Read a page of unedited machine prose aloud and something odd happens to your breathing. It evens out. Every sentence takes roughly the same lungful: fifteen to twenty-five words per breath, clause then comma then clause then full stop. Again. The effect is hard to name on first encounter but unmistakable once named: the rhythm is flat.
Human writing does not behave this way. It lurches. A writer builds a long, qualified, carefully hedged sentence across three lines because the material demands every qualification. Then stops. Two words. The variation is not decoration; it is how emphasis works in prose. Short sentences land because long ones set them up.
Researchers call the property burstiness, and it has a convenient mathematical handle: the coefficient of variation of sentence lengths. Take every sentence in a passage, count its words, and divide the standard deviation by the mean. Lively human prose typically produces a coefficient well above 0.5. Unedited model output regresses toward the average sentence and drifts far lower.
Tellmark measures this under entry S-03, and it earns a special status in the catalogue: the most robust structural signal we track. Lexical tells can be scrubbed with a synonym pass. Rhythm survives light editing, because fixing it requires restructuring, and restructuring is the one thing a hurried writer skips.
The measurement has honest limits, and the catalogue states them. Eight sentences minimum, or the statistic is noise and the tool refuses to compute it. Legal drafting and some academic registers run naturally uniform. A low score is a prompt to read aloud, not a verdict.
The remedy
The remedy is mechanical enough to do on a deadline. Find your longest paragraph. Cut one sentence to under six words. Merge two mid-length neighbours into one long construction with a proper subordinate clause. Read it aloud once more. Where your breath varies, the prose has a pulse.