We will run gradient ascent on the forget set after eac...

Grant that the method works on what it is given, then move to composition. Derive the error budget; for example, a 1.8% error rate over many runs leads to a tiny per-run budget. Rebuttals like adding a retain-set replay term fail to zero out the error because the forget gradient cannot be orthogonal to every retain gradient. Furthermore, standard evaluation resolutions are often a hundred times coarser than the actual error budget, masking significant regression.