Good modelling – the kind done carefully before cheap computers created the illusion of precision – stripped a system down to its core elements, made sure there were more experimental observables than fitting parameters, and produced outputs that inferred meaning to humans. You could feel the mechanism in the equations. The fitted parameters meant something that you could test and measure independently.

Modern climate modelling isn’t this. The models are complex, the historical dataset is tiny and the parameter space is vastly underdetermined, and the results are over-fitted to fuck. Worse, little models get stacked on top of each other without anyone asking what changes at each interface. At every handoff, uncertainty gets laundered silently while the dependent errors build up.

The solution is to stay high level and model a “model” system – a simplified but physically faithful representation stripped to its essential structure. The art is knowing what to leave out. If you can’t explain the dominant behaviour with a handful of terms, you don’t understand the system.

The tragedy is that this modelling discipline is largely gone. Complex models are politically defensible – you can point to the detail as evidence of rigour. Simple models expose their assumptions and are vulnerable to scrutiny. So the incentive runs the wrong way. Complexity protects the modeller. Simplicity exposes them.

At one point the climate scientists went full Cassandra – shifting from “here is what the physics suggests and here is our uncertainty” to “this will happen and you must act now.” Honest uncertainty is more persuasive over the long run than false precision in the service of advocacy. They confused the science with the politics and ended up damaging both. When the precise predictions are missed, the credibility damage spreads back to the underlying science, which may have been perfectly sound. Or not.