This paper introduces a novel measurement framework to evaluate human-AI decision-making, shifting focus from mere model accuracy to the readiness of human-AI teams for safe and effective collaboration. It proposes a taxonomy of metrics and connects them to the Understand–Control–Improve lifecycle to assess calibration, error recovery, and governance in real-world deployments.

This paper challenges the common interpretation of AI models' performance on abstract reasoning benchmarks like ARC, hypothesizing that visual perception limitations, not reasoning deficiencies, are the primary bottleneck. It introduces a two-stage pipeline to separate perception and reasoning, revealing that most model failures stem from perception errors and demonstrating significant performance improvements.