v2.1.0: Corpus Expansion — Cicero, Seneca, Fronto

Corpus Expansion

The database now spans 68 BC to 800 AD with 8,112 letters across 60 collections — up from 6,914 letters in v2.0.0.

New Collections

Collection	Letters	Period	Source
Cicero, Ad Atticum	431	68-43 BC	Latin Library + Winstedt/Gutenberg
Cicero, Ad Familiares	425	62-43 BC	Latin Library + Shuckburgh/Gutenberg
Seneca, Epistulae Morales	124	62-65 AD	Latin Library + Gummere (Loeb)
Fronto, Correspondence	207	139-166 AD	Haines (Loeb) via Wikisource
Ignatius of Antioch	7	~110 AD	New Advent (ANF)
Cyril of Alexandria	4	430-433 AD	New Advent (NPNF)

Translation Quality (continued from v2.0.0)

• Additional 100+ fabrications found and fixed via full collection audits
• Gregory the Great: 30 more fabrications fixed (full 399-letter audit)
• Libanius: 55 more fabrications fixed (full 817-letter audit)
• Augustine: 15 misaligned translations fixed (full 161-letter audit)
• All remaining collections fully audited (1,545 letters, 14 fixes)
• Total collections 100% audited: all 60
• Round 5 verification: <2% fabrication rate

Stats

Metric	v2.0.0	v2.1.0
Total letters	6,914	8,112
With English	6,824	8,022 (98.9%)
Collections	54	60
Date range	100-800 AD	68 BC - 800 AD
Collections audited	100%	100%

v2.0.0: Scholarly-Quality Translation Overhaul

What Changed

Major translation quality overhaul responding to scholarly community feedback. Every translation with Latin/Greek source text has been independently verified.

Translation Quality

• Benchmark score: 3.54 → 4.38 (exceeds 4.0 target and 4.14 human reference)
• 5,162 letters fully audited (100% of letters with source text)
• ~217 fabricated translations found and fixed across all collections
• New scholarly translation pipeline emphasizing fidelity over readability
• 54 author-specific voice profiles for accurate register

Data Fixes

• 3,026 AI metadata headers ("From:/To:/Date:") stripped from all translations
• 1,628 New Advent advertising texts stripped from 19th-century translations
• 535+ Isidore template/paraphrase translations replaced with real translations from Greek
• 190 Hormisdas boilerplate context blocks replaced with real translations from Latin
• 134 duplicate/misattributed records removed (99 Augustine staging, 34 Gregory duplicates, 1 Gregory IX misattribution)
• 343 Isidore letters re-OCR'd from Patrologia Graeca page scans using Gemini Vision

UI Improvements

• Translation source badges on every letter page (green = human, amber = AI-assisted)
• Scholarly mode hides AI translations, shows only pre-existing scholarly sources

Infrastructure

• scripts/scholarly_translate.py — new translation pipeline
• scripts/benchmark_translations.py — blind comparison scoring system
• scripts/scholarly_validate.py — automated quality checks
• scripts/ocr_patrologia_graeca.py — Gemini Vision OCR pipeline for PG volumes
• data/author_prompts.json — 54 author voice/style profiles
• Deployment docs in CLAUDE.md (rsync to cvg-primary)

Database Stats

Metric	v1.1.0	v2.0.0
Total letters	7,049	6,914 (duplicates removed)
With translation	~7,000	6,824 (98.7%)
Fabricated translations	~217	0
AI headers	3,026	0
Benchmark score	3.54	4.38
Collections audited	0%	100%

Audit Coverage

Every collection with Latin/Greek source text has been 100% audited:
Augustine (161), Isidore (630), Symmachus (587), Gregory (399), Cassiodorus (477), Ennodius (297), Hormisdas (249), Libanius (817), Basil (325), Pliny (365), Sidonius (171), Venantius (100), and 14 smaller collections (559).

v1.1.0: Latin/Greek Originals, Reddit Features, Performance Fixes

Latin/Greek coverage 40% to 75%. Manuscript scan links. Places and People pages. Scholarly mode toggle. 7 JS performance fixes. See full changelog on GitHub.

v1.0.0 — Roman Letters Network Dataset

Roman Letters Network Dataset v1.0.0

7,049 letters from the late Roman Empire (100-800 AD)

First public release of the complete dataset, including:

• roman_letters.db — Full SQLite database (61 MB) with letter texts, author data, and metadata
• letters.csv — All letters with sender, recipient, dates, locations, topics, and summaries
• authors.csv — 1,850 historical figures with names, roles, coordinates, and biographies
• collections.csv — 54 letter collections with source URLs and archive links
• README.md — Dataset documentation and schema reference

Key Stats

• 7,049 total letters from 54 collections
• 4,410 translated to English for the first time
• 1,850 people identified
• Languages: English, Latin, Ancient Greek

Citation

Vander Galien, Craig. (2026). Roman Letters Network: A Digital Corpus of Late Antique Correspondence. Zenodo. https://doi.org/10.5281/zenodo.19142059

License

CC-BY 4.0