One inscription, five databases.

一块铭文，五个数据库。

A small marble plaque, 15.5 × 24.5 × 3 cm, was first recorded around 1730 in the Museo Salnitriano in Palermo and now sits in the Museo Archeologico Regionale Antonino Salinas. It carries seven lines of Greek on the left and seven lines of archaic Latin on the right, separated by a deeply incised vertical rule. The text is — fittingly — an advertisement for a stonecutter's shop, in two languages, telling passers-by that "inscriptions are designed and carved here for sacred temples in connection with public works." The inscription literally advertises the production of inscriptions.

一块小型大理石板，尺寸 15.5 × 24.5 × 3 厘米，约 1730 年首次记录于巴勒莫的 Museo Salnitriano，现藏于 Museo Archeologico Regionale Antonino Salinas（萨利纳斯考古博物馆）。板面被一道深刻的竖线分成两栏：左栏七行希腊文，右栏七行带古拼写的拉丁文。文本本身是一则石匠铺的广告，用两种语言告知路人："此处为神圣庙宇及公共工程刻制铭文。"换言之，这块铭文广告的内容，正是"刻制铭文"。

It is the perfect case study for what Heřmánková, Kaše & Sobotková (JDH 2021) are up against. This single object appears, with subtly different metadata, in every major Latin and Greek epigraphic database — except one. Walking through how each database represents it shows in concrete detail why building a clean, comparable, cross-database dataset is so much harder than the abstract ETL-pipeline diagrams suggest.

对 Heřmánková, Kaše & Sobotková（JDH 2021）所讨论的问题而言，这是一个绝佳案例。这一件文物出现在几乎所有主要的拉丁文与希腊文铭文数据库里，只有一个例外，而每个库给它的元数据都略有不同。逐一查看各库的呈现方式，可以具体看到："建立一份跨库可比的干净数据集"为什么远比抽象的 ETL 流水线图所暗示的更难。

§ 1The text铭文文本

The inscription as carved on the stone. Hover over a Greek line to highlight the Latin equivalent — they are not parallel translations but loose biversion, with awkward word-for-word renderings on both sides.

石面上的文字。把鼠标移到希腊文某一行，对应的拉丁文会高亮，两栏并非整齐对译，而是一种"双版本"翻译，两边都带着字字直译的笨拙。

§ 1.5From paper to digital — 140 years of editions从纸到数字，一百四十年的版本史

Before any of the modern databases existed, this inscription had already been published twice in the great 19th-century print corpora. Each editor made different choices. Each digitization that followed inherited some of those choices and silently dropped others.

在任何现代数据库出现之前，这块铭文早已被两次收入 19 世纪的大型印本丛刊，每位编者做出不同的取舍，而其后的每一次数字化又继承一些、悄悄丢掉另一些。

The chain stretches at least eight steps long: physical stone (1st c. CE) → 18th-century manuscript transcription (Torremuzza, Ignarra) → CIL X 7296 (Mommsen, 1883) → IG XIV 297 (Kaibel, 1890) → 20th-century revisions (Wilson, Manni Piraino, Bivona) → five modern digital databases → SDAM's cleaning pipelines → JDH 2021 paper. Look at the two print editions side by side:

这条链条至少有八环：实物石头（公元 1 世纪）→ 18 世纪手抄记录（Torremuzza、Ignarra）→ CIL X 7296（Mommsen，1883）→ IG XIV 297（Kaibel，1890）→ 20 世纪诸修订（Wilson、Manni Piraino、Bivona）→ 五个现代数字数据库→ SDAM 的清洗流水线 → JDH 2021 论文。把两个 19 世纪印本并置：

Original print pages (scans)印本原页（扫描）

These are the actual pages — what every database in this case study ultimately quotes. Compare them with the cleaned facsimiles below: where the editorial discussion above the inscription gets dropped in modern databases, this is the prose lost.

