Indeed that worked! At it produced probably the best results, again if I only applied it to the means (and not the precisions). Interestingly, the testvals chosen have a dramatic impact. I first set them -1 to 1, since the actual means in my code are symmetric about zero, and that perform VERY poorly and took a long time to run. Setting to 0 to 1, as you said, was about as fast as no sorting, and better results (better than tt.sort() as well). Then tried 0 to 2, results not as good and slower again. Obviously some math and/or code I don’t understand under the hood …