I think you missed it too. A drug company wants to know that their drug works. So when they go to the FDA they will get approved. If you have ever seen the raw data that the drug companies give you, you would probably be surprised. They pay you to prove that their drug works and that it will stand up to FDA scrutiny. Using the data that is given to you if you can't prove it you don't get a second gig. Basically if the drug company has hired you they think they can already prove it works. The reason we got it "wrong" (notice it is in quotes) was we were given the data without anything behind it. Just that it was a drug and you will be "presenting" this to the FDA. We weren't told a lot of things about it (on purpose) and weren't allowed to ask questions. Take a stab at it. The exercise was that regardless of what we saw and thought something was there. It may not have been significant to treat A, B and D but it was significant to treat C. We should have realized that the drug company wanted success not failure, not that the drug didn't "work" for A, B, and D but to say that it was working for C. Now it has been a while and I already told you I suck at composition...
I don't get it. In your examples you are still following well defined rules. Everything you described has rules. If you can shape or bend your data to fit those rules you are golden. Make the data fit your model. That is not fuzzy.