K-Veritas notarizes ML runs by checking the machine, not the number

A new class of audit is arriving for machine learning results, and it does not check the answer. It checks the machine.

K-Veritas, an open-source Go binary from 27-GROUP, wraps any training or evaluation command and produces a cryptographically signed PDF at the end. The receipt does not compare the reported number against a ground truth. Instead, the binary samples per-process telemetry at roughly 10 Hz while the run is in flight, then scores whether the hardware channels co-fluctuate like shadows of one process. A coherent run passes; a fabricated one leaves a signature the sampler can read.

The frame is a notarization, not a measurement. The tool certifies that the process you claim ran actually ran, on the hardware you claim it ran on, without rerunning anything. Reviewers stop trusting prose and start trusting a signed envelope that ties code, host, and time together. That is a different rung on the trust ladder than "trust the author" and a lighter one than "rerun the experiment."

The limits travel with the mechanism. 27-GROUP's reader is Linux-only for per-process attribution and cannot retroactively verify runs that predate it. The receipt confirms the run was real. Whether the science inside it is correct remains a separate question.