Analysis Tools

Manuscript Traditions

BibCrit works across seven major textual witnesses to the Hebrew Bible and New Testament, each with its own transmission history and scholarly edition.

How Analysis Works

Each BibCrit tool draws on Claude's extensive training in biblical studies, ancient languages, and the primary literature of textual criticism. When you submit a passage or book, Claude consults that training knowledge to apply the same analytical frameworks used in critical scholarship — Tov's divergence taxonomy for the Divergence Analyzer, Aejmelaeus's translation-technique categories for the Scribal Profiler, the standard Vorlage reconstruction methodology for the Back-Translation Workbench, and so on. See Scholarship & Methodology for the specific sources that inform each tool.

Where the analysis touches manuscript data that can be verified — such as the presence of a Dead Sea Scrolls witness, the alignment of a patristic citation, or the numerical value in a specific tradition — Claude cross-checks its response against its knowledge of the standard critical editions. Results stream to your browser in real time so you can follow the analysis as it develops. Frequently analyzed passages load from a shared cache, so you see prior results immediately.

All analysis is produced from Claude's training knowledge. BibCrit does not query external databases at runtime. Results should be cross-checked against primary sources — BHS/BHQ, the Göttingen LXX, the DSS transcriptions — before citation in peer-reviewed work.

Understanding Confidence Scores

Some results carry a confidence badge or bar. It is the model's own estimate, not a calibrated probability. It is not derived from the corpus or from any count of scholarly opinion, and it says nothing about whether the claim is correct. In our own tests, different tools returned the same figure for very different questions. See Methods.

Scholarship & Methodology

BibCrit's prompts are informed by the standard reference works and methodological frameworks of LXX and textual criticism scholarship. Key sources include:

• Emanuel Tov — Textual Criticism of the Hebrew Bible (3rd ed., Fortress, 2012): the standard taxonomy of divergence types, Vorlage reconstruction methodology, and scribal error classification.
• Anneli Aejmelaeus — On the Trail of the Septuagint Translators (Peeters, 2007): translation technique analysis, literalness scoring, and paratactic vs. hypotactic syntax tendencies.
• Jan de Waard & Eugene Nida — A Translator's Handbook on the Bible: functional equivalence principles used to classify paraphrase and dynamic rendering.
• Biblia Hebraica Stuttgartensia (BHS) and Biblia Hebraica Quinta (BHQ): the critical apparatus entries that classify variants as Vorlage differences, inner-Greek corruption, or translation choices.
• Göttingen Septuagint: critical edition used for LXX textual variants and Greek manuscript attestation.
• Dead Sea Scrolls (4QSam^a, 1QIsa^a, and others): pre-MT Hebrew witnesses that occasionally support LXX pluses and minuses against the MT.

All analysis is produced by Claude from its training knowledge. BibCrit does not call external databases at runtime. Results should be cross-checked against primary sources for peer-reviewed publication.

Open Data API

All cached analyses are freely accessible via a public REST API under the Apache 2.0 license — query, download, and reuse them in your own research or tools.

Example queries:

# Analyses from each tool's current prompt and model (paginated)
GET /api/cache

# Also include results from earlier prompts and models
GET /api/cache?include_retired=true

# Filter by tool
GET /api/cache?tool=divergence

# Filter by tool + passage
GET /api/cache?tool=theological&ref=Isaiah+7:14

# Only discovery-ready results
GET /api/cache?discovery_ready=true&limit=50&offset=0

Each record includes the full structured JSON as generated by Claude, plus metadata: tool, reference, model_version, cached_at. Supported query parameters: tool, ref (substring match), discovery_ready, limit (max 200), offset.

Full interactive documentation for all 14 analysis endpoints, the corpus browser, export formats, and vote routes is at /api/docs (Swagger UI). The machine-readable OpenAPI 3.0 spec is at /api/v1/openapi.json.

A live usage dashboard shows which passages have been analysed, which tools were used, and activity over time — updated in real time from the cache.

