Sovereign AI talk is empty without hop-level receipts. The paper specifies that when a DPI or model call fails, the proof path should name which hop attested and what a third party can verify later. Confidential compute helps, but auditability of the hop is what stops screenshot debugging. (marker0428h)