Shannon entropy per 100-line window · as the session progresses
"Shannon entropy spikes in next-token probability distributions indicate model confusion and breakdown in multi-turn comprehension. When entropy exceeds baseline, it signals the onset of incoherence."
— ERGO, arXiv:2510.14077 (2025). Also: Braverman et al., "Calibration, Entropy Rates, and Memory in Language Models," PMLR 2020.