Careful, contextual analysis of a biblical text to determine its original meaning and implications.

What is Hermeneutics?

The theory and methodology of interpreting biblical texts, including principles that guide understanding across genres.

What is Mixture-of-Experts?

An AI architecture that delegates tasks to specialized models or agents, combining their outputs for a coordinated result.

What is Structured Output?

Formatted, machine-readable results (e.g., JSON) that enforce consistent content and prevent repetitive prose.

Back to Blog

Biblical Studies Theology Technical Updates

How Anselm's Topical Reports Produce Accurate Biblical Theology

Anselm's topical reports deliver accurate biblical theology: a mixture-of-experts that scans the canon, analyzes genres, and synthesizes nonrepetitive insights.

November 28, 2025

Paul Miller

6 min read

Whiteboard ANSELM TOPICAL REPORT showing three phases, expert icons, aggregated results bucket and JSON output

I rebuilt the topical engine from scratch a few weeks ago. The original version was repetitive garbage. This one works.

Here's the technical breakdown of how the Anselm Project generates topical reports.

The Problem

Topical studies are harder than passage studies. With passage reports, the boundaries are clear - Matthew informs Matthew. But ask the AI to study "grace" and you're asking it to synthesize 66 books across multiple genres spanning 1,500 years of writing.

The AI needs to search the entire canon, identify relevant material, analyze it by genre, and synthesize it coherently. Do that wrong and you get repetitive nonsense. Do it right and you get comprehensive biblical theology.

The Architecture: Three Phases

The system uses a mixture-of-experts architecture with massive parallel processing. Every topical report goes through three distinct phases.

Phase 1: Relevance Scanning

Before analyzing anything, the system needs to know which books actually matter for this topic.

I built seven genre experts, each responsible for a specific corpus:

Torah Scholar - Genesis through Deuteronomy
Historian - Joshua through Esther (12 books)
Sage - Job, Psalms, Proverbs, Ecclesiastes, Song of Solomon
Prophet - Isaiah through Malachi (16 books)
Evangelist - Matthew, Mark, Luke, John, Acts
Apostle - Romans through Jude (21 books)
Seer - Daniel and Revelation

Each expert receives a binary relevance scan prompt: "Which books in your domain contain substantial treatment of this topic?"

The AI returns a simple JSON object with true/false flags for each book. True means the topic is a key theme or explicitly discussed. False means it's absent or incidental.

This happens in parallel across all seven experts. Seven AI calls, all at once.

For a topic like "covenant," the Torah Scholar flags all five Pentateuch books, the Prophet flags most prophetic books, the Apostle flags Romans, Galatians, Ephesians, Hebrews. The Sage might only flag Psalms. The Historian flags books dealing with covenant violations and renewals.

The system now knows exactly which books need deep analysis.

Phase 2: Massive Fan-Out

This is where it gets expensive.

For every relevant book identified in Phase 1, the system launches two types of AI calls:

A. Book Deep Dive

Each relevant book gets its own dedicated AI call with the full context window. The system allocates up to 6,000 output tokens per book.

The prompt instructs the AI to analyze how that specific book treats the topic. The AI must return structured JSON with:

A theological thesis (3-4 sentences on the book's unique perspective)
Key passages with exegesis
Original language terms (Hebrew/Greek with transliterations)

But here's the trick: I built book-specific style rules to prevent every analysis from sounding identical.

For Genesis, the AI must open with a patriarch's name or narrative event. For Romans, it must use diatribe style with rhetorical questions. For Psalms, it must reference psalm numbers and use poetic parallelism. For Revelation, it must begin with vision references and apocalyptic symbols.

Every book has unique style enforcement. The AI can't fall back on formulaic openings like "This book treats X as..." or "This book does not primarily focus on..."

Those patterns are explicitly forbidden.

B. Genre Overview

Each active genre expert also gets a separate AI call to synthesize the big picture across all its relevant books.

The Prophet receives: "Your active books are Isaiah, Jeremiah, Ezekiel, Hosea, Amos, Micah. Write a comprehensive theological summary of how these books collectively approach this topic."

The output is a multi-paragraph synthesis focusing on the genre's unique contribution.

All of these calls happen in parallel. If the topic is relevant to 30 books across 5 genres, the system launches 30 book analyses + 5 genre overviews = 35 concurrent AI calls.

