This goes to the qualities of the simulation. Which may
or may not be obvious.
One critical thing is, some 3-op-amp instrumentation
amplifier schemes have an issue with the A=2 amp
clipping / limiting and the A=-1 amp then is added to
a gimp signal and you get a wrong result, from signal
that is within the common mode input range of the
amps (but not common mode output range when
A>1).
A simulation using controlled sources and no limiting
function, will give you clean operation while a real
(or really well modeled at the individual amplifier level)
will show you the bind.
Now slight differences, like a 2*VIO2X-VIOn1X+VIOSUM
offset, may just be about real statistical scatter vs
models with no variability. Gain@VCM vs ideal simulation
model gain (towards the rails, block gain drops a lot).
The less your model looks like transistors with foundry
fully parameterized .model definitions, and more like
B sources / vcvs and passives, the more you are
going to be living in a fantasy land. How to tell, is on
you.