JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Technical guide

Introducing JevEval: Jev-as-a-Judge for LLM Evaluation

Jeffrey Ip introduces JevEval: developers write bounded evaluation questions, Jev returns probabilities, and DeepEval calculates an inspectable weighted score with fixed math.

Original by Jeffrey IpEvaluationDeepEval blogSource reviewed

Before you dive in

What you’ll find in the original

  1. Separate evaluation logic, model decisions, and score calculation instead of asking one generative judge to do all three.
  2. Map Noul probabilities directly, normalize Score distributions over ordered levels, and assign explicit credit to Choice outcomes.
  3. Use strict mode when every applicable requirement must pass; otherwise a weighted mean allows strengths to offset weaknesses.
Worth knowing

JevEval still inherits Jev’s judgment errors and the quality of the questions you write; deterministic scoring does not make the underlying decisions correct.