It is getting late, so this may be way off, and have not the time to read the paper.

This is also assuming finite trials right? Because over infinite trials if you have a non-zero probability of siding with the wrong group of classifiers, you will make infinite mistakes. No matter how small the probabilities go.

It seems it is trading off a better expectation for a worse real worse case.

Also are you thinking of formalising an alternative to the infinite SI worst case you describe?