Was the FWD radiosonde observation representative of its environment?
Were the aircraft observations representative of their environment?
What do we mean by "environment"?
Is the observed temperature equal to an average over a model grid box? For the global model, a grid box around the Dallas area has an average temperature somewhere between the cool values reported by the aircraft and the warmer FWD raob. Temperatures in the rain-cooled air could actually be close to the average value in a grid box in the 22-km Eta model, and the warmer FWD raob may also be close to the average temperature in a grid box in its location southwest of the rainband edge.
However, the model needs to make an analysis of features that it can resolve and for which it should be capable of predicting the future evolution. The analysis ensures this can happen by making only broad changes to the first guess forecast, preventing it from reproducing such sharp gradients. The 700 hPa cold pool was too horizontally narrow to be retained in the 22-km Eta model forecast anyway - it would end up being broadened and washed out over time. The 10-km threats runs and the 12-km operational runs to start in November 2001 may have a chance to include some mesoscale features like this if they could be well analyzed.
If the analysis could surgically remove features too small to be retained, how would that be done? One possibility is to ignore observations representative of the small-scale feature. The problem is being able to distinguish between observations representative of the larger scale and those not. Remember, in this case there were more observations near 700 hPa in the cool air in the rainband than in the warmer air around it, and the observations there agreed with each other. Another possiblity would be to make a "perfect", finely detailed analysis of the atmosphere and then filter it to scales the model can predict. But there is no perfect way to do this filtering, and data coverage is insufficient to make anything resembling such a complete fine-scale analysis.
Fortunately, as models move toward higher resolution, the representativeness problem shifts from mesoscale down to cloud scale. Then the bigger problem will become inadequate data coverage.