Tapdancer, I'm not sure if you've quite understood the situation. If I'm reading you correctly, the confusion arises from the two graphs below (left = leaked draft Figure 1.4 from AR5; right = final). The big problem with the former is that it was improperly baselined - set to go through the average temperature in 1990, rather than the position of the *trend* in that year. This is a very important distinction, since the temperature in any one year is the sum of (trend + noise), and no model can (or can ever be expected to) perfectly capture year-to-year noise. As it happens 1990 was a particularly high outlier above the trend, so baselining in this way artificially shifted the model outputs upwards relative to the data.
The final figure (on the right) is both correctly baselined, and changed to more clearly represent what's going on in the model-data comparison. Each line is an individual model run, and if you trace them you'll see that many have substantial periods of "flatness". The large variation is simply because most year-to-year phenomena ("weather") are chaotic, and the best we can ever hope to do is capture the statistical variation (which the models actually do quite well). When all the models are averaged, the result is the familiar steady upward trend.
Now, in any one year, by definition 5% of the models will be outside of the 95% confidence interval. But here's the thing that a lot of people miss is that they're not the same models each year. Some few might consistently "run hot" or "run cold" - but most of the population of those outliers is just natural variation taking particular model outliers temporarily to one side or another.
