labels; GEPA proposes revised criteria, then the CLI re-evaluates the dataset. A person reviews each label and proposed change before accepting it.
What Jev decides
Illustrative Jev requestExample answers · not a recorded Jev response · Source ↗
Question 1 · alignment
YOUR APP
INSTRUCTION
How severe is this issue?
STATE
One row with text “Export crashes”.
JEV · SCORE
0
App workflow
- Evaluate a dataset
Run the configured Jev function over source rows and measure answer uncertainty.
- Collect labels
Select ambiguous rows and a random audit sample for human labels.
- Propose and compare
Use GEPA to propose a function revision, re-evaluate it, and show the score and definition diff for human review.
Why it is interesting
The useful loop is calibration against labeled edge cases: it turns uncertain Jev judgments into concrete examples for improving a portable function, with human review before any criterion change.