“Unless you incorporate every single variable your model will fail by a greater and greater margin as you extend your timeline.”
Which is impossible with climate because it’s a non-linear coupled chaotic system.
IOW nobody can predict the future.
Don’t they feed the simulation results back into the input — for example: state 0 result are fed back into the same model as initial conditions that generated it to generate state 1 and the results from state 1 is fed back into the model to generate state 2.... etc.
The problem with this methodology is that any error in the previous states is amplified. Errors from model approximations, raw data having errors, etc. all contribute to the accumulated error until you have garbage.
If the errors are significant enough, the system will go unstable and oscillate wildly. So the data and results have to be interpreted and adjusted.