Suggested respondent: a qualified ML researcher paired with owner counsel; named technical interviewee still to be selected. NIST distinguishes removal of training-data influence from ordinary file deletion. Definition.
Ask: “For this model and dataset, which removal claims are technically testable, and which cannot be promised?”
A strong answer should cover: source files, embeddings, caches, checkpoints and deployed versions; testing assumptions; residual risks; costs; and what third parties can verify. Reject a conflation of deleting a training file with proving that a model's behaviour no longer reflects it. Ask what the contract should say when verification remains incomplete.
Approved expert quote: [Not yet collected.]