Residual-Gated Neural Regime Switching

Integrate the detector around a recurrent state-space model, latent ODE, or model-based RL world model. Maintain $K$ transition heads $f_{\theta,1},\ldots,f_{\theta,K}$ sharing an encoder and observation decoder; each head represents a possible regime such as nominal dynamics, actuator degradation, sensor bias, or altered environment physics. At every observation, retain the previous latent state and roll out all heads for one step or a short horizon. For each mode, compute $\hat y_k=h(f_k(z_t,u_t))$ and residual $r_k=y_{t+1}-\hat y_k$; solve the ridge disturbance estimate $\hat q_k=(J_k^\top\Sigma_k^{-1}J_k+\lambda_q I)^{-1}J_k^\top\Sigma_k^{-1}r_k$; recompute the disturbance-corrected score over a FIFO horizon; choose the lowest-score mode only if it beats the runner-up by threshold $\tau$, otherwise retain the previous mode or use a softmax blend. Estimate $\Sigma_k$ and $J_k$ from nominal validation trajectories, and set $\tau$ to the 99th percentile of score gaps under no regime change. Train the heads with prediction loss plus a small disturbance penalty. First test on a 2-D Lorenz or cart-pole simulator with one abrupt parameter change and additive Gaussian noise. Compare a single-head GRU, an oracle switched model, and this detector while sweeping change magnitude and noise. The quantitative prediction is a sharp detection boundary when the whitened residual gap exceeds $\tau$; false alarms should remain near 1%, and post-detection five-step prediction error should approach the oracle error within the detection delay.