If you use BibCrit data in research, please cite:
Fresco Benaim, J. (2026). BibCrit: AI-assisted biblical textual criticism. doi:10.5281/zenodo.19358424

Data Sources

BibCrit's corpus layer is built on eight open academic datasets supplying word-level text across eight manuscript traditions:

• ETCBC — Eep Talstra Centre for Bible and Computer (Vrije Universiteit Amsterdam)
  Provides the Masoretic Text with full morphological tagging (lemma, part of speech, person/gender/number, clause boundaries) via the BHSA dataset. Also the source for the Dead Sea Scrolls corpus (1QIsaᵃ, 1QIsaᵇ, 4QSamᵃ, 11QPaleoLev, 4QDeutⁿ) and the Peshitta Syriac OT (all 39 OT books, 308,863 word tokens). Peshitta lexical data derives from the SEDRA database (Beth Mardutho: The Syriac Institute). All three ETCBC modules are accessed via Text-Fabric (Hagen). Licensed CC-BY-NC 4.0.
  ETCBC/BHSA on GitHub →
• LXX-Rahlfs-1935 — Eliran Wong
  Provides the Septuagint text: Rahlfs's 1935 critical edition in digital form, with lemmas, morphological tags and Strong's numbers, for the 39 books of the Hebrew canon (479,873 words). This is an edition, not a manuscript: BibCrit holds no text of Vaticanus, Sinaiticus or Alexandrinus. Joshua is the B text, Rahlfs's main text; the A text he prints beside it in 15:21–62, 18:21–28 and 19:1–48 is not loaded. Judges is the A text; Rahlfs prints both texts in full, and the B text is not loaded. Daniel is Theodotion only (the Old Greek is not loaded), and Susanna and Bel are excluded. The deuterocanonical books and 1 Esdras are not loaded. The additions Rahlfs prints inside canonical books are loaded: the Prayer of Azariah and the Song of the Three Young Men in Daniel 3, the additions to Esther as lettered verses (Esther 10:3a), and Psalm 151. The source is licensed CC BY-NC-SA 4.0. BibCrit uses it as its LXX corpus for all Greek-side analysis.
  eliranwong/LXX-Rahlfs-1935 on GitHub →
• SBLGNT — Society of Biblical Literature
  The SBL Greek New Testament (Michael W. Holmes, ed., 2010), taken from MorphGNT (James Tauber et al.), which adds the morphological tagging, provides the Greek text for all 27 NT books (137,554 word tokens), without punctuation. Used by the NT Textual Tradition Analyzer and the NT Use of OT Analyzer.
  MorphGNT SBLGNT on GitHub →
• Sefaria — Targum Onkelos & Targum Jonathan
  Sefaria's openly licensed export provides the Aramaic Targum text for Torah (Onkelos, 5 books) and Prophets (Jonathan, 21 books) — 26 books and 240,254 Aramaic word tokens in total. Note: Targum uses the Hebrew square script, not Syriac. Attributed per Sefaria's CC BY-SA license.
  Sefaria-Export on GitHub →
• scrollmapper/bible_databases — Clementine Vulgate
  Provides Jerome's Latin Vulgate (Clementine edition, 1592) across the full Protestant canon — 66 books, 541,537 Latin word tokens. Used by the DSS Bridge, Manuscript Transmission, and Theological Revision Detector to add a 4th-century Latin witness to each analysis.
  scrollmapper/bible_databases on GitHub →
• Samaritan Pentateuch — dt-ucph/sp (University of Copenhagen)
  The Samaritan Pentateuch Text-Fabric module provides the SP text for the five Torah books, used primarily by the Numerical Discrepancy Modeler to compare patriarchal chronologies across MT, LXX, and SP traditions. Accessed via Text-Fabric (Hagen).
  dt-ucph/sp on GitHub →
• Patristic sources
  The Patristic Citation Tracker uses Claude's knowledge of the standard critical editions (Migne PG/PL, Sources Chrétiennes, ANF/NPNF) to surface how Church Fathers through the 5th century cited biblical passages and what text form their quotations reflect.

