From the repository

Apocrypha sources

How every verse in the 17-book deuterocanonical section was sourced and verified, with reader-facing verification steps.

Source: APOCRYPHA_PROVENANCE.md on GitHub

Apocrypha provenance — how every verse was sourced and verified

Public-facing account of how each of the 17 books in the Cartha Open Bible deuterocanonical section arrived in the corpus, verse by verse, and how a reader can verify our text against public-domain scholarship without taking anything on trust.

Audience: readers, scholars, translators, librarians, and anyone asking "did you invent any of this?" The short answer is: no — and here is what you can check.

Principle

Every deuterocanonical verse in the People's Open Bible carries three things on its per-verse YAML:

  1. A Greek source text (the source.text field) with a named edition (source.edition) and the page numbers that edition was drawn from.
  2. A confidence label at the source-text layer (high / medium / low) that reflects our scan-adjudication rubric.
  3. A drafting record naming the model, prompt, and timestamp that produced the English, with lexical decisions exposed.

Every piece is auditable against publicly-downloadable documents. The sections below walk through the four sub-pipelines used across the 17 books.

Sub-pipeline 1 — Swete diplomatic Vaticanus (12 of 17 books)

Books: Tobit, Judith, Additions to Esther, Wisdom of Solomon, Sirach, Baruch, Letter of Jeremiah, Prayer of Azariah + Song of the Three, Susanna, Bel and the Dragon, 1–4 Maccabees, 1 Esdras.

Source: Henry Barclay Swete, The Old Testament in Greek According to the Septuagint, 3 vols. (Cambridge University Press, 1909–1930). Public domain (Swete died 1917). Archive.org identifier: theoldtestamenti03swetuoft_202003.

How we processed it:

  1. AI-vision OCR of each scan page via Azure GPT-5.4 vision with a prompt specialized for polytonic Greek typography.
  2. Multi-witness adjudication: for every verse where our OCR disagreed with at least one of three independent transcriptions (First1KGreek TEI, Rahlfs-Hanhart Marvel.Bible digital, Amicarelli BibleBento), we ran a scan-grounded adjudicator that looked at the actual page image and returned a verdict. The verdict identifies which witness matched the printed page, or marks the reading as "neither" when we re-transcribed directly from the scan.
  3. Confidence rubric:
    • High = every character is clearly visible on the scan; reading matches a candidate exactly or is a confident fresh transcription; no smudges, fading, or ambiguous letterforms in the verse area.
    • Medium = verse substantively readable but specific characters ambiguous; best-guess call made; scan permits an alternative a specialist might prefer.
    • Low = verse not legibly visible on the scan; severe damage or relying on candidates without visual verification.
  4. Rescue passes: initial low-confidence verses were re-run with higher-resolution scans, multi-source Hebrew parallel consultation (Kahana Ben Sira, Neubauer Tobit, WLC for 1 Esdras parallels), explicit confidence-rubric prompts, and content-based per-verse page mapping when running-head parsing missed pages.

Final state (detailed in sources/lxx/swete/QUALITY_BENCHMARK.md):

  • 99.3% of adjudicated verses at high confidence (3,442 / 3,464)
  • 86.1% functional agreement with First1KGreek as an independent validation oracle
  • 6 books reach 100% high confidence (Wisdom of Solomon, 3 Maccabees, 4 Maccabees, Letter of Jeremiah, Additions to Daniel, Judith)

A reader can verify any Swete-sourced verse by:

  1. Pulling the scan page from archive.org (link embedded in source.pages).
  2. Comparing the printed Greek against our source.text for that verse.
  3. Inspecting source.confidence to see how we rated our own reading.

Sub-pipeline 2 — Prayer of Manasseh (Charles 1913 APOT Vol 1)

Book: Prayer of Manasseh (MAN), 15 verses.

Prayer of Manasseh is not in Codex Vaticanus, so Swete's diplomatic edition does not contain it. Our normal pipeline does not apply. We used a dedicated source path documented in detail in PRAYER_OF_MANASSEH.md.

Source: R. H. Charles (ed.), The Apocrypha and Pseudepigrapha of the Old Testament in English, Vol. 1 (Oxford: Clarendon Press, 1913). Public domain (author d. 1931; pre-1929 US publication). Archive.org identifier: theapocryphaandp01unknuoft. Our working PDF has SHA-256 hash recorded in sources/lxx/prayer_of_manasseh/MANIFEST.md.

