What the meter counts, exactly.
The claim being measured: whether the operator's distinctive phrasing shows up in the machine's own output, unprompted. Lexical only. The meter sees phrasing, not judgment, and this page says so before any number does.
Markers: phrase-level n-grams derived mechanically from the operator's writing corpus under a pinned, hash-verified export. The machine does not choose what counts as the operator's; the derivation is re-runnable from the written method. Markers derive from the pre-treatment corpus only, on both sides of every comparison, so a later period can never pick its own detector.
The denominator, pinned: assistant tokens: tokens of machine output, summed across measured windows. Not words, not combined operator-plus-machine text. The claim is about machine output, so machine output is the denominator. The live figure on the So What page renders from the producer's feed; no figure is typed by hand anywhere on these pages, because a transcribed number survived three human reviews here and died at the first producer read.
Quotation is excluded: long verbatim spans are removed before counting, so an echo of a pasted instruction never scores as adoption.
Honest limits: the baseline and current readings come from separate pinned exports with a declared coverage gap between them, stamped on every receipt. Small raw counts are directional, not significant, and the page's fine print says which one you're looking at.