The load test.

The accusation: the instrument just measures CPU load. If that is true, stressing every core must change the bit statistics: the lean, the variance ratio, or the serial correlation.

The defense: load changes the spin count’s magnitude, but the instrument keeps only its parity, which rides the phase noise between independent clocks. Under full load the spin count must collapse (the manipulation check), while the parity statistics stay flat.

The protocol, fixed before the data arrive: alternating blocks of idle and full-core load, each 8-bit segment of the stream tagged by condition, the first two after each switch discarded. Leakage is declared if any of the three statistics differ at more than 3 standard errors. The manipulation check must pass for the run to count. Everything is computed on this device, and the raw trials can be downloaded.