Control errors using CRTN
Apparently, RAW CRTN data isn't as reliable as I thought.
It's unlikely that the RTN data are the problem. Possible, but unlikely.
It sounds like you held the positions of two recovered monuments, set some monuments from those, then calibrated/localized to your set monuments.
By the time you got to staking, you were stacking the absolute positional error of those recovered mons (vs record) and the standard errors of the RTK vectors for your observations of same, on top of the RTK errors of your set mons against their positional error versus your calcs.
The localization is always going to warp things to fit, and my guess is that somewhere along the line something got out of whack enough to really warp your solution.
I rarely localize to calculated positions; if I do it's a single point localization for the sole purpose of improving my search for monuments. I won't stake based upon a field localization unless I (a) have terrestrial grid coordinates that are real world rather than calculated, and (b) I was the one who observed them conventionally.
Simply use a projection, if it doesn't match each day then the RTN isn't capable of doing the survey to the precision needed.