No simple answer as to why. As stated above, likely a combination of satellite geometry and multipath. Network solutions always seem to be a bit more loose when compared to a nearby base/rover observation. To get points really tight between each other the best way in my experience is to observe both simultaneously for a good 5-15 minutes, with good PDOP window, which is where a planning utility comes in handy. Then post process against a base receiver in a favorable location not too far away. If you were to do your observations during a big spike in the PDOP and/or bad multipath, doesn't really matter what you do the results are going to be poor. By doing your observations simultaneously a good many of the random errors will be cancelled out. With RTK you're likely to find a good 1-2 cm. floating around. Just is what it is.
If I'm assuming correctly and your errors are referenced in feet then I don't see a problem with your results. RTK is only reliably good to a tenth of a foot horizontal. What was the time split between observations? The purpose of multiple observations is to use different constellations. A couple of minutes does not cut it. Two hours would be the minimum with four hours preferable.
Multi- path issues or network issues seem to be the two choices. If you have clear skies, then it would be network issues, no doubt there are many traps from user errors, setting errors, network anomalies, the list is probably long.
We don't see this with a base/rover setup; expect <.02' horizontally using R10's and if tight vertical is required the digital level is used. That's RTK control points, base/rover, a few minutes, open sky, never had a bad shot on a control point with RTK and the R10's.
Sure have seen lots of complaints about bad shots in a network.
I tend to agree with John.P. above, the figures you've stated convert to 12mm H 27mm V, while it's not great I've seen alot worse for network RTK, which seems to be the nature of the beast. Maybe you need to use a total station to achieve the accuracys you need.
How far were yo from the closest base?
Give full details. Even post the lat lon, and someone will help you with some of them.
Dittos on the onsight base...
A base within a mile or 2 improves it.
If you want/need high accuracy, set your rtk ctrl pnts at least 500' apart... 1000' is better. Set 3-4 ctrl around your site. Run tight traverse, doubled angles etc. Closed loop. Least squares, (or send to Kent). Then, develop a "best fit" between total station, and rtk.
Rtk is sloppier than total station... Ie, it's always got 0.02' to 0.07' error per shot, but, it's non- cumulative.
Total sta is tighter, but it cumulates.
When used TOGETHER, it can produce wonderful results.
Understanding your tools is essential.
Also, if you always point your gps pole the same direction, it tends to eliminate pole bubble error.
If your pole bubble is off 0.02' @ 2 meters height, and you observe ctrl station number 1, facing north, then ctrl station no. 2 facing south, you just threw 0.04' horizontal error between these 2 pnts.
Jus a sec... Kent is calling on Line 2....