An AI translation of the Bible from the Hebrew and Greek

The Anselm Project Bible was published complete, Genesis through Revelation, translated from per-word Hebrew and Greek data rather than from any English text.

The whole text was rebuilt by a translator that never sees a reference, a book name, or an existing English version, so that no familiar phrasing could be copied rather than translated.

The text was rebuilt again under a full committee: blind translation, independent fidelity audits against the source tokens, and adjudication of every disputed rendering, with the record of each decision kept per verse.

Standardized the capitalization of divine titles across all 31,166 verses: the Holy Spirit, the Father, the Son, the Lord, and related titles are now cased consistently wherever the referent is God, with the referent of every ambiguous title adjudicated from the source tokens.

Spelling was normalized to American English, and the Paraclete of John 14 to 16 is rendered "the Advocate" to match the same Greek word at 1 John 2:1. The same revision repaired quotation marks throughout: speeches that closed too early or never closed, stray marks, and short dialogue that had none, with each speech's boundaries adjudicated chapter by chapter.

Every changed verse carries its revision note in the reader, naming the editor (Claude Fable 5.1) and the adjudicator (Grok 4.3), and no verse was reworded.

Repaired the errors found by a whole-corpus audit of the text against its source tokens, with Claude Fable 5.1 (Anthropic) as editor and no model adjudication. The lost numerals of 1 Samuel 13:1 are now marked rather than read literally as "a year old."

Wrong numbers were corrected against the Hebrew in Genesis 5:7, 1 Kings 4:32, 6:1 and 7:3, Ezra 2:28 and 2:69, Nehemiah 7:70, Ezekiel 48, and Song of Songs 8:12; a dropped place name (Song 8:11) and a dropped negation (1 Chronicles 12:33) were restored; a reading imported from the Psalm 18 parallel was removed from 2 Samuel 22:29; four gloss slips and four name spellings were corrected; and John's double amen now reads "Truly, truly" throughout. Seven verses where the Hebrew numeral is intact but contradicted by a parallel passage (2 Chronicles 22:2 and 36:9 among them) keep their text and carry a textual note instead.

Every changed verse carries its revision note in the reader, with the source tokens it rests on, and the audit scripts now run as part of every release.

The earlier text translated one printed base per testament and had no way to weigh the manuscripts where they disagree. The apparatus built for this revision draws only on open sources: the United Bible Societies' Hebrew Old Testament Text Project (4,515 Hebrew problems with their alternatives), the ketiv and qere of the Open Scriptures Hebrew Bible, the STEPBible TAHOT and TAGNT tagged texts with their per-word record of eight printed Greek editions, the SBLGNT apparatus, the Center for New Testament Restoration's transcriptions of the Greek manuscripts through the fourth century, the Robinson-Pierpont Byzantine text, and Swete's edition of the Septuagint as a versional witness.

From these, 14,877 variation units were assembled and decided blind by three readers from three companies, Grok 4.3 (xAI), GPT-6 Astra (OpenAI), and Claude Fable 5.1 (Anthropic), each seeing the readings, their morphology, and the age and class of the witnesses, but never a reference, a manuscript name, or another committee's rating; the human committees and printed editions were compared only after the vote. A reading was adopted only where two of the three agreed.

On the page, 55 verses now depart from the base text and carry a note giving the reading they left, and 45 verses of disputed passages (Mark 16:9 to 20, John 7:53 to 8:11, Luke 22:43 and 44, Acts 8:37, and Romans 16:24 among them) are set in brackets with their wording kept. The same three readers decided the casing of every ambiguous title (Spirit, Father, Son, Lord, King, Rock, Shepherd) per occurrence from the source tokens, recasing 115 verses.

The 174 pericopes those decisions touch were retranslated over the decided text by the blind committee, audited for fidelity by GPT-6 Astra as a second model family with 93 minimal corrections adopted after a clean re-audit, read by eight blind English reviewers, and re-recorded for the audio Bible. Every changed verse carries its revision note in the reader, the full record of every departure with each reader's reasons is published with the release, and the apparatus tables are released CC BY-SA 4.0 in keeping with their sources.

