Method

Generated 2026-11-14. Counts: 8,509 works · 572 dossiers · 1,756 works with machine-readable text · 14,869 manuscript/print records.

The name is Tergwame (ትርጓሜ), Geʿez for “translation” — the Ethiopian Christian word for rendering scripture and doctrine into another tongue, cognate with targum. The project does that in the other direction too: it records what survives, what English already exists, and what this pipeline has drafted.

1 · Scope

Greek Christian literature from the first century to the fall of Constantinople (1453); Coptic literature from the third to the fourteenth century, including Copto-Arabic; Ethiopic (Geʿez) literature from late antiquity to the eighteenth century. The register counts works, not editions, manuscripts or translations. A work translated twice is one row; a work surviving in five manuscripts is one row, with the witnesses on the Holdings page.

2 · What each status means

Source-by-source grading

Greek status comes from the CTS translation inventory (First1KGreek + Perseus); Coptic status from each Coptic Scriptorium document's translation field; Ethiopic status from Beta Masaheft's translation bibliography and the PEMM English text table. Curated dossiers override the catalogue where they are more specific.

3 · The dossier

Every candidate is a record with nine fields, each traceable:

1. Identification — author, original title, English descriptive title, date, genre, and the Clavis number or catalogue id where one was verified.
2. Translation evidence — the exact series, catalogue queries and negative results behind the status.
3. Existing translations — English volumes, with translator and year; non-English versions are named in the evidence but never counted as English.
4. Electronic text — at least one verified open TEI/HTML source where one exists, with its licence; otherwise an explicit statement that the text is print- or subscription-only.
5. Harvard Hollis — the Primo query, hit count, first records and a Chicago-style citation.
6. Sources checked — every URL used.
7. Notes — caveats, competing editions, and why the work is a gap.

4 · The ingestion pipeline

The pipeline is deliberately separate from the website: /srv/christian-pipeline for code and normalized works, /srv/christian-corpus for the upstream mirrors and catalogue. The ingestion stage stops at a clean source text; a second stage now segments, translates with DeepSeek V4 Flash (deepseek/deepseek-v4-flash:floor), writes commentary, runs QA and publishes. The full programme is in docs/CHRISTIAN-TRANSLATION-PLAN.md.

The apparatus on each passage is two layers, and they answer different questions. Commentary units are model-written and say what the passage says; each one names the kind of work it is doing, so a reader can tell a lexical decision from a scriptural echo, a piece of background from a judgement about transmission. Notes are derived from the record with no model call: what the passage is structurally doing, where the source is fragmentary, which glossary term now carries a fixed rendering, which check fired. A passage whose units all came back as one kind is treated as a failed call and regenerated, because a single undifferentiated note is exactly what this apparatus exists to avoid.

Runbook

cd /srv/christian-pipeline
python3 pipeline/ingest_first1k.py
python3 pipeline/ingest_perseus.py
python3 pipeline/ingest_coptic.py
python3 pipeline/ingest_coptot.py
python3 pipeline/ingest_betmas.py
python3 pipeline/ingest_pemm.py
python3 pipeline/ingest_hmml.py
python3 pipeline/ingest_ia_geez.py
python3 pipeline/enrich_hollis.py

# translation (Coptic first)
python3 pipeline/inventory.py --tradition coptic --exclude-coptot
python3 pipeline/segment.py --all
python3 pipeline/queue_translate.py --tradition coptic --exclude-coptot \
  --workers 3 --budget 5 --daily-cap 5

# rebuild the site data
cd /var/www/naxi
python3 scripts/christian/build_registry.py

Model and cost review

The models page measures low-cost OpenRouter options and the local Gemma 4 deployment on 65 aligned segments from this corpus. DeepSeek V4.1 Flash and V4 Flash lead at chrF 61.0; GLM 5.3 Flash is the value option at 55.4 for about a third of the cost; Gemma 4 31B scores 48.9. Translating all 606 transcribed non-full works (161 MB) costs roughly $11–31 at list prices, or ~$18 of electricity locally. Review, not inference, is the real expense.

5 · The scan lane

Where no transcript exists, the question is whether the work has been photographed and whether that scan can become text. The scan register records 121 works: 65 with an open scan, 21 restricted to viewing, and 35 with no scan located after the named checks. Across them, 178 candidates were verified.

What the OCR tests showed

Quality is scored by function-word and lexicon match, not script ratio: a Latin page OCR'd as Greek can score a Greek ratio of 1.0 while containing no Greek. OCR is not run in bulk; the adapters download one sample page, measure it, and leave the full run to a reviewed job.

6 · Licences and what is withheld

7 · Known limits and traps

• Beta Masaheft is not a Christian-only file. 4,159 of its 10,754 work records are Islamic (IslHornAfr). The pipeline excludes them; the Christian remainder is 6,594 records, of which only 372 carry edition text.
• A catalogue record is not a text. Most Ethiopic works have an incipit and a bibliography, not a transcription. The has_text field is the honest signal.
• “None” is a search result. The dossier names the series and catalogue queries. A translation inside a journal article or a small-press volume can be missed; the register is designed to be corrected, not to be final.
• Expansion packs are not coverage. Modern “commentary chains” (ACCS, The Church's Bible) are fresh selections of translated fragments, not editions of the Greek catenae. They are not counted as translations.
• Non-English versions are not English. A French Sources chrétiennes volume or a German CSCO translation is recorded as evidence of scholarly attention, and does not change the English status.
• Demand-side honesty. The register does not claim that every gap should be translated. It records extent, witnesses and textual quality so that a later decision can be made on evidence.

8 · Reproduce

All inputs are public and unauthenticated. The mirrors are shallow git clones; the PEMM and HMML downloads are checksummed by the adapters; the Hollis cache is under ~/.cache/christian-pipeline/hollis. The site data is generated by one script and is deterministic apart from the timestamp.

Code: /srv/christian-pipeline/pipeline/ · /var/www/naxi/scripts/christian/build_registry.py · /var/www/naxi/api/routes/christian.py. The Tibetan project's pipeline page is the model.