Reference sources — operational policy for translation

This document defines, operationally and legally, how copyrighted and restricted-license scholarly sources may be used during translation of the People's Open Bible. It is a companion to DEUTEROCANONICAL.md (scope/strategy) and METHODOLOGY.md (pipeline).

The core claim: consultation is not reproduction. Scholars and translators have always read the leading critical editions, let those editions inform their judgment, and then produced their own fresh work. That is what every serious modern translation does. This document translates that long-standing practice into concrete rules for our AI-assisted pipeline.

Position assessment — where we stand per book

Reassessed 2026-09-04. The former “optimal,” “near-optimal,” “definitive,” and “sole witness” labels are withdrawn. They confused an available translation input with a demonstrated best critical text. The former numerical prediction of how little Tobit would differ from a Qumran-first translation was untested. This correction does not change the consultation/reproduction policy below.

The IOSCS edition guide distinguishes Swete's edition from broader critical work. The IAA also records 7Q2, a Greek Letter of Jeremiah manuscript: it is plainly not the same thing as a modern Swete volume. See the OT/NT coverage audit for the catalogue-backed comparison plan. None of these rows certifies a completed critical collation or changes the chosen canonical text.

Operational note. ADE and ADA are now best thought of as aggregate source-stream labels. The active translation units are:

• ESG for the Esther additions (A–F)
• PAZ, SUS, and BEL for the Daniel additions

This keeps the source-preparation history auditable while giving Phase 9 clean reader-facing drafting targets.

Net assessment — corrected 2026-09-05: Available working texts do not establish an optimal critical source. The book-specific comparisons above remain necessary; consultation does not recover missing original-language material or settle source rights. This replaces an obsolete paragraph that contradicted the September 4 reassessment. The restoration research log records the correction and the evidence-linked program history.

The three zones

Every scholarly source sits in one of three zones with respect to our pipeline. The zones are defined by how the source influences the output, not by whether anyone looks at it.

Zone 1 — Vendored (safe as both reference and Vorlage)

Clean-licensed sources that we can commit into the repository and use as primary input for translation. Output is free to derive from these.

Zone 1 content MAY appear in the prompt context used to generate the English translation and MAY influence specific word choices.

Zone 2 — Consulted reference (aware of, not derived from)

Copyrighted or restricted-license scholarly sources that inform our judgment without appearing in the output. Examples:

• Yadin 1965, The Ben Sira Scroll from Masada (editio princeps) — Edition of the Masada Hebrew witness, preserving portions within Sir 39:27–44:17. When local notes are available, tools/yadin_masada.py::lookup(chapter, verse) supplies edition reference material for requests within this outer span. Page-index associations do not establish that a requested verse or reading survives. Sirach 4, 49 and 51 are not Masada coverage; the previously stated 413-verse figure is not verified manuscript attestation. Local notes were unavailable at the 2026-10-04 check. Nothing from this consulted edition is vendored; the Zone 2 policy remains in force.
• Fitzmyer, Discoveries in the Judaean Desert Vol. XIX (Qumran Tobit) — reconstructed Aramaic of 4Q196-200
• Beentjes 1997, The Book of Ben Sira in Hebrew — critical edition of all recovered Hebrew Sirach
• Skehan & Di Lella 1987, Anchor Bible 39 — English Sirach with critical apparatus
• Göttingen LXX critical editions (Hanhart and others)
• Rahlfs-Hanhart 2006 Stuttgart revised LXX — reading text
• Leon Levy DSS Digital Library images — Qumran photographs

Rules for Zone 2:

1. Access is legitimate. A translator (human or AI) may read these works, the same way any paid scholarly translator does.
2. The output must not track the source's creative expression. Specifically: where the Zone 2 source supplies a reconstruction of missing or damaged text (filled-in letters, chosen readings among variants, editorial ordering), our English rendering must not word-for-word follow that reconstruction. It may be informed by it — confirming a reading the LXX also attests, or flagging a suggested divergence — but the primary anchor of our English is a Zone 1 source.
3. Fact-level citation is allowed. Footnotes of the form "Qumran 4Q196 supports the Long Recension reading here" cite a fact about what the manuscript attests, not the scholar's creative expression. Feist v. Rural Telephone Service (1991) establishes facts are not copyrightable.
4. Nothing from Zone 2 is committed to the repository. Vendoring would propagate the source's license to every downstream consumer of POB. Consultation is private to the translator's workspace.

