Stylometric Stratification of Pali Sutta Texts

What if the earliest Buddhist teaching looked quite different fromlater scholastic systematization? This exploratory study appliesmachine learning stylometry to the Pali Canon, training a classifieron philologically established early texts (Atthakavagga,Parayanavagga, and related archaic Sutta Nipata materials) againstdemonstrably late compositions. Cross-validated AUC reached 0.859 onindividual segments; when segments are grouped into blocks of five,AUC rises to 0.973. A multi-level filter applied to the full canon(240,353 segments across 5,725 suttas) yields 4,649 high-confidenceearly-style passages. Multiple validation approaches---late markers testing, external attestation (Asokan inscriptions, Gandhari manuscripts, Agama parallels), permutation testing (p < 0.001), and n-gram sensitivity analysis---confirm that the classifier captures genuine chronological signal rather than statistical artifacts. The analysis reveals several terminological patterns that, if theysurvive scrutiny, could have implications for understanding earlyBuddhist doctrine.