Visible CoT can hide how much reasoning stays latent
A model can produce readable chain-of-thought while doing much of its serial work in hidden states. Redwood Research proposes an operational proxy, opaque serial depth, based on the longest stretch of computation without a natural-language-rooted bottleneck. The framework gives evaluators a way to compare architectures and ask a harder safety question: how much of a model’s reasoning remains inspectable after architectural changes?