Push data contracts upstream so APIs don’t drift. Small anecdote that might help. The real cost driver was reprocessing, not inference.