Zone 3 — Forbidden during drafting

A small set of sources that we actively avoid consulting:

• Translations under commercial copyright (including NKJV, NIV, NLT, and ESV). We do not expose these to the drafter — our English must not track their word choices, and the cleanest drafting boundary is not to include them in an LLM prompt.
• Sources where the license explicitly prohibits even internal use (none currently).

An explicitly licensed commercial translation may be used after drafting in an isolated evaluation process. tools/build_translation_divergence.py accepts a private, gitignored NKJV/NIV/NLT bundle and emits only numeric similarity scores. It never emits the licensed wording, never feeds that wording back into the drafting prompt, and never lets a commercial translation determine the review-priority score. This preserves the clean-room drafting boundary while allowing licensed comparative research.

How this flows into the translation prompt

The translator agent (AI drafter + human/AI reviser) receives a structured context per verse. The context includes all Zone 1 and Zone 2 sources available for that verse, with each labeled by zone.

Context block shape

verse_ref: "SIR 1:1"
zone_1_primary:
  greek:
    source: "Swete LXX (our OCR, CC-BY 4.0)"
    text: "Πᾶσα σοφία παρὰ Κυρίου καὶ μετ᾽ αὐτοῦ ἐστιν εἰς τὸν αἰῶνα."
  hebrew:  # present only for SIR, TOB, and 1ES with MT parallel
    source: "Sefaria Ben Sira / Kahana (CC0)"
    kind: "direct_hebrew"
    text: "כָּל חָכְמָה מֵיהֹוָה, וִעמּוֹ הִיא לְעוֹלָמִים."
    note: "Kahana composite edition of Cairo Geniza Hebrew. This is a working source, not one physical manuscript or an adjudicated earliest Hebrew text."
zone_1_secondary:
  - "First1KGreek (CC-BY-SA): <Greek variant, if any>"
  - "Rahlfs-Hanhart (NC, consultation only): <Greek reading>"
  - "Swete-Amicarelli (GPL, consultation only): <Greek reading>"
zone_2_consult:  # reference only -- do not reproduce or track word-for-word
  - name: "Beentjes 1997"
    book_scope: "SIR"
    guidance: "Critical edition of Hebrew Sirach MSS A-F. Consult if Kahana reading seems wrong or if this verse is in a Kahana gap."
  - name: "Skehan & Di Lella 1987 (Anchor Bible 39)"
    book_scope: "SIR"
    guidance: "English Sirach with critical apparatus. Consult for argued interpretive cruxes -- do NOT track their English phrasing."
instructions_to_translator:
  primacy: >
    For Sirach verses with a direct_hebrew witness, translate from
    the Hebrew (Zone 1 primary hebrew). Use the Greek as a consistency
    check and for verses where the Hebrew is lost/damaged. Where Zone
    2 scholarship suggests a different reading from our Zone 1 Hebrew,
    document the disagreement in the apparatus but keep the output
    anchored in Zone 1.
  forbidden: >
    Do not reproduce word-for-word translations from Zone 2. Footnote
    their conclusions as facts where relevant.

For Tobit specifically, the Zone 2 block carries Fitzmyer's DJD XIX reconstruction of 4Q196-200 where the verse falls in the ~20% overlap. The translator knows the reconstruction exists, can factor in its textual-critical verdict, and can note disagreements in footnotes — but the English output anchors in the LXX Long Recension (Zone 1 Greek) + Neubauer (Zone 1 Hebrew back-translation).

