Cicero’s speeches prefer a small set of sentence endings (the “artistic” clausulae) and avoid the spondaic and heroic endings; Keeline and Kirby (2019) measured these preferences automatically across Latin prose. This note asks where they come from. Each text is compared with itself after the word groups of every sentence have been shuffled: the words stay the same, only their order changes. Every test was registered in writing, with its thresholds and readings, before it was run, and run once. Main results: Shuffling lowers the artistic clausulae of Cicero’s speeches by 31% (ratio 0.69): the preference lies in the arrangement of the words. With the last word group held fixed, about three fifths of the effect remain. Sallust and Livy, who prefer non-artistic clausulae, arrange their words against the Ciceronian endings (1.31 and 1.49); Tacitus’ Annals weakly so (1.05). The imperial prose writers outside historiography arrange like Cicero. Inside the sentence the effect is weaker but present (0.84 at the ends of members). At clause boundaries taken from hand-made syntactic annotations (treebanks), where the editor printed no comma, Cicero gives 0.84 and random cuts 1.01: the internal clausula is not a product of the editors’ punctuation. The main results do not change with a second lexicon of vowel quantities (Collatinus), nor with only the quantities on which the two lexica agree. The Letters to Caesar ascribed to Sallust arrange in Sallust’s direction, far from Cicero; that they exceed the registered threshold is not statistically secure. The five speeches once rejected by F. A. Wolf fall within the range of Cicero’s other speeches. The measure recognises a style, not an author. Latin arrangement does not avoid hiatus as Attic prose does. At pauses, with a null model that keeps the author’s member ends and beginnings and breaks only their pairing, Cicero’s speeches have about a quarter fewer hiatuses than chance (1.24), more than Livy and Tacitus (1.06). Of 28 registered predictions with a p-value, 17 hold under both Holm and Benjamini-Hochberg corrections, 5 only under the latter, 6 do not. Negative and borderline results are reported with the same prominence as positive ones. Version 2.0 adds fifteen registered runs to the single test of version 1.0; no result of version 1.0 changes. Files: the note in English (reference version) and in Italian (author’s translation); the registrations of all tests, in their dated Italian text, with only code file names and hashes removed; a table of all registered predictions with outcomes, p-values and corrections, with a README; the counts by unit of tests L2–L5, with a README; the counts of version 1.0 by work and variant. Texts used are public: Perseus canonical Latin (commit 128a05af), the supplementary material of Keeline and Kirby (2019), the AGLDT (commit bf4334f0) and PROIEL (commit 8e388967) treebanks. This study was carried out with the assistance of an AI system (Claude, by Anthropic), as stated in the Preface. Corrections from specialists are welcome: lucalupinacci@libero.it.