In the example above, I just use a Normal distribution as constraint and that is fine. Now another way of seeing the question is, what happens when my measurement process is more complex, skewed to one side. I think here it would look like a log-normal distribution mirrored between model and obs. I guess I am just having trouble to find the analytical parameters to the log-normal distribution that match the reflected log-normal, i.e. moving from pm.Lognormal.dist(mu=np.log(scale), sigma=s), q-loc) to pm.Lognormal('obs', ..., q, ..., observed=...)