Getting trapped was my concern. Now you just need to convince me that the additional constraint on later calculated values doesn't bias the result. I realise that where constraints are the same you can randomise the value calculation order to distribute the bias among the values from set to set. I presume however bias would still be a problem if the constraints were similar, but different? If randomising the value calculation order were used, would that generate bias in the generated sets?

It may be that I need to change technique depending on the disposition of the constraints. Use the original technique where the probability of choosing a good set randomly is high and switch to this technique (or kyle's code?) for troublesome constraint sets.