I spent some time this weekend looking more into the barycentric correction. I couldn’t figure out why the barycentric corrections from barycorr were so much worse than from the old barycentric correction code. It turned out it was because there was a bug with the CF structures — the modified julian dates in the CF structures were stored as doubles, but somewhere in the reduction pipeline they are treated as FLOATs, which only have precision to about the minute level. This corresponds to introducing an error at about the 2 m/s level.