This works repeatedly. My main caveat is, at the end, you sample without noise. I think its because adding log odds noise to the log odds of the original probability vector is not symmetrical once you revert to probability space. However, if i sample noise straight in probability space, example pm.Normal("subj_noise", 0, .05, dims=("subject")), for infrequent probabilities, there is a levelling effect where you can only go so low below .1, while you can asymmetrically increase to 1.