In the current implementation it brakes posterior predictive samplings, because the observations are distributed within -pi, -pi and the posterior is shifted. Since mu can be a distribution itself, I’m not sure how to I can compare the posterior with the observations.