Several independent efforts have been undertaken to create "climate quality" datasets of upper-air temperature changes from both radiosondes and satellites. Although these yield excellent agreement on short timescales they exhibit divergent long-term behaviour. It is this longer term behaviour that is key to policy makers.

At present we are restricted to making Animal Farm-esque "all datasets are equally likely" statements or resorting to at best semi-quantitative arguments to disregard certain solutions. The problem is most critical within the tropics where choice of dataset fundamentally alters the perception as to whether climate models are consistent with the observations.

This talk will summarise recent work using an automated system of climate dataset construction for radiosondes. Instead of taking three years to create a dataset (and sending the investigator(s) half crazy) we can now create a dataset inside a day. Of course, such a dataset will not be as high quality, but we have shown that sufficiently tuned the system can closely recreate HadAT, our current manually produced product.

What the automated system allows us to do is span the range of possible solutions by assessing the sensitivity of the result to a range of fundamental methodological assumptions such as whether, and if so how, to treat day and night launches seperately. Initial results suggest a large range of overall uncertainty, encompassing all currently available datasets, and systematic sensitivity to a relatively small number of decisions.

In this presentation we will focus upon whether using a range of pseudo-worlds created by adding plausible error structures to HadAM3 we can create a more restricted estimate of historical changes. The pseudo-worlds act as a quality check, narrowing down the range of plausible approaches and, at least potentially, as a result the distribution of possible historical changes. We will briefly touch upon the policy implications of our results.