Abstract
This disclosure describes how the custody of AI accelerators, the processors on which AI models are trained and run, is recorded independently and made the root of AI provenance. Each accelerator is registered by its attested hardware identity with an accredited custodian that is independent of the accelerator's maker, of the operator that runs it and of any government. At declared intervals, each accelerator proves to the custodian, by signed attestation and by measurements that cannot easily be falsified, such as signal delays to reference points, that it remains within the region its owner has declared and that applicable law permits. The proof is region-only: it shows that the accelerator is inside the declared region and held by the declared party, without revealing its exact site, its neighbours or anything about the work it performs. Each site that hosts accelerators declares its inventory, and the custodian reconciles the declared count with the attested count; a difference beyond a declared tolerance changes the site's live certification seal, and the reconciled figures are published at site level. The custody history of each accelerator, including every sale, lease, move, repair and destruction, follows it through every change of hands. An accelerator that fails to renew its proof is not disabled: there is no remote switch. Instead its custody status lapses, with declared consequences in the certification of the site and in the provenance of anything computed on it. Finally, when an AI model is trained, the custody status of the accelerators used is recorded in the model's birth certificate, so that every downstream provenance claim, about the model, the data it was trained on and the skills derived from it, rests on compute whose custody can be checked. Records are signed, and access to custody detail beyond the public level requires declared parties acting together under published rules.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Compute Custody™: Independent Custody Records for AI Accelerators as the Root of AI Provenance, in which an Accredited Custodian Independent of the Maker, the Operator and Any Government Holds Attested Records of Where Each Accelerator Is and Who Holds It, Region-Only Proofs Reveal Neither the Exact Site Nor the Workload, Declared and Attested Inventories Are Reconciled as a Condition of Site Certification, Custody History Follows Each Accelerator through Every Transfer, Lapsed Custody Has Consequences in Certification and Provenance Rather Than Remote Disabling, and the Custody Status of the Compute Used for Training Is Recorded in Each Model's Birth Certificate", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12906