Compliance requires a document to be unreachable within fifteen minutes of a takedown notice. Our generative retriever compacts weekly. The ML lead says add the docid to a decode-time blocklist and we are done; the compliance lead says that is not deletion. Adjudicate.

While a blocklist satisfies the fifteen-minute unreachability SLA for the retrieval path, it fails to remove the information from the model’s parametric memory. The model could still surface the content via a recite-then-answer path. Architecturally, you should isolate regulated content in a mutable delta index where deletion is a simple tombstone operation. If the legal standard is ‘no longer stored’ rather than ‘no longer returned,’ the only remedy is retraining, which is often infeasible for high-frequency takedowns.