It’s not a question of “good” data. Slice and dice perfectly random data and sometimes you get spurious correlations. The only way to separate them from real results is to have completely new data.

It’s not even a question of p hacking or bad design. Preform enough experiments and you always get false positives.