Implementation

tools/hebrew_parallels.py already returns Zone 1 Hebrew/MT data per verse. It is extended to also list the Zone 2 sources applicable to each book, so a translation-phase prompt builder can assemble the context block mechanically without per-book special-casing.

The translator-prompt builder itself (Phase 8-C deuterocanon drafting work, not yet written) will:

1. Load the Greek verse from sources/lxx/swete/final_corpus_adjudicated/, with translation-ready overrides from sources/lxx/swete/final_corpus_normalized/ when a known numbering contamination has been formally normalized
2. Call hebrew_parallels.lookup(book, ch, vs) for Zone 1 Hebrew/MT
3. Pull Zone 1 secondary reference readings where available
4. Inject the Zone 2 "consult" block from the book's Zone 2 registry
5. Wrap with the doctrine and style instructions from DOCTRINE.md
6. Call the translator model (GPT-5.4) with the assembled prompt

Derivative-work exposure — concrete test

Before shipping any verse, the reviser asks three questions:

1. Anchor: is the English primarily supported by a Zone 1 source?
2. Zone 2 independence: if the Zone 2 reading were redacted from the translator's context, would the English still be defensible from Zone 1 alone? (This is the "consulted, not derived" test.)
3. Fact vs. expression: is anything we reproduce from Zone 2 a fact (manuscript reading, dating, scholarly conclusion) rather than creative expression (reconstructed wording, editorial phrasing)?

A verse passes if all three answer "yes." A verse that fails gets re-drafted from Zone 1 alone and re-reviewed.

Rationale — why this is the right path

For accuracy. A translator who consults every leading scholarly edition produces better work than one who works from a single source. This is how every serious modern translation is produced (NRSV, NABRE, Orthodox Study Bible). We match that standard.

For license cleanliness. CC-BY 4.0 on our output requires that output be free to redistribute. Derivative works of Zone 2 sources are not free to redistribute. The three-zone discipline keeps our output demonstrably downstream of only Zone 1, while still benefiting from Zone 2 through consultation.

For honesty with readers. Each verse's provenance can be reported in the per-verse YAML with full disclosure: which Zone 1 sources anchored the translation, which Zone 2 sources informed it, and whether any scholarly disagreement is footnoted. Readers can audit exactly what shaped each rendering.

For scalability. The pipeline is mechanical. Zone assignments are declared once per source; translator prompts assemble automatically; derivative-work checks are explicit. No ad-hoc judgment per verse.

Confidence rubric for corpus adjudication

Every verse in the scan-adjudicated corpus carries a confidence level. The rubric below is applied strictly by the adjudicator (AI or human):

HIGH — reading can be defended rigorously:

• Every character of the verse is clearly readable in the scan, including all diacritics (accents, breathings, iota subscripts)
• The reading output exactly matches what is printed, or matches one of the independent transcription candidates exactly
• No visible smudging, faded ink, torn paper, or ambiguous letterforms in the verse area

MEDIUM — reading is defensible but has known ambiguity:

• Verse is substantively readable, but specific characters are ambiguous (e.g. ε vs ω on a worn letter, acute vs circumflex, α vs δ)
• A best-guess call has been made but the scan permits an alternative reading a specialist might prefer
• Verse boundary or punctuation placement is genuinely ambiguous

LOW — reading cannot be visually verified:

• Verse is not legibly visible on the provided scan page
• Severe damage, ink bleed, or missing text
• Reading relies on candidates alone, without visual verification

Calibration principle: it is better to mark MEDIUM and be correct about uncertainty than to mark HIGH and be overconfident. Medium is a disclosure layer, not a defect. A published verse with source_confidence: medium is a trust signal to readers that the source-text reading has a specific, documented uncertainty the translator and reviewer were aware of.

Per-book Zone 2 registry

(Maintained alongside this document; may grow as scholarship is identified.)

These are for private consultation during translation, cited by name in footnotes where their fact-level conclusions matter, and never reproduced in POB output.