You would just get people massaging or generating the data to fit a conclusion prior to sending it out for statistical analysis. If people want to cheat they can find a way. The only way to really do what you are getting at is to construct core labs or CROs to run all experiments on behalf of the investigators. This is not infeasible in many cases (and is already done in narrow ways) but it requires hiring staff scientists to run every experiment rather than grad students or post docs and costs / complexity will explode.
The real way to defend against a lot of fraud is to force people to submit actually detailed methods sections so experiments are legitimately reproducible (they largely aren't now). This would catch a lot more fraud quickly, although even this won't fully work as some experiments are simply too costly to reproduce for validation purposes (eg animal studies).
The real way to defend against a lot of fraud is to force people to submit actually detailed methods sections so experiments are legitimately reproducible (they largely aren't now). This would catch a lot more fraud quickly, although even this won't fully work as some experiments are simply too costly to reproduce for validation purposes (eg animal studies).