Forensic Stylometry of Generated Literary Prose: A Distance-Based Framework for Detecting Register Drift in Commercial Language Models
Madison Charlton · Zenodo (CERN European Organization for Nuclear Research) · 2026
We propose a forensic framework for evaluating whether generated literary prose is consistent with a target register, framed as a multivariate two-sample testing problem rather than a generation-quality problem. The framework treats canonical authors, hand-authored generative grammars, fine-tuned small language models, and prompted commercial language models as four classes of samplers drawing from candidate register distributions, and uses the energy-distance permutation test of Szekely and Rizzo over a 14-dimensional surface stylometric feature panel to compare them. We validate the framework on a corpus of seven texts: four exemplars of American dark realist prose (Bierce, Tales of Soldiers and Civilians, 1891; Crane, The Red Badge of Courage, 1895; Crane, Maggie, 1893; and Charlton, Dendrome, 2023, included with the author's written permission); one historical-reference exemplar of related but distinct register (Hawthorne, Twice-Told Tales, 1837); and two out-of-register controls (Mill, On Liberty; Darwin, Origin of Species). We compare against Burrows's Delta as a baseline and find substantive disagreement on the within-canon rank order, which we interpret as the methodological contribution: function-word stylometry and surface structural stylometry answer different questions about register membership.