Abstract
This disclosure describes a method of preventing raw data from leaving a custody system for data and AI models, even if the job running against the data, the operating system or the operator's software is compromised. Every kind of output that may leave custody, such as a model weight update, an aggregate statistic, an evaluation score or a trained model, is described by a signed output contract that states its exact format, its maximum size, its maximum rate, the privacy budget it consumes from each deposit used, and its permitted recipients. Deposit terms name which contracts they allow. All outputs leave through a one-way hardware egress gateway on an independent safety plane. The gateway's own firmware parses each output strictly against its contract, re-encodes it in a canonical form to remove hidden content, checks its size and rate, checks and debits the remaining privacy budget of every deposit involved, confirms that a leakage test has passed where the contract requires one, and passes only outputs that meet every condition. Anything else is blocked and quarantined, the governance level of the job is raised, and the event is recorded. Because the contract limits form, size and rate, it also limits how much information any covert channel could carry. Every transfer is logged with its hash in a tamper-evident ledger.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Data Bank™: Output Contracts Enforced at a Hardware Egress Gateway, so that Raw Data Cannot Leave Data Custody even through a Compromised Job", Technical Disclosure Commons, (October 13, 2026)
https://www.tdcommons.org/dpubs_series/12168