So the sampling approach described above involves sampling an increasing volume of data, but a true streaming SMC would never see the same data twice. A naive implementation of true SSMC would start with sampling the model given the data at the first timepoint, then use the resulting posterior as inits for a second sampling run given the data at the second timepoint, and so on. However, while the first sampling run should include the priors when incrementing the logp, I think the subsequent runs should be omitting the priors contribution to the logp (correct me if I’m wrong!), and I’m not quite sure how easy it is to achieve that in PyMC. I’ve tried defining my own distribution classes with a logp method that returns zero for use on the second-and-onwards sampling runs, but I suspect I’m doing something wrong as the performance in recovering data-generating parameters is terrible. Thoughts?