measurement-failures

This paper introduces 'evaluation blindness,' a formal framework for silent measurement failures that corrupt AI systems from training to deployment, with case studies, a failure taxonomy validated on 50 real incidents, and a failure budget framework.