PrimaCons Benchmark Evaluates Decision Consistency in Jev Primitives
October 27, 2026
The PrimaCons benchmark reveals systematic decision-making inconsistencies in Jev when using different query formats, such as binary versus multiple-choice. The study finds that these discrepancies can predict decision failure with a ROC-AUC of 0.91, proposing a Cross-Primitive Residual Calibrator to mitigate errors.
HOW THIS AFFECTS YOU
●
builderYou should account for query format sensitivity when designing LLM-based decision engines.
●
researcherThis highlights a significant reliability gap in how LLMs process semantic states across different primitives.