这是真实的页面，本案例所有数据库最终都引自此。把它们与下方的清晰摹本对照阅读：现代数据库丢失了铭文上方的编辑论述，此处即所失之文。

Three things the print editions kept

印本保留的三件事

• Argumentation. Both editors record their reasoning: Mommsen argues "Siculam originem prodit quod bilinguis est, cum litterae sint optimae aetatis" — the Sicilian origin is shown by the bilingualism, the dating by the high quality of the letterforms. None of the modern databases preserves this reasoning chain. They give you a place and a date; the print editions give you why.
• 论证。两位编者都记录了他们的推理：Mommsen 写道 "Siculam originem prodit quod bilinguis est, cum litterae sint optimae aetatis"（双语性显示其西西里出身，字体之精美显示其年代）。任何现代数据库都没有保留这条推理链。它们告诉你"地点"和"日期"，印本告诉你为什么。
• Provenance history. Mommsen writes "Panormi olim apud Iesuitas, nunc in museo publico" — formerly with the Jesuits in Palermo, now in the public museum. EDR has only "Palermo?" with a question mark. The 1762 → 1883 → modern chain is in the print but lost from the digital.
• 出土流转史。Mommsen 写"Panormi olim apud Iesuitas, nunc in museo publico"（曾在巴勒莫耶稣会，现藏公共博物馆）。EDR 仅记"Palermo?"。1762 → 1883 → 现今的流转链在印本里清晰，在数字层却已丢失。
• Editorial judgment. Mommsen calls the stonecutter's bilingualism infantia ("infancy / incompetence"). Kaibel demurs: "nec Graecus opinor nec Romanus homo" ("a man neither Greek nor Roman, in my opinion"). Two scholarly judgments coexist on the same artifact. The modern databases collapse both into a single normalized "type: epitaph" or "type: advertisement."
• 编者判断。Mommsen 把石匠的双语水平称作 infantia（"幼稚/不通"）。Kaibel 不同意，他说 "nec Graecus opinor nec Romanus homo"（"我以为他既非希腊人也非罗马人"）。同一件文物上有两种学术判断并存。而现代数据库把两者都压缩为统一的 "type: epitaph" 或 "type: advertisement"。

Three things the print editions changed (and the digital inherited)

印本改动了的三件事（数字层一并继承）

• The line "QVM OPERVM" became "CVM OPERVM." Both Mommsen (1883) and Kaibel (1890) silently regularize the archaic QVM on the stone to CVM. Look at line 6 of both columns above — the orange marker on the CIL flags it. The actual stone says QVM (visible in the photograph in the previous section). EDCS, EDR, and PHI 140601 all inherit "CVM" from the print editions. Only PHI 175744 and I.Sicily preserve "QVM" — and only I.Sicily encodes both forms with EpiDoc <choice> markup. The textual normalization that begins on Mommsen's printing block in 1883 is still propagating in 2024.
• "QVM OPERVM" 被改作 "CVM OPERVM"。Mommsen（1883）与 Kaibel（1890）都默默把石上古拼写 QVM 规范化为 CVM。请看上面两栏第 6 行，CIL 上的橙色框出了这一点。原石上写的是 QVM（前一节的照片可见）。EDCS、EDR、PHI 140601 都从印本继承了 "CVM"。唯有 PHI 175744 与 I.Sicily 保留 "QVM"，而只有 I.Sicily 用 EpiDoc <choice> 同时记录两种形式。1883 年 Mommsen 印刷版的文字规范化，一路延续传播到 2024 年。
• Word-spacing got invented. The actual stone has no word dividers. Mommsen (CIL) introduces them in some lines (e.g. NAOIϹ ΙΕΡΟΙϹ). Kaibel (IG) runs everything together (ΝΑΟΙϹΙΕΡΟΙϹ). The modern databases mostly follow Mommsen, but the 19th-century split is still visible in PHI 175744 vs PHI 140601's transcription differences.
• 词间空格是被"发明出来的"。原石上没有词间分隔符。Mommsen（CIL）在某些行加入空格（如 NAOIϹ ΙΕΡΟΙϹ），Kaibel（IG）则全部连写（ΝΑΟΙϹΙΕΡΟΙϹ）。现代数据库大多沿用 Mommsen 的版本，但 19 世纪的这个分歧依旧可在 PHI 175744 与 PHI 140601 的转写差异中看到。
• The Greek got a "diplomatic" version. Kaibel below the columns supplies a single normalized Greek sentence with proper accents, breathings, and spacing: Στῆλαι ἐνθάδε τυποῦνται καὶ χαράσσονται ναοῖς ἱεροῖς σὺν ἐνεργείαις δημοσίαις. Mommsen does not. This is the line the modern databases (PHI especially) descend from for their text. The "canonical" Greek text in PHI today is Kaibel's 1890 reading, four iterations away from the stone.
• 希腊文得到一个"规范"版。Kaibel 在两栏下方补上一行带重音、气音、词间空格的规范希腊文：Στῆλαι ἐνθάδε τυποῦνται καὶ χαράσσονται ναοῖς ἱεροῖς σὺν ἐνεργείαις δημοσίαις. Mommsen 没有。这条规范行就是现代数据库（尤其是 PHI）的希腊文文本来源。今天 PHI 中那条"标准"希腊文，其实就是 Kaibel 1890 年的读法，距离原石已转录四次。

What the digital era discarded entirely

数字时代彻底丢掉的东西

The Latin commentary paragraphs you can see in both print snapshots — Mommsen's argument about marble quality, Kaibel's quotation of Mommsen's verdict, both editors' interpretations of the bilingual stonecutter — are gone. None of the five modern databases has a structured field for "editor's interpretive commentary." EDR records "Textus secundum (6)" — meaning "text follows reference 6 [Manni Piraino 1973]." That's it. The argumentative scholarly culture of 19th-century epigraphy was replaced by a database row. I.Sicily's TEI <commentary> field is the only modern tool that admits prose commentary back into the structured record — and it's the only modern tool that does, which is why I.Sicily looks so much richer than the others on this case study. (For the underlying TEI block structure that supports this — msDesc, physDesc, history, apparatus, named entities, certainty markers — see the Atlas · EpiDoc deep-dive series.)

