autoregressive state transitions, where predictions are conditioned on past
predictions, are the predominant choice for both deterministic and stochastic
sequential models. However, autoregressive feedback exposes the evolution of
the hidden state trajectory to potential biases from we