Answers are grounded in this paper and cite their sources.
No accepted work is lost, and no completed result is delivered twice, across a durable broker and GPU runtimes.. Methods and Systems for No-Loss Work Reconciliation Between a Durable-Log Broker and Accelerator Inference Runtimes Using Consumer-Reported Hardware Residency. U.S. Provisional Patent Application No. 64/147,527 (patent pending). Embodied in the broker of the Inference Fabric. Every piece of work is in exactly one of five states at any time: durable storage, consumer memory, queued for GPU, running on GPU, or durably complete. At any instant all accepted work is the disjoint union of those five sets, so nothing is ever unaccounted for. Completion is marked by a durable tombstone, and the durability order is fixed: the output is persisted, then the tombstone, and only then is the input retired. Recovery is a classification. Canonical page: https://coreyjames.ai/research/family-two-durable-work-and-lossless-recovery/. Repository document: https://github.com/coreyjames-ai/portfolio/tree/main/research/patent-families/family-two-durable-work-and-lossless-recovery, mirrored at https://gitlab.com/coreyjames-ai/portfolio/-/tree/main/research/patent-families/family-two-durable-work-and-lossless-recovery. Generated from the coreyjames.ai inventory on 2026-10-28; every figure and evidence line in this report is one on the canonical page it cites.
No takes yet. Share an insight, caveat, or question.
Corey James (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: