Unorganized Machines is a technology company developing verifiable storage infrastructure. Its approach stores a file’s reconstruction recipe rather than relying only on keeping a full copy: compatible machines can use the recipe to recover the original bytes, which are then checked against a recorded fingerprint. The company’s working prototype is being tested, and it is seeking design partners to evaluate real-world workloads, recovery performance, costs, and hardware requirements.

The initial focus is on AI laboratories that produce large volumes of model checkpoints, compute and storage providers assessing whether recomputation can replace storage, and auditors or regulators who need to verify recovered data. In one prototype test, a 164,538-byte recipe reconstructed a 1,048,576-byte original. The company’s current work builds on earlier research into arithmetic and checkpoint recovery; its prototype paper is in preparation. Longer term, Unorganized Machines aims to develop receipts that help others check how a model was trained, using exact, repeatable computation.