Anselm Project Bible

The APB is a fresh translation based on the following standard critical editions of the original language texts:

• Old Testament: The Unicode/XML Westminster Leningrad Codex (WLC), which represents the Masoretic Text (MT).
• New Testament: The Greek New Testament: SBL Edition (SBLGNT), a modern critical text.

The APB is a direct translation of the specified base texts (WLC and SBLGNT) and does not depart from them to adopt alternate readings from other manuscripts.

Other important textual traditions like the Septuagint (LXX) and the Dead Sea Scrolls (DSS) are not used for textual adoption. The translation reflects the WLC and SBLGNT as received.

• Ketiv/Qere: The translation follows the Masoretic scribal tradition of prioritizing the Qere (what is to be read) in the main text. For example, in Psalm 100:3, the APB reads "...it is he who made us, and we are his," following the Qere ("and His") rather than the Ketiv ("and not").
• Psalm Titles: Titles given in the Hebrew text (e.g., "A psalm of thanksgiving") are included as part of verse 1, following a common English convention.
• Psalm Numbering: The APB uses the psalm numbering from the Masoretic Text (WLC) exclusively and does not include alternative numbering from the Septuagint/Vulgate.
• OT in the NT: The APB faithfully translates its base texts as-is, without attempting to harmonize them. When the NT (SBLGNT) quotes the OT using a Septuagint reading, the APB translates that reading. When translating the original OT passage (WLC), it translates the Masoretic text. (e.g., Hebrews 10:5 reads "...a body you prepared for me," while Psalm 40:6 reads "...you opened my ears.").
• Deuterocanon: The Deuterocanonical books (Apocrypha) are not included. They may be considered for inclusion as a separate volume in a future update.

Version 2.0 implements a "blind translator" approach to prevent AI bias. The translation AI has no idea it's working with biblical text - at least from the prompting into the AI. All references to Scripture, the Bible, theology, and religion were removed from translation prompts.

The AI receives ancient language text and translates it to formal English without any awareness of the content's nature. This prevents the AI from defaulting to traditional renderings and forces strictly literal translation from the source text.

Result: A translation that is more literal and in many places clunkier than previous versions, but significantly more honest. The clunkiness is proof the system is working in theory - when the AI isn't cheating by recognizing biblical patterns, you get word-for-word accuracy even when it sounds awkward in English.

The APB v2.0 uses a four-step translation pipeline:

1. Generate Candidates: The system generates five translation candidates for each pericope from the source text, with the AI strictly forbidden from introducing concepts not lexically present.
2. Judge Selection & Name Extraction: A separate judge AI reviews all five candidates and selects the best translation based on literal accessibility, ideological fidelity, dignified English, and internal consistency. During this step, the judge also extracts all proper names from the selected translation.
3. Name Standardization: A specialized AI standardizes the extracted proper names to modern English Bible conventions (e.g., YHWH → "the LORD", Elohim → "God"). This ensures consistency across all 66 books.
4. Syntax Polishing: A polisher AI rearranges words for natural English syntax without changing word choices, substituting synonyms, or adding/removing content. Only minimal articles and prepositions required by English grammar may be adjusted.

Each step saves its output incrementally. If something crashes at step 3, you don't lose the work from steps 1 and 2. This creates a transparent audit trail where you can see the literal translation, the syntactic changes made, and the rationale for each step.

Hybrid AI Approach: Version 2.0 uses both xAI's Grok 4.1 and OpenAI for different tasks:

• Linguistic Work (Grok 4.1): Translation generation using fast-non-reasoning variant, and judge selection using fast-reasoning variant. Grok handles all the actual translation and evaluation.
• Structured Output (OpenAI): Handles JSON formatting and structured output validation. OpenAI doesn't crash nearly as much with structured outputs, providing reliability where it matters.

Grok is significantly cheaper than GPT-5, allowing full-power processing constantly for the linguistic work without balancing cost versus quality. The system returns data fast enough to complete the entire Bible translation in three days.

Output Validation: Uses Pydantic schemas for all structured outputs. The system defines exactly what structure it expects and enforces that structure at the generation level, eliminating JSON parsing errors.

The evidence-building system uses modular decision bots. Each bot has a specific job and runs independently:

• MorphologiesBot: Generates grammatical parsing for every Hebrew and Greek word (tense, voice, mood, case, number, gender). Runs first because everything else depends on it.
• ProperNamesBot: Builds a comprehensive database of every person, place, and divine name in Scripture.
• KeywordConsistencyBot: Tracks how key theological terms get translated across different contexts.
• AmbiguityBot: Flags passages where the source text is genuinely unclear so the system can document translation choices.

The architecture is modular - you can run individual bots on specific books or chapters. You can add new bots without touching the core pipeline. Each bot saves its results to a separate JSON file.

• Divine Name: The tetragrammaton (YHWH) and its variants are standardized to "the LORD" in small-caps during the name standardization step. The Hebrew term Adonai is rendered as "Lord" in title-case. Compound names like "YHWH God" become "the LORD God." This follows the convention used by ESV, NIV, NASB, and KJV.
• Name Consistency: The name standardization step ensures consistent rendering of proper names throughout the translation (people, places, divine titles) according to modern English Bible conventions.
• Idioms: The AI was instructed to preserve the original Hebrew and Greek idioms, translating them literally or with their most traditional English equivalent (e.g., "Gird up your loins" in Job 38:3; "my horn is exalted" in Psalm 92:10).
• Measures & Currency: Ancient units of measurement, weight, and currency are transliterated directly into the text (e.g., "cubits," "shekels," "denarius"). No in-text conversions or footnotes are provided.

Complete Transparency: Every verse in the APB reader includes access to:

• The source text (Hebrew or Greek)
• The literal translation with standardized names
• The polished translation for natural English readability
• Polishing reasoning explaining what syntactic changes were made and why
• All translation candidates generated by the AI with their rationales

Artifact Files: The JSON artifact file generated for each pericope serves as a complete record of the entire four-step translation pipeline, preserving all candidates, decisions, and reasoning for scholarly review.