An interrupted upload can leave a saved progress offset ahead of missing data or behind bytes that actually arrived. This report examines an implemented model-blob uploader that compares payload chunks before resuming, persists only contiguous completed ranges and conditionally reuses a prior full-payload audit after checking current sealed-account metadata. Ten unchanged existing tests pass against local buffers, controlled asynchronous callbacks and synthetic program-account records. They cover the first differing chunk, legitimate zero data, partial final chunks, out-of-order completion, failure draining, audit identity checks and a byte-comparison helper for a genesis-preloaded program. The surrounding deployment script and on-chain upload handler are inspected rather than executed. The evidence supports bounded recovery and admission predicates, not a completed model deployment or verified inference. Audit reuse trusts a supplied local audit record and RPC observations; its hash fields and boolean flags are not authenticated evidence by themselves. A program seal prevents later writes under the inspected implementation but does not verify the payload hash or model correctness. The study separates these assumptions from fresh tests and treats user control, infrastructure affordability and accountable evidence as research questions rather than demonstrated social outcomes. Jarett Dunn. Technical report, version 1.0. Preprint, not peer reviewed. Codex assisted source recovery, inspection, local test execution and writing. New report prose is CC BY 4.0; upstream software retains its licenses and attribution.
No takes yet. Share an insight, caveat, or question.
Jarett Dunn (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: