Plumb re-runs a paper's own code on your compute, re-derives each claim from the actual run, and returns a per-claim verdict: REPRODUCED, DIVERGED, or UNVERIFIED. Execution decides, never an LLM. Open source; the honest number is published with it.