For a scope or spectrumanalyser it is a different thing. This because the accuracy of those things is relative low. A good scope is 2% so amplitude calibration only needs a 0.2 % reference and that is not hard to add. For frequency the same thing, a normal cristal oscillator will be a few magnitudes better as the timebase accuracy. In a Tek 271X SA they use an ovenised Xtal and a TL431 as amplitude reference. More then enough.
A multimeter needs to have a voltreference (or older types a frequency reference for the VF converter)
This Vref can be a simple resistor divider (I have seen two cheap DMM using this) or something like a LTZ1000 and there is a lot in between. The Vref here dictates for the major part the total performance. But you could make a construction that used very cheap unstable components and use the meter to measure its own reference in the hope that is a bit better as the Vref and so compensate the crappy components. But normally when you use a good reference, you use good parts for the rest and when you save on parts you do not throw in a more expensive Vref (that also needs a better design pcb etc)
And for calibration the standard needs to be around 10x better as the internal reference. Besides that, that reference needs to be aged enough (many people forget that) and calibrated. You need to know its value. Even a LM399 has a poor initial accuracy, but the value it is after burning in stays very stable and that is important. It could be used for some sort of self check. But in that case you must asume the Vref is the most stable part and the drift is only caused by things like resistors.
I have seen Vref designs where they used the Vref for a sort of self calibration. I do not remember exact how but is was rather clever, a sort of ratio measurement and bridge like setup. If both test points had the right ratio the output should be 10V. I once made that and it worked but the Vref (1n821 if i remember well) used had a very bad tempco and I could not find the sweetspot, so you kept adjusting it.
a separate Vref as in house standard is nice but never forget thwe initial accuracy, the Rout, load it can handle and calculate the numbers. A LM399 on its own is not usefull but combined with a good chopperamp, good resistors and a buffer can make a good reference. Problem is you need to calibrate it. Even the standards at nist need calibration