A repair on September 17 and 18 corrected what the readers had been shown. The Hebrew problems file names no manuscript for most of its alternatives, yet the apparatus had labelled every one of them a medieval Hebrew manuscript; those labels are gone, the Dead Sea Scrolls and Septuagint witnesses cited on the Crux Registry pages now reach the readers, and the Greek rendering of the divine name is counted as support again. Every unit whose evidence changed (1,363 of them) was decided again by the same three readers, blind, and every place where an early witness stands against the kept text now carries a note giving the reader's reasons. The same repair found that the decided text had dropped the adopted words from 35 verses before translation; those words were restored, and the 34 pericopes holding them were retranslated, audited, reviewed by eight blind English readers, and re-recorded. A further check the same day compared every departing verse against the reading the committee had adopted, read blind by two families of models with a third to break disputes, and found five verses whose English still followed the rejected wording; those five were corrected by the same route.

A review at the end of September found that long stretches of the earlier text matched a modern published version word for word. The translating model had recited wording it remembered instead of translating, and the fidelity audit could not see it, because remembered wording from a faithful translation traces back to the source words as well as honest work does. So every one of the 2,058 passages was translated again, from the same tagged Hebrew, Aramaic, and Greek word data and the critical text decided for 3.5.

The committee for this revision was one model, GPT-6.1 Sol (OpenAI), run blind through the Codex command line with no tools and no memory between calls. Each passage went through nine steps in order: an analysis of how the words combine, one draft written from every attested sense of each word rather than a single gloss, a segment review, a fidelity audit, a terminology pass, a style pass, a proofread, a formal-English pass, and a blind meaning review that sees only the words and the English. A recall gate after every step fails any run of 30 or more words shared with the earlier text, and a meaning gate sends a flagged clause back to the words up to three times. No other Bible entered the pipeline at any point, not as input, not as an example, and not as a check.

After translation, each book was made consistent with itself (718 verses in the Old Testament, 200 in the New), 136 Old Testament and 127 New Testament verses were settled to one rendering of a key term, and 317 Old Testament and 85 New Testament verses were made to word the same source the same way where books run parallel. A blind English reader with no view of the source then read every line; each fix it proposed was retranslated from the words and had to pass the audit, the meaning review, and the recall gate again. A finishing pass set the quotation marks, kept the 45 disputed verses in brackets, and applied the casing adjudicated in 3.1 wherever the committee had not decided the referent itself.

Only afterwards was the finished text compared against four published versions, from local files the pipeline never saw. The check measures each passage's longest run of shared words and its overall similarity against the ceiling that a literal rendering of plain narrative reaches on its own. Of the 2,058 passages, 1,680 are independent, 350 are close, and 28 are recalled. A close passage shares a little more wording with one version than the four versions share with each other, most often a run about four words longer. The longest of those runs were read, and each renders plain Hebrew or Greek word for word. Each of the 28 recalled passages was read and ruled on, and all are lists of names, census rolls, itineraries, or short literal sentences with one possible wording. Isaiah 52:13 to 53:12, the passage that exposed the problem, now has no verse identical to any of the four.

The text on the page is the committee's. No person edited a verse. Where a gate stayed flagged after its rounds, a chair model resolved the passage from the committee's last draft, and that ruling is logged against the verse. The 3.1 and 3.2 conventions, American spelling, adjudicated casing, and the gap marked in 1 Samuel 13:1, were applied again by rule, and every change made by rule is on the ledger with the verse.

Before release, every one of the 14,877 places where the manuscripts disagree was decided again. The 3.5 decisions rested on an older first reader, Grok 4.3, and at 1,621 places on readings shown malformed, with a Greek edition's added words standing alone in place of the full wording. With the readings corrected, GPT-6.1 Sol read all 14,877 blind, and Claude Fable 5.1 read 2,483 of them as a second model family: every place where Sol left the base text, every question about a whole passage, and a sample of the rest. A departure needed both readers and a first reader at least 0.8 confident. No third reader was used, so the base text stands at the 293 places where the two disagreed, where the confidence fell short, or where the same words were put twice and the readers departed only once. The readers adopted 124 departures, against 90 in 3.5: 80 are new, and 46 of the earlier ones went back to the base text. Romans 16:25 to 27 came out of brackets, which leaves 42 bracketed verses. The English changed with the text in 45 verses, among them Deuteronomy 32:8, 1 Samuel 17:4, Psalm 22:16, Psalm 145:13, Isaiah 53:11, Mark 6:22, and Luke 24:36. Each of those verses was retranslated from the words, passed the audit, the meaning review, and the recall gate, and carries its textual note in the reader.