JevBench runs a Jev-class typed decision model, one that answers choices, scores and yes/no questions with calibrated probabilities instead of generating text, through a shared set of tasks.
Results come back on five axes: how smart the model is, how cheap, how fast, how reliable its calibration is, and how open the model and its weights are, so different decision models can be compared on the same footing.
It is aimed at anyone choosing or building a Jev-class model who wants numbers instead of vibes.
What you get
- Full source code under the MIT license
- The JevBench v1 task suite and its five-axis scoring
Jev
TypeScript
Node.js
Claude Code
Codex
CodeRabbit Top10
Python
Go
OpenRouter