Relational graphic structure supports diachronic retrieval of wholly unseen ancient Chinese character identities

Ancient Chinese character identities can persist while graphic forms change across historical stages. We tested whether models can retrieve later forms of identities wholly absent from learning. Eighty metadata-derived candidate relations were independently reviewed by five specialists; 64 identities comprising 9540 quality-controlled images across oracle-bone, bronze, Spring–Autumn and Warring States scripts were retained. Five-fold whole-identity evaluation excluded each test identity from training, validation, descriptor standardisation and model selection. Frozen DINOv2 achieved identity-macro MRR of 0.469, supervised contrastive learning reached 0.554, and adding 24 image-derived structural descriptors increased MRR to 0.591 (difference 0.0369, 95% CI 0.0199–0.0540). Topology and skeleton organisation carried most transferable structural signal. These findings indicate that explicit relational graphic structure can improve controlled whole-identity retrieval of unseen ancient Chinese character identities. They do not constitute automated decipherment and should be interpreted as evidence-surfacing support for expert palaeographic judgement.