Mamba Hidden-State Norm as a Regime-Change Sensor in Time-Series Data

This record documents exploratory experiments on interpretingMamba hidden states as a *sensor* for regime changes in time-series data. The key observation is that the L2 norm of hidden states behaves as a smoothedrepresentation of the input signal, similar to a learned low-pass dynamical filter. An EWMA baseline is used to estimate slow regime trends, and deviations(delta between hidden norm and baseline) highlight abrupt regime transitions. This work is heuristic and exploratory in nature.The goal is not prediction accuracy, but building intuition and toolsfor monitoring regime changes in operational data. Future work will focus on normalization across items and basis-functioninterpretation of hidden dynamics (e.g., HiPPO-like structures).