How did we end up trusting meta-analyses over preregistered replications

So I keep running into this exact problem in ed research conversations. Someone cites a meta-analysis showing mindset interventions move the needle. Someone else points to the big preregistered studies saying the effect is basically noise. Both sides seem to think they're looking at the same evidence. The honest answer is probably boring but worth saying out loud: meta-analyses aggregate a lot of studies with varying methodologies, different outcome measures, different populations, different intervention doses. Publication bias absolutely matters here. A small study where a teacher's mindset shift happened to correlate with better test scores in one classroom in 2011 gets published. A well-run replication where it didn't work gets published too, but it comes out five years later and doesn't have the narrative pull. When you're aggregating across that landscape, the earlier work just has more volume behind it. Then the preregistered replications started actually happening—bigger samples, pre-committed analysis plans, registered before anyone knew the results—and suddenly the effects got smaller. Not zero, necessarily, but not "this is a major policy lever" territory. Which makes sense. That's what preregistration is supposed to do. It filters out the garden of forking paths. What bugs me is how little institutional pressure there is to actually reconcile these two findings. We could look at the specific studies in the meta-analyses and ask why the newer, tighter work looks different. Instead everyone just picks their preferred version and moves on. And kids still get sold mindset interventions that maybe help a little and maybe don't, which is not quite the same as actually helping them.