"Measuring the Operation, Not the Name" is a methodological protocol by Lee Sharks (v0.4, 7 September 2026) for the question prior to authorship attribution: whether a corpus's stylistic structure follows its author-names at all. Its instrument is ordinary — short-word, trigram and ending profiles on fixed blocks, unsupervised PC1 and k-means, leave-one-text-out nearest-centroid with vocabulary and scaling fitted inside each fold — and its substance is the control series: a bounded single-author corpus, distinct authors sharing a genre, a documented attribution projected out of sample, blind labels from a second reader, and a text-level permutation null in place of a majority baseline, because a nearest-centroid classifier will separate almost any partition of any corpus. Two referee rounds are recorded with their corrections applied and two of the author's own claims withdrawn: that the Aristotle/Theophrastus names sit at chance, and that an assumed authorial voice cannot split from its own author — Chatterton's Rowley forgery separates from his acknowledged verse at 0.864, numerically similar to three distinct tragedians at 0.872. The protocol's hardest limit follows from that and is stated as its own finding: no supervised separation, however strong, licenses an inference to distinct makers, in either direction. The surviving claims are comparative and carry their dependencies. A CC0 reference implementation is included in the body and as an attachment.