Helmut is not wrong; we're all saying the same thing. U(out) = Vbe(Q2) - Vbe(Q1). Helmut was just kind enough to elaborate on the temperature stability and such.
Measured, with real transistors.
Uo is already pretty "good," in the sense that the temperature coefficients cancel pretty well, but the voltage itself is small and strongly dependant on Vbe matching between the transistors. That makes Uo essentially unpredictable.
To square up a TTL waveform to precision levels, it looks to me.
Cheers, James