One model, three corpora — Greek inscriptions · Latin inscriptions · Greek papyri一个模型，三种语料 — 希腊铭文 · 拉丁铭文 · 希腊纸草
A single neural torso, trained jointly on all three corpora, shares one embedding
space — so retrieving "parallels" can cross the boundary between inscriptions and papyri,
Greek and Latin. Runs entirely in your browser, offline. Type a text, mark lost letters
with ?, choose the corpus, and it restores, dates, locates, and finds cross-corpus parallels.单一神经主干，在三种语料上联合训练，共享同一向量空间，故“平行”检索可跨越
铭文与纸草、希腊与拉丁之界。完全在浏览器内离线运行。输入文本，以 ? 标残缺字母，
择其语料，即可补字、定年、定位并检索跨语料平行。
Restoration补字
Each fill is evidence-typed: model-only = AI probability alone, confirm before use · parallel / cross-corpus-supported = a comparandum agrees · abstain = too uncertain to restore. how this is judged →每处补字均标注证据：仅模型＝仅凭 AI 概率，用前须确认 · 平行／跨语料佐证＝有可比文本印证 · 知止＝过于不确定。判定依据 →
When年代
Where地域
Cross-corpus parallels · most similar across all three corpora跨语料平行 · 三语料中最相似者
Honestly. A pilot joint model: a 3,308,986-parameter
character transformer (4 layers, 256-dim) with a corpus embedding, CPU-trained across the three
corpora — a 58,000-text cross-corpus index (Greek inscriptions from I.PHI +
InsAph/IRCyr, Latin from I.Sicily + IRT + EDH (CC BY-SA 4.0), Greek papyri from the DDbDP, via
EpiDoc→JSON). Measured letters-only on the model's own validation set:
restoration 59% top-1 / 81% top-3;
date within 50 yr 46%; region (of 225)
30%. The point isn't to beat DeepMind's specialists — it's a working pilot:
one small model reads across three corpora at once, with calibrated confidence. A rigorous leakage-free
re-eval (top-1 0.55, ECE 0.06) and the production architecture are on the validation page:
cross-language linking goes through the identity / concept layer, not one shared vector space.诚实说明。 试点联合模型：一个 3,308,986 参数的字符 Transformer
（4 层，256 维）附语料嵌入，跨三语料 CPU 训练；58,000 条跨语料索引（希腊铭文取 I.PHI,
拉丁取 I.Sicily，希腊纸草取 DDbDP，经 EpiDoc→JSON）。于模型自有验证集测得（仅字母）：补字首选 59%/
前三 81%；定年 50 年内 46%；地域 (共 225)
30%。要旨非胜过 DeepMind 专模，而为一可用试点：一个小模型可同时读跨三语料，且置信已校准。严格无泄漏复评（首选 0.55, ECE 0.06）与生产架构见验证页:
跨语言之链经身份／概念层；不取单一共享向量空间。