You need at least two completely independent teams who don't see each other's code so you'll have a rough idea how much developer bias is responsible for the signal. Right now they are all copying code from each other, a single error in these and all the climate models are effected. Replication here should mean different scientists writing different code arriving at the same estimates.

It needs to be open source, and it needs to be way more organized, no physical constants which aren't constant or which have bad precision or which have different values in different parts of the code.

Actual error modeling built in or at least a way to measure and quantify the accuracy.

Some models only use CO2 input for radiation and not for many other parameters that depend on CO2. Because the code is a mess and constants that shouldn't be constants are all over the place. And getting "constants" right is even more important than having high resolution to the simulation.

Getting real physical data to replace all the ad hoc constants is extremely important. And validating the correctness of the cloud modeling with real world data is also important.

And you can't write huge spaghetti code, it needs to be readable and comprehensible to someone other than original programmer so that people can validate the assumptions.