Climate models disagree with each other about how much the Earth will warm. The disagreement is large enough to matter for policy — the difference between “manageable” and “catastrophic.” Where does the disagreement come from?
Not CO2. We know how much CO2 is in the atmosphere. Not ocean circulation. We can model that reasonably well. Not ice sheets, not deforestation, not methane. All of those have uncertainty, but they're not the dominant source. The dominant source of disagreement between climate models is clouds.
George Matheou, a climate physicist, puts the sensitivity coefficient bluntly: if a model is off by 2 to 3 percent in cloud cover, the warming prediction changes by several degrees Celsius. More than half the variation between climate models traces to how they handle clouds. The ghost in the machine isn't exotic. It's visible out any window.
The problem is resolution. Clouds form at scales of meters to kilometers. Global climate simulations run at scales of tens to hundreds of kilometers. The clouds fall through the grid. To simulate clouds directly at global scale would require, by Tapio Schneider's estimate, a hundred billion times the computing power currently available. So modelers approximate: they add parameterized correction terms, adjusted by hand until the model matches historical data. The adjustment is partly science and partly intuition. Different intuitions produce different models. Different models produce different futures.
The binding constraint on climate prediction isn't the physics we don't know. It's the physics we know but can't compute. The Navier-Stokes equations describe cloud formation perfectly well. The equations are not the bottleneck. The bottleneck is that solving them at the necessary resolution, for the necessary domain, exceeds any plausible computational budget. The knowledge exists. The capacity doesn't.
This is a specific instance of a general pattern: the variable that matters most is the one you can't measure or compute well enough. Not the hardest one. Not the most exotic. The one at the wrong scale — too small for your simulation, too large for your experiment, too frequent for your sampling rate. The ghost is always the thing that falls between the resolution of your instruments. Clouds aren't mysterious. They're just the wrong size.