I have no idea whether I’m for or against this suggestion - feeling a bit ambivalent actually - but I think it is important to note that many users aren’t likely to value the integrity of the datapoints themselves as long as they can be used to steer the users towards their goals. Personally I like statistics and graphs, but I don’t really care that much about them, and I’d never use Beeminder without it solving actual problems for me aka getting things done.
I like the quantified self notion, and I’m a nerd^3, but still I can’t be truly bothered about actually collecting those data (this is where auto goals can come in) unless the goal itself requires me to do it, and this is where Beeminder excels. Besides the nice graphs, I feel that Beeminder have far too few tools to analyze the dataset to be noticably useful in that direction. Even the obvious average lifetime entered value rate is missing under the Goal stats heading. 
Fake values aren’t likely to cause major disruption to the data set either (if it does it is because the user is going nuclear on the goal anyway), but if the user feels that that aspect is something that will haunt them later, they’ll probably enter them with a comment that makes them identifiable and thus cleanable.