According to Anthropic on Twitter, the company measures 'incoherence' in AI models through a bias-variance decomposition of errors. In this framework, bias refers to consistent, systematic errors where the model reliably achieves the wrong goal, while variance refers to inconsistent and unpredictable mistakes. Anthropic defines incoherence as the proportion of total error attributed to variance, offering a quantitative approach to evaluating the unpredictability in AI model outputs. This methodology allows AI industry professionals to better assess and improve model reliability, as reported by Anthropic.