Legal needs a licensed corpus out of the product in 30...

The dispute should be reframed from whether to keep the data to where the data lives. Retraining the model on low-risk data while serving the encumbered corpus from a retrieval index allows for compliance with the 30-day deletion deadline, as index deletion is instantaneous. Furthermore, the four-point benchmark drop should be decomposed into language competence (which is lost) versus retrievable knowledge (which can be recovered via RAG). Unlearning is rejected as a strategy because it is technically infeasible to enumerate every sequence to be removed within the required timeframe.