SelectStar launches DAVI judgment AI, claims 89.2% accuracy
SelectStar's new DAVI model outperformed TypeSafeAI's Jev on 231 JevBench questions and cut model execution costs by about 28% through routing.
SelectStar has unveiled DAVI, an AI model specialized for judgment tasks, the company announced on October 26, 2026. According to digitaltoday.co.kr, DAVI is designed to select an appropriate AI model based on question difficulty and to evaluate whether AI answers are appropriate.
In a comparison with TypeSafeAI's Jev on 231 JevBench questions, DAVI scored 89.2% accuracy versus Jev's 86.6%. On high-difficulty problems, DAVI reached 77.5% against Jev's 73.0%, and on Korean-language feedback evaluation it scored 72.5% versus 66.8%.
DAVI can also predict the likelihood that an AI agent will succeed at a task. In a routing experiment on 870 hard Korean question-answer pairs, DAVI-based routing achieved 71.1% accuracy, compared with 70.3% for using Anthropic's Opus model alone.
Model execution costs with DAVI-based routing were about 28% lower than using Opus alone, according to the company.
SelectStar plans to apply DAVI to its AI reliability evaluation platform Datumo Eval in November.
Quick answers
What is DAVI?
DAVI is an AI model from SelectStar specialized for judgment tasks such as selecting an appropriate AI model based on question difficulty or evaluating the appropriateness of AI answers.
How did DAVI perform against Jev?
On 231 JevBench questions, DAVI scored 89.2% accuracy versus Jev's 86.6%. On high-difficulty problems it scored 77.5% versus 73.0%, and on Korean-language feedback evaluation 72.5% versus 66.8%.
When will DAVI be used in Datumo Eval?
SelectStar plans to apply DAVI to its AI reliability evaluation platform Datumo Eval in November.