All corpus data is used in accordance with the respective open licenses. See the BibCrit repository for full attribution.

Frequently Asked Questions

• Methodology & Scholarly Use
• Can BibCrit analyses be cited in peer-reviewed publications or dissertations? ▾

  With appropriate qualification, yes. BibCrit is a research aid — not a primary source or a peer-reviewed commentary. Its analyses reflect Claude's training on published scholarship (Tov, Ulrich, Kraft, Nickelsburg, VanderKam, Metzger, and others) but have not been independently verified by human specialists.

  Recommended citation practice: use BibCrit to identify hypotheses, map the terrain, and generate leads; then verify significant findings against the primary critical editions (BHS/BHQ, Göttingen LXX, DJD series) and the specialist literature BibCrit cites. When you do cite BibCrit directly, use the DOI: 10.5281/zenodo.19358424.

• How should confidence scores be interpreted? ▾

  As the model's own estimate, not as a probability. It does not come from Bayesian inference, a corpus frequency or a survey of scholars, and it is not calibrated against anything. A high figure is not evidence that a claim is right. Always check the reasoning and the primary sources.

• How does BibCrit distinguish a different Vorlage from inner-Greek or scribal variation? ▾

  The Divergence Analyzer applies Tov's four-category typology: (1) different Vorlage — the LXX translator had a Hebrew text that already differed from proto-MT; (2) inner-Greek transmission error — the difference arose in the Greek manuscript chain after translation; (3) translator technique — a deliberate rendering decision by the translator (e.g. theological tendency, lexical substitution); (4) scribal error in either tradition. The Back-Translation Workbench separately attempts to reconstruct the Hebrew that would have produced a given Greek reading, then evaluates whether that Hebrew is attested in DSS or SP.

  The model's typology classification is always provisional. For high-stakes divergences (e.g. Isaiah 7:14, Deuteronomy 32:8, Psalm 22:17), the scholarly debate is centuries old and BibCrit surfaces the leading positions without adjudicating them definitively.

• How are Hays' criteria applied in the Second Temple Literature Bridge? ▾

  Each allusion in the STL Bridge is evaluated against Hays' six criteria from Echoes of Scripture in the Letters of Paul (1989): availability (was the source text accessible to the author?), volume (how extensive is the verbal/thematic overlap?), recurrence (does the author echo this source elsewhere?), thematic coherence (does the echo fit the surrounding argument?), historical plausibility (is this connection plausible within the author's social/historical context?), and satisfaction (does the allusion enrich the reading?). The confidence score aggregates across all six criteria, weighted by the primary methodology of Nickelsburg, VanderKam, Collins, or Knibb depending on the work being assessed.

• Are analyses deterministic? Will the same passage always return the same result? ▾

  Once cached, yes — every subsequent request for the same passage and tool returns the identical JSON from the cache (Supabase + local disk). Cache keys are SHA256(reference|tool|prompt_version|model_version), so results are stable until a prompt or model version changes.

  The first generation of any passage is stochastic (LLM sampling). If you need reproducibility for a specific result, note the model_version and prompt_version fields returned in the response. If a prompt version is incremented (e.g. divergence_v3), all existing cache entries for that tool are invalidated and regenerated on next request.

• How does BibCrit handle passages where the scholarly consensus is genuinely divided? ▾

  Contested passages receive multi-hypothesis treatment. Rather than committing to a single classification, the model returns competing hypotheses each with a confidence weight and a scholarly note identifying the principal proponents. Confidence scores on contested readings are typically in the 0.45–0.65 range to signal ambiguity — that range is a feature, not a limitation.

  For the most debated passages (Deuteronomy 32:8 MT vs. LXX/DSS; Psalm 22:17 kāʿărî/kārû; Isaiah 53:11), BibCrit surfaces the Barthélemy, Tov, and Ulrich positions rather than arbitrating. The assessment section explicitly names where consensus is weak.

