---
license: cc-by-sa-4.0
language:
  - grc
task_categories:
  - feature-extraction
tags:
  - nuberea
  - biblical-studies
  - new-testament
  - textual-criticism
  - cbgm
size_categories:
  - 100K<n<1M
configs:
  - config_name: apparatus_full
    data_files:
      - split: train
        path: data/apparatus_full/train-*.parquet
  - config_name: reader_apparatus
    data_files:
      - split: train
        path: data/reader_apparatus/train-*.parquet
  - config_name: reader_apparatus_approximate
    data_files:
      - split: train
        path: data/reader_apparatus_approximate/train-*.parquet
  - config_name: witnesses_public
    data_files:
      - split: train
        path: data/witnesses_public/train-*.parquet
  - config_name: witness_catalog
    data_files:
      - split: train
        path: data/witness_catalog/train-*.parquet
  - config_name: witness_verse_texts
    data_files:
      - split: train
        path: data/witness_verse_texts/train-*.parquet
  - config_name: variation_units
    data_files:
      - split: train
        path: data/variation_units/train-*.parquet
  - config_name: major_edits_curated
    data_files:
      - split: train
        path: data/major_edits_curated/train-*.parquet
  - config_name: witness_verse_alignment
    data_files:
      - split: train
        path: data/witness_verse_alignment/train-*.parquet
  - config_name: witness_apparatus_links
    data_files:
      - split: train
        path: data/witness_apparatus_links/train-*.parquet
  - config_name: syntax_graph_index
    data_files:
      - split: train
        path: data/syntax_graph_index/train-*.parquet
  - config_name: lineage_nodes
    data_files:
      - split: train
        path: data/lineage_nodes/train-*.parquet
  - config_name: lineage_edges
    data_files:
      - split: train
        path: data/lineage_edges/train-*.parquet
  - config_name: apparatus_claim_delta
    data_files:
      - split: train
        path: data/apparatus_claim_delta/train-*.parquet
  - config_name: cbgm_variation_units
    data_files:
      - split: train
        path: data/cbgm_variation_units/train-*.parquet
  - config_name: cbgm_readings
    data_files:
      - split: train
        path: data/cbgm_readings/train-*.parquet
  - config_name: cbgm_reading_support
    data_files:
      - split: train
        path: data/cbgm_reading_support/train-*.parquet
  - config_name: cbgm_compare_witnesses
    data_files:
      - split: train
        path: data/cbgm_compare_witnesses/train-*.parquet
  - config_name: cbgm_reading_uncertainty
    data_files:
      - split: train
        path: data/cbgm_reading_uncertainty/train-*.parquet
status: active
---
# NuBerea NT Textual Analysis

New Testament textual-criticism data over the 188-witness CNTR transcription
base, SBLGNT apparatus, MACULA token alignment, and Open-CBGM v2.0 analysis.
This query-facing package exposes the formal variation-unit, reading-support,
and witness-comparison surfaces alongside the established apparatus and witness
tables.

## Interpretation boundary

The published `lineage_edges` rows are **undirected textual-affinity edges**.
They are useful for closest-relative and coherence questions, but they are not
formal ancestor/descendant relations. The recovered model contains no reading
relations, no potential ancestors, and no directed DOT edges. No directed
local-stemma or textual-flow claim should be inferred from these tables.

`cbgm_reading_uncertainty` is an AI-facing evidence-ranking layer, not the old
SBLGNT footnote-presentation policy. Its selection frequencies are deterministic
perturbation diagnostics, not calibrated probabilities or confidence intervals.
It uses 512 Bayesian-bootstrap draws over heuristic witness-affinity blocks.
Rows with fewer than three blocks are marked provisional; single-reading units
do not receive a stability estimate. The `evidence_requirements` payload states
the minimum and observed evidence for each datum on each row.

