Methods
What each BibCrit tool's results rest on: what code computes from the corpus, what text the model is given, and what comes from the model alone.
Three kinds of evidence
- Corpus-computed
- Code reads or counts it from the corpus. The model plays no part, and the result is repeatable.
- Hybrid
- The model works from corpus text it is given, and code may check or compute part of the result. Interpretation still comes from the model.
- Model knowledge
- The substance comes from what the model recalls from training. BibCrit holds no source to check it against. Treat every specific claim as unverified.
No tool below is corpus-computed from end to end. The corpus-computed parts are in the second column. The label on each tool describes its overall result.
Tool by tool
| Tool | Computed or checked by code | Corpus text given to the model | From the model | Reviewed | Prompt and model |
|---|---|---|---|---|---|
|
MT/LXX Divergence Analyzer Hybrid |
Nothing. The output is not checked. | MT, LXX, and any scroll text in the corpus. | Classification of each divergence, hypotheses, citations and confidence. | Revised | divergence v3claude-opus-4-7 |
|
Back-Translation Workbench Hybrid |
Compares each suggested Hebrew word with the MT and the scrolls in the corpus, and corrects false claims of agreement. Shows the status the model first gave. | MT, LXX, and any scroll text in the corpus. | The suggested Hebrew, its basis and reasoning. Scrolls outside the corpus, marked as such. | Revised | backtranslation v2claude-opus-4-7 |
|
Ancient Witness Bridge Hybrid |
Works out each witness's label and variant counts from the model's per-variant data, and marks which scrolls are in the corpus. Does not check that data against the texts. | MT, LXX, Dead Sea Scrolls, Samaritan Pentateuch, Peshitta and Vulgate, where the corpus has them. | Each variant, its type, whether it is substantive and which traditions share it. Two self-reported confidences, the argument and the assessment. | Revised | dss v14claude-opus-4-7 |
|
Scribal Tendency Profiler Hybrid |
Per book: verse-length correlation, Greek words per Hebrew segment and share of καί. Rough measures of unaligned text. | Sample MT and LXX passages, and the computed measures. | Evidence labels (clear, debated, none found), notes on examples, and named scholars. | Revised | scribal v2claude-opus-4-7 |
|
Numerical Discrepancy Modeler Hybrid |
Every number: reads the figures from the MT, LXX and SP, then computes the differences and repeated offsets. One text per tradition; numbers under 10 are left out. | The computed table, not the passage. | Competing explanations, unranked. | Revised | numerical v4claude-opus-4-7 |
|
Theological Revision Detector Model knowledge |
Nothing. The output is not checked. | For a passage: MT, LXX, Targum, Peshitta, Samaritan Pentateuch and Vulgate, where the corpus has them. For a whole book: none. | Whether motive explains a difference better than technique, source or copying, and citations and confidence. For a whole book, every reading is from memory. | Revised | theological v3claude-opus-4-7 |
|
Patristic Citation Tracker Model knowledge |
Counts the citations listed and their text forms. Does not verify them. | MT and LXX, or the SBLGNT for New Testament passages, as reference. | Every Father, work, location, paraphrase and text form. BibCrit has no patristic corpus. | Revised | patristic v4claude-opus-4-7 |
|
Manuscript Transmission Hybrid |
Draws a hand-curated diagram with a published source for each edge, lists the scrolls in the corpus, and rejects invalid edges the model proposes. The model cannot change the diagram. | The diagram, the scroll list and the first verse of the Vulgate. | Commentary, key divergences and any proposed edges. | Revised | genealogy v4claude-opus-4-7 |
|
Chiasm Detector Hybrid |
Counts the content lemmas each paired unit shares and ranks the pairing against every other pairing of the same units. Tests repeated words only, not themes. | MT. | Whether there is a structure, the units, the centre, the evidence and named scholars. | Revised | chiasm v2claude-opus-4-7 |
|
Pentateuchal Source and Redaction Analysis Hybrid |
Nothing. The output is not checked. | MT. | Observations, and separate readings under the documentary, supplementary and compositional frameworks. | Revised | source v2claude-opus-4-7 |
|
Targum Comparator Hybrid |
Refuses a passage with no Targum text in the corpus, without calling the model. Otherwise checks nothing. | MT, LXX and the Targum (Torah and Prophets). | Comparison, categories of change and confidence. | Revised | targum v2claude-opus-4-7 |
|
NT Textual Tradition Analyzer Model knowledge |
Nothing. The output is not checked. | SBLGNT only. No apparatus, ECM data or manuscript transcriptions. | All manuscript evidence and the judgement of how settled the text is. | Revised | nt_text v3claude-opus-4-7 |
|
NT Use of OT Analyzer Hybrid |
Looks up the MT and Rahlfs text of each passage the model names and shows that text. Does not check the alignment. | The New Testament passage, then the MT and LXX of the passages named in a first call. | Which Old Testament texts are used, alignment and confidence. Greek revisions are from memory. | Revised | nt_ot v2claude-opus-4-7 |
|
Second Temple Literature Bridge Model knowledge |
Counts the allusions listed per work. Does not verify them. | MT and LXX. No Second Temple texts. | Every allusion and everything said about the Second Temple works, with confidence. | Revised | stl v2claude-opus-4-7 |
|
LXX Manuscript Witnesses Model knowledge |
Works out which codices survive at the passage, overrides the model on this, and flags readings given for a lost codex. | Rahlfs' edition only. No text of Vaticanus, Sinaiticus or Alexandrinus. | Every reading of the three codices, from memory, and confidence. | Revised | lxx_ms_variants v2claude-opus-4-7 |
“Reviewed” refers to a review of each prompt against these evidence rules, prompted by peer-review criticism of the tool. “Revised” does not mean the output is correct, or that a specialist has validated it.
How to read a result
- Confidence figures are the model's own estimates, not calibrated probabilities. They are not evidence.
- Each result shows the model, prompt version and date that produced it. Results are stored, so the same question returns the same stored answer until the prompt changes.
- Scholars, manuscripts and page references named by the model are not checked against the works. Verify them before citing.
- Results are generated automatically. No scholar reviews them before they appear.
Corpus texts
| MT | ETCBC / BHSA |
| LXX | Rahlfs 1935 (eliranwong) |
| DSS | ETCBC / DSS |
| SP | dt-ucph / sp |
| GNT | SBLGNT via MorphGNT |
| PESH | ETCBC / peshitta |
| TARG | Sefaria (Onkelos, Jonathan) |
| VUL | Clementine Vulgate (scrollmapper) |
Known limits
- Each tradition is one text, so variants within a tradition are not in the corpus. Rahlfs' edition, for example, is not the Greek manuscripts.
- Rahlfs' lettered verses, such as 1 Kings 12:24a-z and the additions to Esther, are separate verses. Ask for them by label (1 Kings 12:24a). They have no Hebrew verse to pair with, and Esther 1:1 is found at Esther 1:1s. Daniel is Theodotion's text. Susanna and Bel are not in the corpus.
- The Vulgate here is the Clementine edition of 1592, not a 4th-century text. Its Psalms are renumbered to the Hebrew; other books keep its own numbering, which differs in places (e.g. Joel, Malachi, Esther, Daniel).
- The corpus holds only some Dead Sea Scrolls. A scroll missing from the corpus does not mean no scroll has the reading. Correction: until November 2026 the text labelled 1QIsaᵃ here was 1QIsaᵇ. Results from before then that cite 1QIsaᵃ should not be relied on.