An incident-style teardown of a DePIN AI compute run: what a failed GPU job can prove, what it cannot prove, and which receipt fields make the dispute debuggable.
