Model projections require data for verification
Coupled model intercomparison projects (e.g., CMIP4) aggregate many forecasts of future effects of CO₂. Often the average is usually considered the most probable scenario, but there is a significant problem with this approach:
Consider the dozen researchers who attempted to model ammonia ignition vs temperature. Their predictions were varied, but all generally trended in the same direction.
While it might seem reasonable to take the average of a dozen model predictions and suppose it is most likely to reflect reality, this may not be the case.
Averaging model predictions is akin to the "middle ground" logical fallacy: when several scientists suggest the number is high when several others suggest it is low, the attractive compromise must be "in the middle."
A better approach: take experimental measurements & compare with the model predictions. In the case of ammonia ignition delay time, we see that the ensemble average is not a good approximation of reality.
The best action is actually to disregard all but the two models that accurately predict experimental results.
The last 10 years of climate modeling on the world's fastest supercomputers has not produced agreement on cloud and albedo feedbacks. The correct model might be at one extreme or the other.
For Earth, the experimental data required to validate model predictions of CO₂ forcing is measurement of the changed infrared emission spectrum at top-of-atmosphere.
The changed infrared spectrum from Earth has been modeled many times. Examples: A) LOWTRAN (Charlock-1984) B) MODTRAN4 (Huang-2010) C) MODTRAN5 (Brindley-2016)