Graphic Compensation in Ancient Greek Documentary Hands: A Computational Paleographic Analysis from Handwritten Character Recognition

Historical scripts are structured visual systems whose internal organization has traditionally been explored through qualitative paleographic analysis. However, many of the principles shaping their evolution remain difficult to test empirically across corpora. One such principle, formulated by Jean Irigoin, is the hypothesis of graphic compensation, according to which graphic systems tend to maximize visual distinctiveness between letters that share similar phonetic functions while allowing greater visual proximity in other cases. In this work, we develop a computational approach to explore the visual structure of historical scripts and apply it to the cursive Greek writing of Hellenistic papyri. Building on a previously introduced representation-learning framework for ancient Greek letterforms, we analyze the behavior of a ConvNeXt-V2 character recognizer trained with lacuna-based fragmentation and dynamically learned supervised contrastive loss in order to extract the similarity relations that structure the script. Rather than treating classification errors as poor technical results, we interpret them as paleographically meaningful patterns that reveal the internal organization of the graphic system. The resulting analysis shows that several of the strongest confusions occur between visually similar but phonetically distinct letters, including Alpha–Lambda, Omicron–Theta, Iota–Rho, Epsilon–Sigma, Tau–Upsilon, and Eta–Pi, providing quantitative support for Irigoin's hypothesis and suggesting that the principle of graphic compensation continues to operate even in the less controlled domain of cursive handwriting. This tendency is supported by a global pairwise comparison of same-category and vowel–consonant confusions, by stability checks across alternative training variants, by evaluation on damaged characters, and by a papyrus-level split designed to reduce leakage between train and test material. More broadly, this study proposes a methodological framework for computational paleography, in which model behavior can be used to extract interpretable descriptions of script structure from large collections of cultural heritage material. Finally, we suggest that the learned relationship between phonetic difference and visual confusion may itself serve as paleographic background knowledge and be incorporated into future HTR pipelines, helping recognition systems resolve visual ambiguity when visual evidence and linguistic expectations compete during decoding.