How we processed it:

  1. OCR of Charles pp. 636–640 (the Prayer of Manasseh main-text + apparatus + footnotes) via Gemini 3.1 Pro Preview with the strict text-extraction prompt (systemInstruction + responseMimeType: text/plain). 5 pages, 0 failures, 1,480 Greek characters across the apparatus + footnote lemmata.
  2. Greek text reconstruction: Charles's edition prints the English translation as main body, with Greek scattered through the critical apparatus and footnotes (he cites "A" for Codex Alexandrinus, "T" for Codex Turicensis, "Const. Apost." for Apostolic Constitutions 2.22, "Syr." for Syriac, "Lat." for Latin Vulgate, "Moz." for Mozarabic Psalter). We passed the OCR output to Gemini 3.1 Pro with an explicit scholarly prompt: "reconstruct the continuous Greek text, verses 1-15, as witnessed by Codex Alexandrinus (A) per Charles's apparatus." The Greek text itself is public domain by age (~2000 years); Charles's apparatus lemmata are public domain by 1913 publication. Our reconstruction is at sources/lxx/prayer_of_manasseh/corpus/MAN.jsonl — 15 verses, 292 Greek words.
  3. Rahlfs-Hanhart cross-check: we verified our reconstruction against the Rahlfs-Hanhart 2006 Ode 12 digital text (a Zone 2 consultation source, NC-licensed, never copied into our output). Word-overlap per verse was 62–92%, with mismatches traceable to known manuscript-family verse-division conventions and minor Alexandrinus-vs-eclectic orthographic variants. This does not introduce Rahlfs text into our corpus; it simply confirms that our reconstruction sits inside the expected range of variation for this prayer.
  4. Direct visual verification against Charles 1913 pp. 636–640 (2026-04-22): every verse division was checked against Charles's editorial numbering in the margin of his printed edition; every substantive word was checked against his apparatus lemmata or against the expected Codex Alexandrinus reading. Specific findings recorded in sources/lxx/prayer_of_manasseh/corpus/MAN.jsonl v1.reconstruction_note:
    • Verse boundaries match Charles's editorial numbering exactly.
    • v7's "inserted" clause ("Thou, O Lord, according to thy great goodness hast promised repentance...") is correctly omitted per Charles's apparatus: "Const. Apost., Syr., Lat., Moz.: om A T".
    • v8 reads ἐμοὶ τῷ ἁμαρτωλῷ (Codex A) rather than Rahlfs's ἐπ᾽ ἐμοὶ τῷ ἁμαρτωλῷ.
    • v9 originally had passive ἐπληθύνθησαν; this was corrected on 2026-04-22 to active ἐπλήθυναν per Charles's apparatus for Codex A: "κυριε επληθυναν αι ανομιαι μου". The v9 English draft was re-run against the corrected Greek.

Licensing note: MAN's source.edition field is charles-1913-apot-vol1, not lxx-swete-1909. The drafter (tools/draft.py) has a book-specific override for this so the provenance is honest about where the Greek came from.

Sub-pipeline 3 — Psalm 151 (Swete, but manually located)

Book: Psalm 151 (PS151), 7 verses.

Psalm 151 is in Codex Vaticanus (at the end of the Psalter), but it sits at Swete vol 2 page 432 which our automated Phase-8 OCR hadn't reached yet. The book had been declared in tools/lxx_swete.py with a zero page range; the page discovery was tail-of-phase-9 work.

Source: Swete LXX vol 2 page 432. The full Psalm 151 — running head [CLI] with the characteristic intro "Οὗτος ὁ ψαλμὸς ἰδιόγραφος εἰς Δαυείδ καὶ ἔξωθεν τοῦ ἀριθμοῦ, ὅτε ἐμονομάχησεν τῷ Γολιάδ" ("This is the autograph psalm of David, outside the number, when he fought Goliath") — sits cleanly on that page.

How we processed it:

  1. Located the page by browsing Swete's Psalter pagination (end of Psalter = p432 based on running heads Ψαλμοί CXLIV–CLI).
  2. OCR via tools/transcribe_source.py --source swete --vol 2 --page 432 using Azure GPT-5.4 vision with our standard Swete prompt. Clean output, 1,393 characters, all 7 verses with verse markers preserved.
  3. Split into PS151.jsonl with 7 verse-keyed records.
  4. Mirrored into sources/lxx/swete/final_corpus_adjudicated/PS151.jsonl so the existing drafter pipeline resolves PS151 via lxx_swete.iter_source_verses("PS151").
  5. Drafted all 7 verses via the standard GPT-5.4 drafter with no book-specific overrides (PS151 is a regular Swete-sourced book; its only unusual feature was not being in the original DEUTEROCANONICAL_BOOKS page-range registry).

Licensing: PS151's source.edition is the standard lxx-swete-1909 — same as every other Swete-sourced deuterocanonical book.