两个印本可见的拉丁文评注段落，Mommsen 关于大理石优质的论证、Kaibel 对 Mommsen 判断的引用、两位编者对石匠双语能力的解读，在数字时代全部消失。五个现代数据库中没有一个有"编者诠释性评注"这个结构化字段。EDR 仅记 "Textus secundum (6)"，意为"文字从第 6 号参考文献（Manni Piraino 1973）"，仅此而已。19 世纪铭文学的论证性学术文化被一行数据库记录所替代。I.Sicily TEI 的 <commentary> 字段是唯一允许散文评注重新进入结构化记录的现代工具，这也正是为什么 I.Sicily 在本案例研究中看起来比其他库都丰富得多。

§ 2Five database views五个数据库的视角

Five major epigraphic databases (and one notable absence) record this same inscription. Click any to open its actual record.

五大铭文数据库（外加一个引人注目的"缺席"）都收录了这块铭文。点击卡片打开真实记录。

§ 3Side-by-side comparison逐项对比

Same physical object. Five databases. Pick any field to see how they disagree.

同一件实物，五个数据库。每一栏都能看出彼此分歧。

§ 4The seven issues七项问题

§ 5A merge simulator合并模拟

If you fed all five database records into the SDAM LI_ETL deduplication pipeline, here's what would happen.

如果把全部五条记录都喂给 SDAM 的 LI_ETL 去重流水线，结果会是这样：

§ 6Why this hinders the JDH paper's analyses为何这阻碍了 JDH 论文的分析

The Heřmánková–Kaše–Sobotková paper is one of the most rigorous attempts to do macro-history with this kind of data. This single inscription shows where every single one of its careful methodological moves still has to absorb cost.

Heřmánková–Kaše–Sobotková 三人的论文是用此类数据做宏观史最严谨的尝试之一。这一块铭文恰恰展示：他们每一个谨慎的方法论举措，依然要付出代价。

• The "epigraphic habit" curve (Fig 1): one inscription with a 200-year date range contributes a flat plateau of 0.005 inscriptions/year × 200 years to the empire-wide aggregate. Across hundreds of thousands of inscriptions this averages out — but only if the date ranges are consistent. Here, four different ranges from four databases would each shift the local curve.
• "铭文习俗"曲线（图 1）：一条带 200 年区间的铭文，对帝国整体的贡献是 0.005 条/年 × 200 年的均匀平台。汇总到几十万条铭文时，平均化能掩盖个体差异，但前提是日期范围本身一致。这里四个数据库给四个不同范围，每一个都会把局部曲线推向不同方向。
• Provincial distribution (Figs 4–6): EDR's "Sic?" with a question mark cannot be aggregated cleanly with EDH's confident "Sicilia" attribution. A clean province-rank-order plot must either drop "Sic?" (losing data) or treat it as confident "Sicilia" (overcounting).
• 行省分布（图 4–6）：EDR 标注的 "Sic?"（带问号）无法干净地与 EDH 确凿的 "Sicilia" 聚合。要画一张干净的"行省次序图"，要么丢掉 "Sic?"（损失数据），要么把它当作确凿的 "Sicilia"（超统计）。
• Type distribution (Fig 2): I.Sicily classifies this inscription as function.advertisement (EAGLE term 128). EDR doesn't have an explicit type for "advertisement" — it falls under cetera (other). PHI doesn't classify by function at all. The same physical object would be counted as "advertisement" in one analysis, "other" in another, "no type" in a third.
• 类型分布（图 2）：I.Sicily 把这块铭文分类为 function.advertisement（EAGLE 词表 128）。EDR 没有明确的"广告"类型，落入 cetera（其他）。PHI 根本不按功能分类。同一件实物，在一种分析中计为"广告"、在另一种中计为"其他"、在第三种中"无类型"。
• The bilingual question: I.Sicily encodes this as biversion.duplicating Latin + Greek with a structured taxonomy (textLang mainLang="la" otherLangs="grc"). EDCS has latina-graeca as a flat string, EDR has latina-graeca as a flat string, PHI catalogs it under "Greek" because it's in PHI Greek. Aggregating "how many Latin inscriptions are bilingual?" is therefore a labyrinth.
• 双语问题：I.Sicily 把这块铭文编码为 biversion.duplicating（双版本对译），并使用结构化的语言分类（textLang mainLang="la" otherLangs="grc"）。EDCS 标 latina-graeca，但只是平面字符串。EDR 也是 latina-graeca 字符串。PHI 把它归为"希腊文"因为它在 PHI Greek 数据库里。要回答"多少拉丁铭文是双语的"，会陷入字段对照的迷宫。

