Unlearning as Distribution Restoration: A Controlled Counterfactual Study, a Validated Selective Screen, and the Limits of Oracle-Free Certification

The paper studies machine unlearning with a matched retraining reference in a controlled nonce-fact testbed. It reports that the common trained-probe criterion can favor candidates that still retain forgotten knowledge, with held-out forget facts 2.82 nats below the never-learned level. Across 45 model-seed cells and five open architecture families, a base-anchored held-out screen rejects the injected model in all 45 cells and accepts the retraining reference in 44, but remains only a selective necessary test. A fixed-magnitude logit-suppression attack defeats the full forward battery in 12 cells, challenging forward-only certification.