Sub-pipeline 4 — Additions to Daniel split

Books: Susanna (SUS, 64 verses), Bel and the Dragon (BEL, 42 verses), Prayer of Azariah + Song of the Three (PAZ, 67 verses).

These three works are historically aggregated as the "Greek Additions to Daniel" (book code ADA in Swete) but were split in our translation directory per scholarly convention. Each is drafted under its own slug. The source Greek is the same Swete vol 3 adjudicated corpus as the other LXX books; the split is purely organizational.

Drafting layer (common to all 17 books)

English drafts are produced by tools/draft.py with GPT-5.4 (Azure OpenAI, deployment gpt-5-4-deployment). For each verse the drafter:

  1. Reads the adjudicated Greek source and verse number.
  2. Loads the relevant POB doctrinal and translation-philosophy excerpts (DOCTRINE.md, PHILOSOPHY.md).
  3. Loads any per-book apparatus or witness parallels (e.g., Hebrew parallels for Sirach / Tobit / 1 Esdras from sources/lxx/hebrew_parallels/).
  4. Produces a schema-valid YAML with:
    • English translation
    • Translation philosophy (optimal-equivalence by default)
    • Per-word lexical decisions with rationale
    • Footnotes for alternative readings
    • Model + prompt-hash provenance for reproducibility

Every English word in every deuterocanonical YAML is traceable to a specific source edition + adjudicator verdict + drafter invocation.

What a reader can check, per verse

Take any deuterocanonical verse page on cartha.com/peoples-open-bible and click through to its provenance. You will see:

  • The Greek source text (source.text)
  • The edition it came from (source.edition)
  • The scan page numbers (source.pages)
  • The adjudication verdict and confidence (in the adjudication sub-object for verses that were scan-adjudicated)
  • Our English rendering (translation.text)
  • Every significant lexical choice (lexical_decisions)
  • The AI model and prompt that drafted the verse (ai_draft)

The scan pages are linkable to archive.org. The lexical decisions are critique-able. Nothing is hidden behind "trust us."

Honest disclosure of residual work

Not all 17 books are claimed as final:

  • Consistency lint (tools/consistency_lint.py) has not been run across the deuterocanonical set yet. It enforces uniform handling of recurring terms (e.g., κύριος rendering, proper-name transliteration). Known pending task; does not affect per-verse correctness, only cross-verse consistency.
  • Systematic revision-methodology pass per REVISION_METHODOLOGY.md has not yet been run. Phase-1 Pauline drafts have had full revision; the Apocrypha are at first-pass quality.
  • Scan-adjudication polish on the remaining 17 medium-confidence verses (mostly proper-name transliteration lists in 1 Esdras + some Sirach + Greek Esther Additions) is tracked in REVISION_LATER.md.

A reader relying on a specific verse for scholarly citation should consult the per-verse source.confidence field. Any verse marked "high" has been visually verified; "medium" means the adjudicator was substantively sure but not certain; "low" means we could not visually verify the reading and defaulted to a well-attested scholarly form.

The three-zone policy (scholarship consultation)

Consultation of copyrighted scholarly editions is governed by REFERENCE_SOURCES.md. In brief:

  • Zone 1: public-domain or CC-licensed sources we vendor into the repo and derive from (Swete, Charles 1913, Sefaria Kahana, WLC, First1KGreek, etc.).
  • Zone 2: copyrighted scholarly editions we consult without copying (Beentjes Hebrew Sirach, Fitzmyer DJD XIX Qumran Tobit, Yadin Masada scroll, Göttingen critical LXX, Rahlfs-Hanhart, Skehan & Di Lella, etc.). We may cite fact-level findings ("4Q196 attests the Long Recension here") but never reproduce their text.
  • Zone 3: modern commercial English translations (NIV, NLT, ESV, NRSV). Not consulted during drafting to prevent derivative-work exposure.

The three-zone discipline is why our output is cleanly CC-BY 4.0-redistributable even though our translators consulted modern scholarship during drafting, the same way every modern translation always has.

Summary

The Apocrypha section of the People's Open Bible is:

  • Translated from original Greek (not retranslated from other English translations).
  • Sourced from public-domain editions (Swete 1909, Charles 1913).
  • Scan-grounded per verse (99.3% at high confidence; medium and low verses explicitly labeled).
  • Reproducibly drafted (model + prompt-hash + source page list on every verse).
  • CC-BY 4.0 licensed (use freely with attribution).

What you read in the app is what the public-domain page prints, as rendered into English by a documented, reproducible pipeline. Nothing is black-box; everything is auditable.