Abstract
This disclosure describes an offline archive tier for governed data custody, in which data deposits, their versions and the data behind AI models are kept on removable media, such as magnetic tape or optical discs, that are physically disconnected from any network when not in use. Every cartridge carries a signed manifest that binds each segment on it to the persistent data identifiers of the deposits and versions it holds, to their terms and permitted regions, and to their content hashes, and the registry records which cartridge holds what, where it is and who handled it. Data on the media remains encrypted under per-deposit keys held by depositors, so that a deposit can be deleted from every archived copy by destroying its keys, without recalling or rewriting the media. For each certified AI model version, the exact data it was trained or tuned on is sealed into a lineage snapshot in the archive, so that the model's training can be audited later, compromised or poisoned data can be traced to every model that used it, and a model can be rebuilt from its last known good data. Physical media move only under a chain of custody: sealed containers, two-person handling, location records and tamper-evident seals, each event signed and recorded. Nothing read back from the archive reconnects directly: it is restored onto attested clean hardware, checked against the hashes in the manifest and the registry, scanned against current advisories, including notices of poisoned data, and only then admitted to live custody. AI systems never have direct access to the archive; retrieval is performed by certified custodians, with the depositor's approval where terms require it, and every retrieval is signed and recorded. Sampled restore drills and periodic integrity reads prove that archived data remains readable, and media are migrated to new generations with every hash re-verified. Archives are replicated across sites in different locations, such as parks of an infrastructure grid, as strategic reserves from which a region's data and models can be rebuilt after a catastrophic loss.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI DataBank™: A Governed Offline Archive Tier for AI Data on Removable Media, with Every Cartridge Bound to Data Identifiers, Sealed Lineage Snapshots of the Data behind Each Certified Model, Chain of Custody for Physical Media, Verified Re-Entry through Quarantine, Periodic Proof of Readability, and Archives Spread across Sites as Strategic Reserves", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12646