• Corpus & Data
• Which critical editions back the MT and LXX corpora? ▾

  The MT corpus uses the ETCBC BHSA dataset (Biblia Hebraica Stuttgartensia with morphological encoding), the standard computational edition of the Leningrad Codex, distributed via the Text-Fabric framework from the Eep Talstra Centre for Bible and Computer (Vrije Universiteit Amsterdam) under CC BY-NC 4.0. The LXX corpus uses Rahlfs's 1935 edition as digitised in eliranwong/LXX-Rahlfs-1935, with morphological tags, covering 39 books. BibCrit does not change the wording of either corpus, though the LXX ingest renumbers Psalms to match the MT. The word objects are passed to Claude with reference metadata attached.

• Which DSS manuscripts are in the corpus? How are lacunae handled? ▾

  Five scrolls are loaded: 1QIsaᵃ (Isaiah 1–66, ~95% preserved), 1QIsaᵇ (Isaiah 7–66, fragmentary), 4QSamᵃ (Samuel fragments), 11QPaleoLev (Leviticus fragments in paleo-Hebrew script), and 4QDeutⁿ (Deuteronomy fragments). All are drawn from the ETCBC DSS Text-Fabric module.

• Does the STL Bridge use an actual corpus of 1 Enoch, Jubilees, etc.? ▾

  No — the STL Bridge is a pure AI tool. There is no CLTK or Open Scriptures corpus loaded for Second Temple literature. The model draws on its training knowledge of the Nickelsburg Hermeneia commentary on 1 Enoch, VanderKam's Jubilees edition, Collins' Apocalyptic Imagination, Knibb's Ethiopic Enoch, and the standard critical editions of Sirach, 4 Ezra, and Tobit. This means allusion identification for very obscure passages is less reliable than for well-studied ones; always verify specific passage references against the primary editions (Knibb 1978, VanderKam 2018 CEJL, Skehan–Di Lella AB Sirach).

• Tools & Workflow
• What is the difference between the Divergence Analyzer and the Ancient Witness Bridge? ▾

  The Divergence Analyzer (/divergence) focuses exclusively on MT vs. LXX, producing a word-level apparatus with divergence classification, back-translation hypothesis, and competing scholarly explanations. It is the primary tool for studying the LXX translation and its Vorlage.

  The Ancient Witness Bridge (/dss) is a five-tradition comparison tool: MT, LXX, Dead Sea Scrolls (corpus-backed), Samaritan Pentateuch (corpus-backed), and Peshitta (corpus-backed via ETCBC). It is designed for passages where DSS or SP alignment patterns are the central question — e.g. whether a Qumran scroll supports the LXX against MT, or stands independently.

• How do I export results for use in Zotero, Endnote, or LaTeX? ▾

  Every analysis result includes SBL Footnote and BibTeX export buttons in the toolbar below the results. The Divergence Analyzer additionally offers RIS (Zotero, Endnote, Mendeley) and TEI XML export. The Download button produces a plain-text file with the full structured analysis. All exports are also accessible programmatically via the REST API — e.g. GET /api/divergence/export/bibtex?ref=Isaiah+7:14.

• Can I access the raw JSON output or integrate BibCrit into a research pipeline? ▾

  Yes. The analysis cache is publicly readable, with earlier versions available through include_retired=true, via GET /api/cache?tool=divergence&ref=Isaiah+7:14. Each record returns the complete structured JSON with all fields, plus model_version, prompt_version, and cached_at metadata. The SSE stream endpoints (/api/{tool}/stream) can be consumed programmatically — they emit section events as each JSON key completes, then a final done event with the full payload. See docs/api-reference.md for full schema documentation.

• How does the Scholar Rating system work? ▾

  The ▲/▼ voting buttons below each result submit anonymous upvotes and downvotes to the BibCrit vote API (POST /api/vote). Votes are stored per tool + reference and inform the Discovery feed's quality ranking — highly-rated analyses surface more prominently. Your vote is stored in localStorage so it persists across sessions on the same device. The system is designed to let the scholarly community surface the most methodologically sound results, not to crowdsource truth — a 0.9 confidence score from Claude with 50 upvotes still requires you to check the primary sources.