You have a good idea: Parallelizing op-amps is a technique used to improve performance - to minimize noise. Unfortunately, it doesn't really improve their precision - if anything, it usually worsens it, since it's progressively harder to keep all those op-amps at the same temperature.
For DC precision improvements, then the usual thing to do is to start with a precision op-amp, and then work to cancel the errors.
There are several DC error sources. The usual culprits include:
- gain error due to finite open-loop gain,
- gain error due to feedback circuit component ratios,
- finite CMRR,
- thermal voltage offsets in the input stage and in the input wiring,
- voltage offsets due to conversion of bias and offset currents into voltages.
For example, CMRR can be improved by orders of magnitude by bootstrapping the power supply of the op-amp. The finite open-loop gain errors can be improved by cascading two or even three stages. Their AC compensation must be done right to keep it from oscillating. Thermal offsets can be stabilized by putting the whole circuit in an oven. Bias and offset current issues are dealt with by having sufficiently low impedance input sources. And so on.
Even relatively lousy op-amps like TL071 can be cascaded and bootstrapped to yield a buffer that performs well enough at DC that you will have trouble measuring the errors without a 6.5 digit voltmeter at least. And even then - at 6.5 digits, the thermal errors from just the banana cable terminal temperature differences are easily seen. When microvolts are the last digit, just grabbing a short to the voltmeter input in your hands, so that it warms up a bit, is enough to see the drift! To work on such circuitry, it typically takes a climate controlled lab, and lots of time to get things to thermally stabilize. An oven for the op-amps helps too - but then it's tricky getting the signals into and out of the oven without introducing thermal errors...
Finally, note that static offset errors are trivial to calibrate out. What matters is how they change over time and ambient temperature. Even the best precision DC buffers have significant static offset errors until they get calibrated/trimmed. So, the SPICE model telling you that there is an offset error doesn't mean all that much: does this offset error change with voltage? Temperature? Aging?
Most SPICE models that you didn't pay real money for won't be useful at all for evaluating ultimate precision. For precision DC work, it usually takes real hardware to figure stuff out - or expensive simulation tools and models under NDA.
SPICE with typical public models can give you some idea of thermal sensitivities, if the models are suitably detailed, but it really takes transistor-level models to capture it all, and those are not very popular for obvious reasons. To a hobbyist who didn't win a big sum at a lottery, they are simply out of reach.
the NMOS is only accurate down to millivolts, & starts being inaccurate in the microvolts
The NMOS is way less "accurate" than that. That's why you're using an op-amp's gain to control the output voltage in a closed loop. This removes NMOS inaccuracy from the equation - mostly. The op-amp's DC open loop gain has a large impact on the accuracy, and for such applications you want the highest open-loop DC gains you can find. On the other hand, there will be some offset error that you'll need to trim out, as well as a temperature coefficient that you'd want to compensate in some way if necessary.
Remember that SPICE has limited accuracy: you're placing a lot of faith in simulation results that may well be garbage at the ppm levels you're interested in.
I would advise to start working with real circuits, and then try replicating their behavior in SPICE. You'll see that it's really easy to have SPICE models of something, but not necessarily anything close to what you have on the breadboard even though the schematic may look the same.