§ 7What I.Sicily models wellI.Sicily 提供的范例

It is fair to single out one record as a benchmark. I.Sicily's TEI for ISic000470 demonstrates, in a single open-access XML file, what comprehensive data-aware epigraphy looks like:

单独把一条记录作为基准并无不公。I.Sicily 为 ISic000470 提供的那份 TEI 文件，集中展现了"数据意识"完备的铭文学应具备什么：

• Identifier crosswalk — every external ID (EDR, EDCS, PHI ×2, TM, DOI, all the print refs) is recorded in publicationStmt.
• 编号对照表：在 publicationStmt 中记录每一个外部编号（EDR、EDCS、两个 PHI、TM、DOI、所有印本参引）。
• Linked authority files — Pleiades for the ancient place (Panhormus = pleiades 462410), GeoNames for the modern (Palermo = 2523920), Eagle Network vocabularies for inscription type and material, ORCID for editors.
• 权威文件链接：古地名链接 Pleiades（Panhormus = 462410）、现代地名链接 GeoNames（Palermo = 2523920）、铭文类型与材质链接 Eagle Network 词表、编辑者链接 ORCID。
• Original alongside cleaned — <choice><orig>qum</orig><reg>cum</reg></choice>: the EpiDoc convention for keeping both forms recoverable from the same record.
• 原值与清洗值并存：<choice><orig>qum</orig><reg>cum</reg></choice>：EpiDoc 让原始形式与规范形式同时从同一条记录可恢复。
• Material analysis as scholarly contribution — pXRF candidate quarry list with the specific scholar (Alessia Coccato) credited.
• 把材质分析当学术成果：pXRF 候选采石场列表，并具名（Alessia Coccato）。
• Versioned provenance — <revisionDesc> records every edit to the record, dated and signed by ORCID.
• 带版本号的修订记录：<revisionDesc> 记录对该条目的每一次修改，带日期与 ORCID 签名。
• License + DOI — explicit CC-BY 4.0, with a Zenodo DOI for citing this specific record.
• 许可 + DOI：明确 CC-BY 4.0；附 Zenodo DOI，可引用本条目特定版本。

If every Latin and Greek epigraphic database recorded inscriptions to this standard, the JDH paper's data-construction story would be much shorter — and macro-historical aggregation across databases would be tractable instead of artisanal.

如果所有拉丁文与希腊文铭文数据库都按这一标准记录铭文，JDH 论文的"数据构建"叙事会大大缩短，跨库宏观汇总也将由"手工活"变为"可计算"。

§ 8Sources for this case study本案例的出处

The five database records used

使用的五条数据库记录

1. I.Sicily ISic000470 — primary record, with linked TEI EpiDoc XML.
2. ISic000470 raw TEI XML — the source of every detail in this walkthrough.
3. EDR 140617
4. EDCS-22000882
5. PHI 175744 (IGLPalermo 139)
6. PHI 140601 (IG XIV 297)

Print editions cross-referenced

交叉参考的印本

1. Mommsen, T. (1883). CIL X, no. 7296.
2. Kaibel, G. (1890). IG XIV, no. 297.
3. Manni Piraino, M. T. (1973). IGMusPalermo, no. 139.
4. Bivona, L. (1970). ILMusPalermo, no. 74.
5. Wilson, R. J. A. (1990). Sicily under the Roman Empire, p. 314 fig. 266.
6. Tribulato, O. (2011 / 2012). On the bilingualism of this inscription.
7. Consani, C. (2021). On the inscription as the work of a Latin speaker translating literally into Greek.

The methodological frame

方法论框架

1. Heřmánková, P., Kaše, V., & Sobotková, A. (2021). Inscriptions as data: digital epigraphy in macro-historical perspective. Journal of Digital History, 1(1). doi.org/10.1515/jdh-2021-1004 — see the Paper Edition for a section-by-section walkthrough.
2. Kaše, V., Sobotková, A., & Heřmánková, P. (2023). Modeling Temporal Uncertainty in Historical Datasets. CHR 2023. CEUR-WS

Companion editions

配套版本

1. Paper Edition — JDH 2021 walkthrough.JDH 2021 论文逐节导读。
2. Reference Edition — deep technical companion to all 37 SDAM repositories.SDAM 37 个仓库的深度技术配套。
3. Visual Edition — 19-slide interactive intro to ETL.ETL 的 19 张交互式幻灯片导览。
4. Landing page — choose your starting point.选择入口。