Much of this thread is going in circles:
Thermocouples need temperature (not to mention some very good amplifiers).
Temperature of phase transitions depends on pressure.
You can use triple points of appropriate materials (water's is a hair over 0C), but then you have to prepare such samples... they also have to be extremely pure, chemically, otherwise the melting point will be slightly depressed, or the vapor phase will be depressed.
Electrochemical voltages depend on all these, once again. At least the voltage has a logarithmic dependency on temperature and concentration, but it's still sensitive.
And you need an extremely good thermal bath, no small feat of its own, and certainly not compact.
Semiconductors implement a voltage reference by using a difference of exponentials, which are ultimately related to the bandgap of the material. Which also varies with temperature, but in a way predictable for a given material; and another bandgap can be adjusted to give a signal proportional to temperature, rather than nominally canceling it out, and that signal used as compensation to Nth order.
Which, really, isn't much *better* than taking any half-assed method, measuring the contributing variables (like temperature), and applying a compensation to keep it straight. But it's notable that semiconductors are particularly well suited to this (the reference circuit can be made very independent of applied voltage, so that ambient temperature is basically the extent of it), and have decades of development behind them, perfecting these designs*.
*Except for the crap they threw in the ATMegas. Those things are shit. The ADC barely meets useful specs, the reference isn't, and POR/BOD are dV/dt sensitive. The stuff in the XMEGAs looks like they actually got someone who knows their head from their ass, but almost no specs have guaranteed ranges, and the documentation stinks.
Tim