AlphaDiana Measurement System for Verifiable Agent Reasoning

AlphaDiana provides a modular framework to evaluate agents by decoupling datasets, environments, and scorers. Testing reveals that agent harnesses do not consistently improve reasoning accuracy, as gains on unsolved problems are often offset by performance losses on easier tasks.