Anselm Project Bible 3.5

Version 3.5 of the Anselm Project Bible is out and in the reader. Through version 3.2 the APB translated the Hebrew and Greek exactly as two printed editions gave them, and wherever ancient manuscripts disagree about a word or a verse, it took whatever reading those editions had chosen. In 3.5, every place where the manuscripts disagree was collected, and three AI models, Grok 4.3, GPT-6 Astra, and Fable 5.1, decided each one without being told what passage they were reading. To explain what that means, this post starts with how the NIV, NASB, NLT, and ESV committees settle the same questions.

How the NIV, NASB, NLT, and ESV get their text

No modern translation committee translates from a manuscript. Each one translates from a critical edition: a printed Hebrew or Greek text that an editorial committee has already assembled from the manuscripts, with footnotes recording the readings it set aside. For the Old Testament the major translations use the Biblia Hebraica (BHS, and now BHQ), which prints the Leningrad Codex and lists variants underneath. For the New Testament it is the Nestle-Aland Novum Testamentum Graece and the United Bible Societies' Greek New Testament, which share a text and are edited by the same committee.

The translation committees then depart from those editions at a small number of places. The ESV preface says the translators followed the Masoretic Text and, in exceptional and difficult cases, the Dead Sea Scrolls, the Septuagint, the Samaritan Pentateuch, the Syriac Peshitta, and the Latin Vulgate. The NIV's Committee on Bible Translation says the same in slightly different words, and its footnotes ("Some manuscripts do not have...") are where the departures show. The NASB and the NLT work the same way. In every case, two committees stand between the manuscripts and the English: the editorial committee that built the critical edition, and the translation committee that decided where to follow it.

What the sources are

The Masoretic Text is the Hebrew Bible as copied by the Masoretes, the Jewish scribal families who fixed the consonants, added the vowel points, and counted the letters between the sixth and tenth centuries. The Leningrad Codex, copied in 1008, is the oldest complete manuscript of it and the base of every printed Hebrew Bible since 1937. The Masoretes also recorded places where the written word (the ketiv) was to be read aloud as a different word (the qere); there are about 1,300 of these, and each is a small textual decision.

The Dead Sea Scrolls, found at Qumran from 1947 on, are Hebrew copies from the third century BC to the first century AD, a thousand years older than the Leningrad Codex. Most agree with the Masoretic Text closely. Some, especially in Samuel and Jeremiah, preserve a different text, and a few readings that modern translations adopt (Deuteronomy 32:8, 1 Samuel 10:27) come from them.

The Septuagint is the Greek translation of the Hebrew Bible made in Alexandria in the third and second centuries BC. It is the Bible the New Testament writers usually quote. Where it differs from the Masoretic Text, the question is whether the translator was working from a different Hebrew text or translating freely, and a committee has to decide that case by case.

The Samaritan Pentateuch is the copy of the five books of Moses kept by the Samaritan community, which separated from the Jewish community before the time of Christ and preserved the Torah in a different script. I've had the chance to go to Samarita in Israel about ten years ago, and that part sticks in my head distinctly. It differs from the Masoretic Text in about six thousand places. Most are spelling and grammar; some expand the text from parallel passages; a handful are sectarian, such as the commandment to build the altar on Mount Gerizim. Where the Samaritan text agrees with the Septuagint or the Scrolls against the Masoretic Text, it carries weight.

The Peshitta (Syriac), the Vulgate (Jerome's Latin, about 400 AD), and the Targums (Aramaic paraphrases) are translations, and they witness to the Hebrew text their translators had in front of them.

For the New Testament, the witnesses are Greek manuscripts: papyri from the second and third centuries, the great fourth-century codices Sinaiticus and Vaticanus, and then thousands of later copies, most of which carry the Byzantine text that became standard in the Greek church and stands behind the King James Version. Printed editions compress all of that into a text with an apparatus. Nestle-Aland and the UBS edition are the standard; the SBL Greek New Testament, the Tyndale House edition, Westcott and Hort, and the Robinson-Pierpont Byzantine text are the others a committee compares.

