Regulations such as India's DPDP Act and the EU's GDPR grant a right to erasure. For AI systems this is hard: proving that specific data was removed, and that its removal did not disturb unrelated knowledge, is non-trivial. Building on the proposal in TR-06, we implement and measure selective unlearning on a content-addressed storage engine. Removing one contributor's data changed the storage root hash (cryptographic evidence of the change), removed 100% of that contributor's entries, and left a second contributor's data bit-for-bit intact.
Every stored entry has a content hash. The storage engine computes a Merkle-like root hash over all live entries. To unlearn a contributor, all entries tagged with that contributor's identifier are removed across every namespace (personal notes, chat history), and a new root hash is computed. The change in root hash is evidence that the stored state changed; comparing per-contributor entry counts before and after establishes that the removal was both complete (for the target) and selective (leaving others untouched).
We ingested eight entries for contributor Alice (five notes, three chat messages) and five entries for contributor Bob, then unlearned Alice.
| Measurement | Before | After |
|---|---|---|
| Alice's entries | 8 | 0 |
| Bob's entries | 5 | 5 |
| Root hash | be42e712... | 24638f6f... |
Alice's entries went to zero; Bob's remained at five, unchanged. The root hash changed from be42e712d88fb38c06e47e55d7eab3d7 to 24638f6f7a056c5f68e72d481f1b0693, providing tamper-evident proof that the stored state was altered. The combination — target fully removed, unrelated data preserved, root hash changed — is the empirical definition of selective, provable unlearning. The deletion proof (the post-deletion root hash) can be anchored to the blockchain so that the fact and time of erasure are themselves permanently auditable.
An honest scope note: this demonstrates provable unlearning of stored memory (the retrievable knowledge a model reasons over), not the removal of information already absorbed into a neural network's weights. For the retrieval-augmented architecture described in our other reports, where knowledge lives in external memory rather than the weights, memory-level unlearning is the operative mechanism. Weight-level unlearning for parametric models remains a harder, separate problem.
Selective, provable unlearning of memory is implemented and measured. The combination — target fully removed, unrelated data preserved, root hash changed — provides tamper-evident proof of erasure. Preliminary results, not peer reviewed.