When the Commentary Chose a Word: A Masked-Probing Study of the Tuan's Yuan/Da Distinction across Four Classical-Chinese Models

First public release accompanying the working paper: When the Commentary Chose a Word: A Masked-Probing Study of the Tuan's Yuan/Da Distinction across Four Classical-Chinese Models (Raymond Lao, 2026). This release archives the paper together with the complete code and data used to produce its results, so that the experiment can be reproduced exactly. Contents Lao_2026_masked_probing_yuan.pdf — the paper (preprint). yuan_masked_probing_AB.py — the end-to-end experiment: data, marker-masking with self-checks, embeddings, cosine similarity, Mann–Whitney U with rank-biserial r, and the scatter figure. outputs/ — the exact results reported in the paper (per-hexagram cosines, group means and direction, statistical tests, token diagnostics, and the scatter figure). requirements.txt — pinned dependencies. Method in brief Nine hexagram Tuan passages are grouped along the yuan axis — Group A (yuan-retained) and Group B (da-substituted) — and compared, by the cosine similarity of masked-language-model representations, with an anchor built from the Qian and Kun glosses of yuan. The verdict markers are then surgically masked and the comparison repeated, separating a genuine semantic relation from mere shared word-form. Four classical-Chinese encoders are tested (GuwenBERT, SikuBERT, GujiBERT, BERT-Ancient-Chinese). Reproducibility The pipeline is deterministic. See README.md for setup and run instructions. The classical text is in the public domain (verified against the Chinese Text Project); the code is released under the MIT License.