misterprmiller/anselm-project-bible

The Anselm Project Bible (APB) — v3

A complete AI translation of the Bible: all 66 books of the Protestant canon, translated from original-language data — not paraphrased from existing English versions.

• First complete version published: late October 2025 (announcement, Show HN, Nov 9 2025)
• v2.0 (blind translator): December 2, 2025
• v3 (audit-committee release): June 5, 2026 — release apb_v3_final, generated 2026-06-02
• Read online (free, no account): https://www.anselm-project.com/bible/
• Methodology: https://www.anselm-project.com/bible/about

As far as the author can determine, the APB is the first complete AI translation of the entire Protestant canon. If you know of an earlier complete one, write to info@anselm-project.com and this record will be amended.

Contents

File	Description
apb_v3.jsonl	One row per public verse: {book, chapter, verse, text, testament, translation, version}
apb_v3_audit_trail.jsonl	The complete committee audit trail: one row per pipeline stage per pericope — 22,965 raw stage records with model, prompt version, input hash, token counts, timestamp, and the stage's full result
audit_trail_summary.json	Aggregate trail counts: stages, audit verdicts, publish-gate results, token totals
dataset_metadata.json	Release provenance and counts
export_apb_dataset.py	The script that produced the verse JSONL from the site's authoritative release file
export_audit_trail.py	The script that produced the audit trail from the raw checkpoint directory

Scale: 66 books · 2,058 pericopes · 31,166 verses · 22,965 audit-trail stage records.

How the translation was made

Every stage runs on Grok (grok-4.3, xAI), but no single model pass produces the text. One model is run as eight separated committee roles — philology, apparatus, two-candidate translation, deterministic selection, adversarial back-translation audit, terminology, style, and proofreading — with deterministic (non-AI) gates between stages and a deterministic publish gate at the end.

Two properties define the approach:

1. Blind translation. The translating stages are never told the text is Scripture — they receive per-word tokens, decoded morphology, and lexicon glosses with no canonical framing. This prevents the model from reciting the English Bible it already knows instead of translating the source (the historical failure case: adding "the Son of God" to Mark 1:1 where the Greek source data lacks it — v3 renders it "The beginning of the gospel of Jesus Christ.").
2. Token-level traceability. An adversarial audit back-translates the English and reverses any word that has no source token. Design goal: every English word provably traceable to a source token. The per-verse translation record is inspectable in the web reader's Audit panel — and the complete raw stage records ship in this dataset (next section).

The audit trail (raw committee records)

apb_v3_audit_trail.jsonl is the complete, unedited run record of the v3 committee: one row per pipeline stage per pericope — 22,965 stage records covering all 2,058 pericopes across all 66 books, exactly as the pipeline checkpointed them at generation time (May–June 2026).

Field	Meaning
book, ref	Public book name and pericope reference (e.g. "1 John 1:1-4")
book_dir, pericope_key, pericope_id	Stable keys for joining rows
stage	Pipeline stage (see below)
model, prompt_version, schema_version	What produced the record
input_hash	SHA-256 (16 hex chars) over the stage's canonical input payload plus the prompt version — identical inputs always hash identically
input_tokens, output_tokens	Model tokens spent on the stage; 0/0 marks the deterministic (non-AI) stages
completed_at	UTC timestamp when the stage checkpointed
result_json	The stage's full raw result, serialized as a JSON string — parse with json.loads

Core stages in pipeline order: philology_literal, translate, source_locked_english, candidate_select, terminology_standardize, style, parallels_v1, audit, publish_gate, final — plus the repair machinery where it fired: audit_redraft, philology_contextual, reconcile, segment_review, repair, publish_gate_repair, polish_flagged (three tiers), translate_repair.

The record is published as written, failures included:

• Every pericope has exactly one final audit-stage record: audit (1,537) or audit_redraft (521 — the first audit forced a redraft and the audit re-ran). Audit rows retain the auditor's full back-translation text plus its additions / omissions / imported-concepts findings, clean verdicts included.
• Stage verdicts are left as the auditor wrote them: 358 pericopes' last audit-stage verdict is flagged. The downstream repair-stage rows record what was done about each, and the deterministic publish gate (publish-gate-2026-06-01a) records the final scan: 2,058 of 2,058 passed.
• 157.8M model tokens are accounted for stage-by-stage in the token columns.

import json
rows = [json.loads(l) for l in open("apb_v3_audit_trail.jsonl", encoding="utf-8")]
audits = [r for r in rows if r["stage"] in ("audit", "audit_redraft")]
mark = next(r for r in audits if r["ref"].startswith("Mark 1:"))
print(json.loads(mark["result_json"])["back_translation"])

Source texts

• Old Testament: Open Scriptures Hebrew Bible (Westminster Leningrad Codex; text public domain, lemma/morphology CC BY 4.0)
• New Testament: SBL Greek New Testament with MorphGNT morphological analysis (SBLGNT text under the SBLGNT license; MorphGNT parsing CC BY-SA)
• A small number of traditional verse slots absent from the local source snapshot (e.g., John 7:53–8:11) are supplied from the official Faithlife/SBLGNT text and explicitly marked as textual supplements in the web release.

License

The APB English text is released under CC BY 4.0 — free to use, share, adapt, and train on, with attribution to "The Anselm Project (anselm-project.com)". The SBLGNT/MorphGNT licenses above govern the source texts, not this English translation.

Citation

Miller, Paul. The Anselm Project Bible (v3), 2026. First complete version published October 2025. https://www.anselm-project.com/bible/about