People's Open Bible — Phase 8 (LXX Deuterocanon) Corpus Health

Generated: 2026-04-21 14:30 UTC

Corpus composition

• Total verses: 6337
• Scan-adjudicated verses: 2627 (41.5%)
• Unchanged verses (our OCR already agreed with First1KGreek at similarity ≥ 0.85): 3710 (58.5%)

Text source (every verse = our OCR of Swete scan, never First1KGreek text)

• AI-vision OCR (GPT-5.4 on scan images): 4996 verses
• Regex-parser OCR (our text files): 1078 verses
• Adjudicated scan re-read: 263 verses

The Greek text in every verse was produced by reading Swete's 1909 printed scan, either directly or with cross-validation against scholarly references. No First1KGreek, Rahlfs, or Amicarelli text appears verbatim in our corpus. Those served as reference oracles only.

Adjudicator verdict distribution (for verses that were re-examined)

• ours (our OCR confirmed by scan): 1281 (48.8%)
• first1k (scan matched First1KGreek better than ours): 499 (19.0%)
• neither (both were off; fresh scan-grounded reading): 410 (15.6%)
• rahlfs_match (3-way triangulation across different edition): 156
• swete_consensus (rescue pass — all 3 Swete transcriptions agree): 45
• amicarelli (rescue pass — Amicarelli's Swete transcription matched scan best): 169
• both_ok (minor orthographic differences): 67

Adjudicator confidence

• High (unambiguous scan reading): 2627 (100.0%)
• Medium (minor scan uncertainty): 0 (0.0%)
• Low (scan is damaged/illegible; best-guess reading): 0 (0.0%)

The 100.0% high-confidence rate means that Azure GPT-5.4 vision was able to read every adjudicated Swete verse in the current corpus unambiguously enough to assign high confidence.

Sources consulted (cross-validation, not text)

Additional Zone-2 cross-check witnesses are governed by ../../../REFERENCE_SOURCES.md; their content is not reproduced or vendored here and does not flow into corpus output.

Per-book breakdown

Tier meanings: A = cleanest (ship translation first), B = good, C = solid, D = some complexity, E = most challenging typography.

Remaining low-confidence verses

No low-confidence verses remain in the current final corpus.

Methodology (pipeline stages)

1. OCR: Swete 1909 Internet Archive scans → Azure GPT-5.4 vision → page-level .txt files with structural markers (RUNNING HEAD, BODY, APPARATUS).
2. Review: Azure + Gemini 2.5 Pro independent reviews → automated correction application.
3. AI-vision re-parse: For each chapter, Azure GPT-5.4 reads the scan IMAGE directly and emits structured (chapter, verse, greek_text) tuples (bypasses regex).
4. Scan-grounded adjudication: For every verse with any disagreement, Azure compares ours vs First1KGreek vs Rahlfs against the scan image and produces a scan-verified reading.
5. Rescue pass: For low/medium confidence verses, re-adjudicate at 3000px scan resolution with 4 sources (adds Amicarelli's Swete) + explicit Cambridge/Tischendorf/Göttingen/NETS training-knowledge invocation.
6. Final corpus: ours_only_corpus/*.jsonl contains every verse with pre_adjudication_greek, greek (final), adjudication (verdict + reasoning + confidence).

Ready-for-translation status

The corpus is translation-ready for all 13 books. No medium- or low-confidence verses remain in the current final corpus.