How a committee scores a reading

Two kinds of evidence go into every decision. External evidence is about the witnesses: how old they are, how many independent lines of copying they represent, and how widely spread they are. Internal evidence is about the scribes: copyists tend to harmonize a passage to its parallel, expand a short reading, smooth a hard grammar, and add a familiar phrase, so the reading that explains how the others arose is usually the earlier one. A committee weighs the two together, and where they point the same way the decision is easy.

The UBS committee published its confidence as a letter. An A reading is certain, a B reading is almost certain, a C reading means the committee had difficulty deciding, and a D reading means great difficulty. The United Bible Societies' Hebrew Old Testament Text Project, which worked through the Old Testament from 1969 to 1980, used the same letters and listed the factors it weighed at each of its 4,515 problems: assimilation to a parallel, a harder reading misunderstood, a theological correction, and so on. Those ratings and factors are what a translation committee consults when it decides whether to follow the edition or a footnote.

What the APB does with the same sources

Version 3.5 built its apparatus from open sources only: the HOTTP problems with their Hebrew alternatives, the ketiv and qere of the Open Scriptures Hebrew Bible, the STEPBible tagging of both testaments (which records for every Greek word which of eight printed editions carry it), the SBL Greek New Testament 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. The Samaritan Pentateuch dataset is licensed for non-commercial use only, and the APB is completely free to everyone. (Further, Samaritan readings enter reports only where an openly licensed source reports them, never directly.) The Scrolls have no open transcriptions, and their readings arrive through the HOTTP entries. Every source and license is on the About page.

The build produced 14,871 variation units: one record for each place where the witnesses differ, with every reading, its grammar, a gloss, and each witness described by class and date. The manuscript names, the edition names, and the HOTTP and UBS ratings were written to a separate file that no model reads.

Now, we have to discuss our new committee: three readers decided every unit. Grok 4.3 (xAI) read every unit. GPT-6 Astra (OpenAI) also read every unit independently. Claude Fable 5.1 (Anthropic) read the units where the first two disagreed or where either reported low confidence. A reading was adopted only where two of the three agreed, and the base text stood where they split three ways. Each reader saw the surrounding words, the readings, and the witnesses by class and age ("early Greek manuscript tradition, 2nd to 4th century, three witnesses" against "later majority tradition, 9th to 15th century"). None saw a reference, a book name, a manuscript name, or another committee's letter grade. A model that recognizes Mark 16 will decide Mark 16 from what it remembers of the commentaries, and keeping the readers blind is how the APB avoids that. After the vote, each decision was compared with the HOTTP ratings and the printed editions, and that comparison is kept in the release record.

The readers weighed the same two kinds of evidence a human committee weighs, and their reasons are on the page. At 1 Corinthians 13:3 the base text's "and if" (καὶ ἐὰν) was set aside for the contracted κἂν that the early witnesses share; the note gives both forms and says why. Fifty-four verses adopt a reading other than the printed base, each with a note giving the reading it left. No verse is deleted: where the committee judged a passage a later addition, it stays on the page in brackets with a note. Mark 16:9 to 20, John 7:53 to 8:11, Luke 22:43 and 44, Acts 8:37, and Romans 16:24 are bracketed, along with seventeen other verses.

Divine-title casing went through the same readers. Every occurrence of the words behind Spirit, Father, Son, Lord, King, Rock, and Shepherd was put to them as a question of who is meant, so a verse can carry the same word two ways where the Hebrew does. 1 Samuel 16:14 now reads "The Spirit of the LORD departed from Saul, and a harmful spirit from the LORD terrified him." One hundred fifteen verses were recased.

The 174 pericopes touched by a textual or casing decision were retranslated by the version 3 committee over the decided text. GPT-6 Astra audited each one against the source for additions, omissions, and imported ideas, and ninety-three took a minimal correction after a clean re-audit. Eight blind English readers then flagged 207 verses as bad English, and those verses reverted to the 3.2 wording; where a reverted verse also carried a decision, the note stays on the page and the verse is on a list for me to work through by hand. The audio for the 170 pericopes whose wording changed was re-recorded.

Open any verse in the reader, press Audit, then Repairs, to see the decision, the reading it left, and the readers' reasons. The Crux Registry still holds the sixty-five best-known disputes with every witness named. The text is on Hugging Face as version 3.5 under CC BY 4.0.