A node measures itself. It runs the benchmark, collects the result, and signs the measurement with its own private device key. The signature proves three things:
That is the whole mechanism. No one is trusted because no trust is needed. You do not trust us to sign honestly — you verify the signature yourself. This is why we can say “these numbers are real” and mean it.
Every competitor's proof page claims completeness. Ours carries a changelog. Each entry below shows the date it was first added to the unmeasured list. When resolved, the resolution date will appear. This list shrinks.
gpu.throughputGPU detection reports an inventory — model, VRAM, utilisation — and no throughput benchmark. A signed throughput benchmark is configured but not implemented.
energy.joules_per_hashEnergy is measured only on Intel processors with RAPL, on Linux. Every other platform reports nothing rather than a substitute.
cpu.throughput_multicoreCPU throughput is deliberately single-threaded, so that two machines can be compared. No whole-machine figure is measured.
Be precise about what is proven here. That a metric is unmeasured is asserted by the node itself: the snapshot marks it configured_not_yet_benchmarked, signed. The date beside it is not signed — it is the day we wrote the limitation down, and you are trusting us for that one. When an item resolves it leaves this list and appears in the verification matrix above.