Thoretical question: Can you do secure computation on a non-secure base?
In practical terms think of Amazon's compute cloud, or the Cern LHc Grid. There's a job submission and results interface, but how could you trust the results?
The IBM Mainframe's are possibly the best example we have where they use two cores in parallel to execute each instruction, and if a difference is found a third core is called in to break the tie.
On a larger scale that seems to be only solution, although a random partial overlap of results might be sufficient. Of course that only works if the data & results aren't sensitive.
Got any ideas?