Exact machine un-learning: remove any customer from your models. Prove it.

B2B AI vendors train shared models on many customers’ data, and several say in writing that a departed customer’s data cannot be removed (HubSpot: “It is not possible to delete previously used data from trained models.”). Lethe is training infrastructure that makes removal exact and provable:

1. Every training record is tagged with the customer it came from.
2. Customers are split into isolated shards (√N), each fine-tuned on its own customers (SFT on tickets, then DPO on accept/reject feedback). The shards are merged into one serving model with a binary merge tree.
3. Removing a customer retrains only its shard and re-merges the path to the root. Removals can be queued and batched, so shared merges and the root are redone once.
4. Every training run leaves a signed receipt. Each removal issues a deletion certificate with a non-membership proof for every one of the customer’s records, which the customer can check with their own data.

Hackathon project. All data is synthetic, all customer names are fictional, and TEE attestation is
simulated.