You could run those experiments by constructing cases, or re-constructing known cases.
Even for people who are wrongfully convicted, we can't always say the jury made the wrong call. In many of those cases, the conviction is overturned due to new evidence being discovered, or because the police or prosecution wrongly hid exculpatory evidence from the jury, or because the defendant's lawyer was grossly incompetent – in those cases, it is entirely possible the jury made the right call given the evidence actually presented to them, even if it turned out ultimately to be the wrong one.
Yes, in the general case you can't know that. That's why you need to construct artificial cases where you do know.
Very similar to getting labelled training data for your machine learning.
I wonder whether you could get the benefits of weighing the evidence collectively but reduce the influence that 1 or 2 dominant jurors might have.
If 2 or 3 independent groups reached the same conclusion, does that increase the chances of them being correct?