In my own analysis of NeverWorld2 output (idealised configuration with a spherical grid), I found that the diagnostics using "pow2" were always off, especially when dealing with time-averages. One example is the diagnostic `e2` in the averages file which is the square of the interface height. Unfortunately, there are some issues with FMS, which lead to `e2` being off by a factor of (72/86400) in the 1/4 degree run, (192/86400) in the 1/8 degree run and (540/86400) in 1/16 degree run. I found that the ratio of the "off-factor" is inversely proportional to the the time-step in each run (DT=1200 in 1/4 degree, DT=450 in 1/8 degree, DT=160 in 1/16 degree).
I found this through trial-and-error so unfortunately don’t have a general solution that works in all cases.