This is why topical reports cost 3 credits instead of 1. The token consumption is massive.

Phase 3: Master Synthesis

After Phase 2 completes, the system has genre summaries but hasn't yet created the master introduction.

The synthesizer receives only the genre overviews (not the individual book analyses) and generates three critical components:

Definition - A precise 2-3 sentence theological definition
Executive Summary - One paragraph overview of the concept's significance
Redemptive History - A 500-word narrative essay tracing the concept from Creation through the Old Testament, culminating in Christ, and reaching consummation in the New Testament

This is written as flowing prose, not disconnected observations.

The AI is explicitly told: "Do NOT put the topic in quotation marks. Write naturally and theologically. Avoid meta-commentary."

The Gatekeeper

Before any of this runs, there's a gatekeeper.

The gatekeeper validates two things:

Is this a legitimate biblical/theological topic? (Rejects "pudding," "ninjas," "math homework")
Is this a universal topic that appears in almost every chapter? (e.g., "God," "Jesus," "sin")

If the topic is universal, the system flags it with a warning but still processes it. If it's invalid, the system rejects it immediately.

This uses GPT-5-nano, the cheapest model available. It's a simple binary check.

The Final Structure

The orchestrator stitches everything together in canonical order:

Topic
├── Metadata (books analyzed, processing time, token usage)
├── Definition
├── Executive Summary
├── Redemptive History
└── Sections (in canonical order)
    ├── Torah Scholar
    │   ├── Genre Overview
    │   └── Books (Genesis, Exodus, etc.)
    ├── Historian
    │   ├── Genre Overview
    │   └── Books (Joshua, Judges, etc.)
    ├── Sage
    ├── Prophet
    ├── Evangelist
    ├── Apostle
    └── Seer

Each book contains its theological thesis, key passages, and original language terms. Each genre contains its synthesis of how that corpus treats the topic.

Token Usage

The system tracks token consumption across all phases. Phase 2 is the heaviest - when you're analyzing 40 books simultaneously, the token count adds up fast. This is why topical reports cost 3 credits instead of 1.

Why This Works

Three reasons:

1. Parallel Processing

Phase 1 scans all genres simultaneously. Phase 2 analyzes all books and genres simultaneously. The system doesn't wait for sequential AI calls. If 40 books are relevant, all 40 analyses run at once.

2. Specialized Prompts

Each book gets its own style rules. Each genre expert has a distinct persona. The synthesizer has different instructions than the analysts. No single prompt tries to do everything.

3. Structured Output

Everything returns JSON. The system validates against schemas. If the AI tries to return malformed data or skip required fields, it gets rejected.

No Caching

Just like the passage reports, I removed caching. Every topical report is generated fresh. This ensures prompt refinements immediately affect output quality.

The Result

A topic like "righteousness" produces:

A clear definition rooted in both testaments
Redemptive-historical narrative showing how the concept develops
Torah analysis covering covenant righteousness and legal frameworks
Historical analysis showing righteousness in the rise and fall of kingdoms
Wisdom analysis exploring experiential and philosophical dimensions
Prophetic analysis tracing judgment, hope, and messianic trajectory
Gospel analysis showing how Christ embodies and fulfills righteousness
Apostolic analysis covering justification, sanctification, and eschatological righteousness
Apocalyptic analysis revealing ultimate vindication

Each section includes original language work, key passages, and theological synthesis.

If you want to see what these actually look like, check the Share Gallery. Topical reports are marked clearly and anyone can view them without an account. Or generate your own topical report to explore any biblical concept.

God bless, everyone.

Three-column flowchart: blue left—Bible ref and inputs; green center—prompt, AI call, validation; gold right—JSON output

Scripture References

Genesis 12:1-3
Exodus 34:6-7
Psalm 23:1-6
Romans 3:21-26

Key Terms

Exegesis: Careful, contextual analysis of a biblical text to determine its original meaning and implications.
Hermeneutics: The theory and methodology of interpreting biblical texts, including principles that guide understanding across genres.
Mixture-of-Experts: An AI architecture that delegates tasks to specialized models or agents, combining their outputs for a coordinated result.
Structured Output: Formatted, machine-readable results (e.g., JSON) that enforce consistent content and prevent repetitive prose.