@pentavalentcarbon For the coupling constants, I think it makes it better. What typically happens is that a particular experimental coupling is 7 Hz. My lowest-energy conformer has 1 Hz. There is a slightly-higher-in-energy conformer with 9, so the averaged value is somewhere around 3.5. But obviously, so long as this conformer is not the ground state, the average cannot go above 5.

And, yes, pcJ-2 for the couplings; for the shifts cc-pVTZ, but because I'm doing linear scaling on the shifts, the performance across a wide bunch of basis sets seems to be pretty uniform.

The DP4 results aren't that bad, actually (I got the script working today). The molecule above has three stereocentres, so there are four possible diastereomers. Out of the 24 possible assignments (four experimental data sets ==> four diastereomers), the correct one is predicted with 55% certainty, which is not particularly high, but is higher than all the other wrong assignments. I think it might help to just try it on more data points.