Choice or arrangement? A word-group shuffling test for ancient prose rhythm: a scale of null orders, known-answer controls in Greek and Latin, and the code to reproduce it

A rhythm that is rare or frequent in a prose author can come from the words the author chose or from the order in which they are arranged. This preprint describes a test that separates the two. Each text is compared with itself after its word groups (a content word with the small words that lean on it) have been shuffled at random within the sentence, twenty times: the words stay the same, only their order changes. The ratio between a measure after shuffling and the same measure in the original order shows whether the arrangement works for a rhythm (below 1), against it (above 1), or neither. The principle is K. J. Dover’s (1995; 1997), who compared Greek passages by hand with a single transposition of their words. A scale of null orders. Because a random order can be far from anything an author could have written, the same effect is measured with alternative orders ever closer to the original: free within the sentence, within the colon, among neighbouring groups only, within the clause (from hand-annotated dependency treebanks), in projective orders built from the syntactic tree, and with the last word fixed. For hiatus at pauses, an exact pairing null keeps the words that close and open each member. Known-answer controls in Greek and Latin, with opposite signs. The test recovers: Blass’s law in Demosthenes: shuffling raises runs of three or more short syllables by 61%, against −11% to +2% in eight other Attic orators. The effect survives every null and is halved only by the most conservative one. Late Plato’s avoidance of hiatus (1.39) and preference for the paeonic clausula (0.59). Cicero’s artistic clausulae (0.69) and the opposite sign in Sallust and Livy (1.31, 1.49). A stratum of Against Timocrates (111–150) that lacks the habit (1.02 against 1.71 in the rest of the speech). A prediction registered before any measure on six Greek prose authors succeeded for Polybius, Plutarch, Herodotus and Thucydides, and failed for Lysias. The test does not identify authors: it says whether a rhythmic preference lies in the order of the words. Every test was registered and dated before it was run. All numbers come from two earlier records: on the Demosthenic corpus (doi:10.5281/zenodo.23661370) and on Cicero’s clausulae (doi:10.5281/zenodo.23517823). The registrations and the tables of all registered tests, with Holm and Benjamini-Hochberg corrections, are deposited there. Files Prose_rhythm_shuffle_method_preprint_EN.pdf: the preprint (English). Prose_rhythm_shuffle_method_preprint_IT.pdf: the same in Italian. prose-rhythm-shuffle-1.0.0.zip: the code package (Python, MIT licence). It contains no texts: it downloads them from Perseus, V. Gorman’s treebanks and J. Winge’s Latin Macronizer at fixed commits, checks their SHA-256 hashes and recomputes the main tables, comparing every number with its published value. In the reference run it reproduced all 235 values checked; the report of that run is included. The work was carried out with the assistance of an AI system (Claude, by Anthropic); the preface of the preprint describes the division of work.