The current jointcal tests either use a perfect reference catalog in the fits, and then compare with it, or use an imperfect refcat in the fits, and also compare with it. We believe that much of joincal's improvement is due to its "improving" an imperfect reference catalog, and thus these tests don't tell us whether our understanding is correct, nor whether it really "works" at all, or if it's just over-fitting.

In order to confirm that jointcal really is causing an improvement by "constraining the reference catalog", we need to compare a perfect reference catalog with the single frame and jointcal astrometric fits performed on a non-perfect reference catalog. Gaia is as closed to perfection as we've got, and the validation datasets have less perfect refcats already.

Steps:

1. Download a 3-4 degree radius Gaia catalog at the location of the 4-visit validation_data_hsc data.
2. Take the jointcal results for the 4-visit HSC validation_data and run plot_jointcal_results.py using the above Gaia data as the reference catalog.
3. If jointcal "works" as we understand it to, the "absolute astrometry" (computed against Gaia's perfection) should improve with the post-jointcal fit.

If 4 visits aren't enough, another possibility would be to reprocess twinkles with non-perfect refcats and then do the same comparison with the known-perfect refcat.