The unread oracle bone characters seeded: corpus, readings and open data

Read the Linear A reference records to match the house style, then researched the oracle-bone corpus: about 150,000 inscribed pieces from Yinxu, the Jiaguwen heji edition of 41,956 rubbings, the holders (National Library of China, Academia Sinica, British Library) and the radiocarbon dating of 2021 [@guo1978heji] [@bl2015oracle] [@liu2021radiocarbon].

Traced the history of decipherment from Wang Yirong's recognition in 1899 through Luo Zhenyu's 1914 readings, Wang Guowei's 1917 confirmation of the Shiji king list, Dong Zuobin's 1933 periodization and its diviner-group revision, to Qiu Xigui's account of the form-chain method [@luo1914kaoshi] [@wang1917xianwang] [@dong1933duandai] [@shaughnessy1982periodization] [@qiu2000chinese].

Collected the counts. Published tallies of distinct graphs run from 3,763 to about 4,500 with between 1,160 and 2,200 read, while HUST-OBC lists 9,411 unread image classes against 1,588 read; the gap is the variant-merging problem [@ma2026computational] [@wang2024fudan] [@li2024survey] [@wang2024hustobc].

Documented the Anyang bounty: 100,000 yuan per accepted reading since 2016, Jiang Yubin's 2018 award for chun, the 2023 batch including Zhou Zhongbing's tai, and three awards by 2025 [@chinadaily2017reward] [@bjreview2019vocation] [@henan2024deciphering] [@fang2025race].

Reviewed the computational work: OBC306, HUST-OBC and EVOBC datasets, the OBSD diffusion model (ACL 2024 best paper), OBI-Bench's finding that multimodal models decipher at an untrained-human level, fragment rejoining from Zhang and colleagues (2020) and the Diviner system, and the Yinqi Wenyuan and Tencent platforms [@huang2019obc306] [@guan2024evobc] [@guan2024deciphering] [@chen2025obibench] [@zhang2020rejoining] [@digitalorientalist2024ai] [@tencent2024corpus].

Found CC0 images on Wikimedia Commons: Heji 1 in the National Museum of China (lead image) and the British Library's Or. 7694/1595 lunar-eclipse scapula; uploaded both and created the artifact and copy records. The Commons API was rate-limited through the shared proxy, so file pages were read from their raw wikitext.

Wrote the mystery with five sections and a 15-row literature table, five key questions, six answers to the central question (names, lost words, variants, descendants, model-assisted, none of these) and one on rejoining, with evidence tiered conservatively.

Not uploaded: HUST-OBC (CC BY-NC, several GB) and EVOBC (no licence stated); the datasets are recorded with their source URLs instead. The Yinqi Wenyuan site blocks automated fetching, so its description rests on the 2025 China Daily report.