JevBench maps typed-decision model tradeoffs.

JevBench’s strongest contribution is treating decision models as production components, where speed, confidence, cost, and reproducibility matter alongside raw accuracy.

• –GPT-5.6 Luna reached 97.1% accuracy, while Jev 1.13.0 scored 96.3%; overlapping confidence intervals make this a close comparison rather than a definitive podium.
• –Reporting five separate axes avoids hiding meaningful tradeoffs behind one composite score.
• –The benchmark measures typed outputs and probabilities, making calibration and integration reliability more relevant than prose-generation quality.
• –Its 242-item mix includes published, held-out, and imported decisions, but the small system pool and single-server setup limit how broadly results should be generalized.
• –The MIT-licensed harness and published decisions give developers a practical starting point for reproducing or extending the evaluation.