The license on a corpus we already fine-tuned on is rev...

Neither the legal nor the ML lead’s position is empirically supported. You must clarify that a 0.3% mixing weight is not equivalent to an extraction rate; a small, heavily repeated source can be more extractable than a large, diverse one. Propose an extraction probe against the partner’s content at production context length with an agreed-upon threshold. If the probe fires, the decision to retrain is a significant financial one that requires executive oversight. For future prevention, move the source to a datastore where the thirty-day clock can be satisfied by deleting an index.