Method v3 makes the sensitivity dimensions explicit. Coherence weighting is
ablated against uniform witness weights; affinity blocking is ablated against
independent witness multipliers; and every observed witness is removed once.
Each factor distinguishes selection change, stability change, and score-margin
change. `plausible_readings_90` is the smallest descending selection-frequency
set whose cumulative frequency reaches 0.90. Chronological weighting is marked
unimplemented because all 188 `witnesses_public.date_range` values are null;
root proxies, graph position, manuscript class, and affinity are not substituted
for chronology or ancestry.

## Configs

| Config | Rows | Description |
|---|---:|---|
| `apparatus_full` | 6,966 | Full SBLGNT critical apparatus with variants, witness support, and coherence summaries. |
| `reader_apparatus` | 1,132 | Reader-facing substantive apparatus notes. |
| `reader_apparatus_approximate` | 87 | Reader apparatus notes linked by approximate alignment. |
| `witnesses_public` | 188 | Public witness metadata for the 188-witness corpus. |
| `witness_catalog` | 188 | Extended witness catalog and attestation statistics. |
| `witness_verse_texts` | 42,482 | Per-witness, per-verse CNTR transcription texts. |
| `variation_units` | 26,065 | Evidence-first variation units: a unit wherever >=2 of 196 aligned witnesses (188 CNTR + 8 editions NA28/NA27/THGNT/SBL/WH/Treg/TR/Byz) read differently after orthographic folding, anchored on SBLGNT token spans; readings carry manuscript and edition support, `support_tier` (edition/multi/singleton), SBL note links, CBGM overlay; 26 explicit coverage rows. |
| `major_edits_curated` | 28 | Hand-curated NT scholarly-seam registry (t0 curated-gold). Its textual entries are also served in `apparatus_full` as `curated_major_edit` notes (`note_id` suffix `:cme`); this config remains the curated source of record and the curated-locus guard for `variation_units`. |
| `witness_verse_alignment` | 42,482 | Per-witness, per-verse alignment against the SBLGNT base. |
| `witness_apparatus_links` | 7,511 | Witness readings linked to apparatus and formal variation units. |
| `syntax_graph_index` | 4,718 | Index of SBLGNT syntax graph artifacts. |
| `lineage_nodes` | 188 | Witness nodes with coverage and affinity-graph statistics. |
| `lineage_edges` | 977 | Undirected witness-affinity edges; not ancestor/descendant relations. |
| `apparatus_claim_delta` | 21,722 | Machine-derived claim-impact comparisons for NT apparatus notes. |
| `cbgm_variation_units` | 2,675 | Exact formal variation units with book, chapter, verse, and OSIS anchors. |
| `cbgm_readings` | 7,024 | Greek text and metadata for every TEI reading; 237 defective readings retain explicit engine-exclusion flags. |
| `cbgm_reading_support` | 16,151 | Witness support for every TEI reading; 247 defective support rows retain explicit engine-exclusion flags. |
| `cbgm_compare_witnesses` | 9,156 | Decoded per-book pairwise witness coherence and agreement statistics. |
| `cbgm_reading_uncertainty` | 2,675 | AI-facing coherence-weighted reading selection, perturbation stability, plausible alternatives, explicit coherence/affinity/witness ablations, and minimum-evidence diagnostics for every formal unit. |

## Formal CBGM recovery

The complete 27-book model was rebuilt with Open-CBGM standalone commit
`7e338f2984b77a1f8b49253e1109c289bfd1e0a2` and core commit
`2917ba4de5f62f97f59edd451b8c8a6dcb7f134d`. It reproduces every published
witness-pair/book value (`pass`, `eq`, `expl`, `uncl`, and `perc`) and every
positive node-extant count exactly. The `cbgm_recovery/` directory carries the
hash and row-count publication manifest.

Queryable formal surfaces use exact book/chapter/verse/OSIS anchors. Prefer an
exact `osis_ref` predicate for a verse; a trailing wildcard deliberately widens
the query to a reference prefix.

## License and attribution

CC-BY-SA-4.0. The strictest upstream constraint is the CNTR witness
transcription corpus. Other inputs include the SBLGNT apparatus (CC-BY-4.0),
MACULA SBLGNT (CC-BY-4.0), and Open-CBGM (MIT).