I had a little start-up idea a while ago .. albeit only tangentially related to this one.
Make a deal with JC Decaux, or some similar out-of-home advertising company to place cameras (strategically) around the City of London.
Nominally to provide personally tailored advertising, the significant secondary purpose is to use face recognition to identify individuals-of-interest: specific traders, fund managers and so on.
This enables us to analyse facial expression, gait, maybe body temperature to determine mood, then look for correlations in the stocks and markets that these individuals trade.
I think that this will be legal, since all the information that you are using is (nominally, at least) legal, and gained in a public place.
After all, if it is OK for the authorities to place the whole population under close surveillance, they cannot possibly object if we turn around and do the same thing to their paymasters, can they?
The best type of test depends on the type of software being developed. For the sort of statistical software that I have been involved with, I think that system level tests (with synthetic and/or real data) give tremendous bang for the buck. This is particularly true if the data is high volume, relatively homogeneous (in some sense), and most of the top-level interfaces are fixed fairly early on. Many other projects are not like this, and so may benefit more (proportionally) from different approaches to testing.
I was very, very impressed by the translation - kudos goes to the Google translate engineers. If this high quality of translation can be achieved consistently, it marks a pretty significant step in human progress.