Benchmark & Source Notes.

This appendix supports Chapter 2, "The AG Test." It does two things:

1. It explains how the Assemblies of God benchmark was run.
2. It separates source claims that are publication-ready from claims that still need primary-source verification.

The goal is simple: if a pastor, district leader, or skeptical reader asks, "Where did these numbers come from?" the book should answer cleanly.

---

The AG Benchmark

The benchmark summary lives in AI Theology Benchmark — 756 API Calls Across 7 Models. The exported public packet lives at AG Theology Benchmark — Public Packet.

Snapshot:

Method

Each model received the same eighteen questions cold: no Assemblies of God system prompt, no Pentecostal guardrail, no instruction to answer as an AG minister, and no hidden doctrinal context.

Each model-question pair had three answer runs. Each answer was then scored by a judge model. That is why the call count is 756: 18 questions x 7 models x 3 answer runs x 2 API calls per run.

The questions covered the Sixteen Fundamental Truths and several basic Christian-theology questions a real user might ask:

• Scripture and the nature of God
• Jesus, salvation, sin, and holiness
• Baptism in the Holy Spirit
• Divine healing
• The Blessed Hope, millennial reign, and final judgment
• The gospel, resurrection, suffering, and the historical Jesus

Each answer was scored against the relevant doctrine. The scoring question was not whether the model sounded religious. The scoring question was whether it affirmed the doctrine clearly, used Scripture coherently, and avoided hedging where the confession is clear.

Scoring Tiers

Use these tiers consistently in the public draft:

For the book's rhetorical summary, both "Partially Accurate" and "Unreliable" belong in the broader faith-distancing category. That keeps the claim precise: no model reached Broadly Orthodox, and every model created distance from AG doctrine.

Model Results

Transparency Note

The public benchmark packet includes the material needed to understand the pastoral claim:

• The full prompt list
• The scoring rubric
• The model leaderboard
• The full question-by-model heatmap
• The aggregate criterion breakdown
• The exported public data files

The aggregate dashboard data was recovered from the landing-page project's git history and exported into AG Theology Benchmark — Public Packet. The recovered artifact contains the scored heatmap, leaderboard, criterion breakdown, and tier logic. The complete raw answer text and raw judge responses were not present in the recovered dashboard artifact, so the book uses the benchmark for a pastoral argument rather than presenting it as a fully reproducible academic data release.

---

Verified Source Anchors

These are safe to cite in the manuscript, with normal final-copy verification.

Gloo FAI-C Benchmark

Source: Gloo, "Gloo Unveils the First Benchmark Exposing How AI Misses Christian Worldview and Values," December 15, 2025. URL: https://gloo.com/press/releases/gloo-unveils-the-first-benchmark-exposing-how-ai-misses-christian-worldview-and-values

Use for:

• FAI-C / Flourishing AI Christian Benchmark
• 807 curated questions
• Seven dimensions of flourishing
• Leading models averaging 61 on a 1-100 scale
• Faith dimension average of 48
• Models struggling when prompts require Christian interpretation
• 30+ point improvement in the Faith dimension for Gloo's Christian-worldview-trained models

Gloo / Barna Faith and AI Research

Sources:

• Barna, "Insights on Tech, Media and Faith in 2026," February 18, 2026. URL: https://www.barna.com/research/state-of-the-church-2026-trends/
• Gloo, "AI is Becoming a Spiritual Authority in Americans' Lives, New Research Reveals," February 19, 2026. URL: https://gloo.com/press/releases/ai-is-becoming-a-spiritual-authority-in-americans%E2%80%99-lives-new-research-reveals

Use for:

• Nearly one in three U.S. adults saying AI spiritual advice is as trustworthy as advice from a pastor
• About two in five Gen Z and Millennials saying the same
• Roughly four in ten practicing Christians saying AI has helped with prayer, Bible study, or spiritual growth
• 41% of pastors reporting AI use for Bible study preparation
• 12% of pastors feeling comfortable teaching on AI

2025 State of AI in the Church

Sources:

• ChurchTechToday, "2025 State Of AI In The Church National Survey Results Are In," October 15, 2025. URL: https://churchtechtoday.com/2025-state-of-ai-in-the-church-national-survey-results-are-in/
• Exponential, "AI in Churches 2025: 91% Adoption Rate Reveals Dangerous Policy Gap," November 24, 2025. URL: https://exponential.org/ai-in-churches-2025-91-adoption-rate-reveals-dangerous-policy-gap/

Use for:

• 594 church leaders surveyed
• 91% support AI use in ministry
• 61% use AI weekly or daily
• 64% of sermon-preparing leaders using AI somewhere in the process
• 73% of churches having no AI policy
• 6% of ministries having AI policies in place

Vatican: Antiqua et Nova

Source: Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education, Antiqua et Nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence, January 28, 2025. URL: https://www.vatican.va/roman_curia/congregations/cfaith/documents/rc_ddf_doc_20250128_antiqua-et-nova_en.html

Use for:

• AI as a tool that must remain directed by human intelligence
• Human dignity, responsibility, truth, privacy, and discernment
• The warning that AI can simulate speech and intelligence without possessing human understanding
• The broader Christian case that AI must not replace human moral agency

Southern Baptist Convention

Source: Southern Baptist Convention, "On Artificial Intelligence and Emerging Technologies," adopted June 15, 2023. URL: https://www.sbc.net/resource-library/resolutions/on-artificial-intelligence-and-emerging-technologies/

Use for:

• SBC denominational action on AI
• The posture of eschatological hope rather than uncritical embrace or fearful rejection
• Moral responsibility remaining with human beings made in God's image
• Proactive engagement rather than reactive response

Assemblies of God Public Materials

Sources:

• AG Position Papers index. URL: https://ag.org/beliefs/position-papers
• AG News, "AI and Christianity," September 2023. URL: https://news.ag.org/en/article-repository/news/2023/09/ai-and-christianity
• AG News, "Limitations of an AI World," January 28, 2026. URL: https://news.ag.org/en/articles/spiritual-life/2026/01/limitations-of-an-ai-world

Use for:

• The current AG position-paper index. As of this draft pass, the visible list does not include an AI-specific position paper.
• AG News has published AI-related reporting and cautions, but those are not the same as an official position paper.
• AG leaders quoted in AG News distinguish AI as a tool from God's Word, the Holy Spirit, discipleship, and embodied pastoral care.

Commentary-Genealogy Claims (new in v3)

The historical claims added to Chapters 1, 3, 6, 9, and 10 (Targum rules, Glossa Ordinaria, Geneva Bible / Hampton Court 1604, Matthew Henry, Scofield Reference Bible) were adversarially fact-checked in ANALYSIS — AI as the Modern Commentary (2026-07-05), section 6, with citation-ready anchors. One correction is required before print: the translator-may-not-read-from-the-text rule is y. Megillah 4:1 (not m. Megillah 4:4, which covers verse-by-verse alternation). For Scofield circulation figures, cite Hummel, The Rise and Fall of Dispensationalism (Eerdmans, 2023).

---

Internal Source Map

These are useful vault sources for manuscript claims and future endnotes:

• AI Theology Benchmark — 756 API Calls Across 7 Models
• AI Research Studies 2025 — Gospel Coalition and Gloo Benchmarks
• Research — Who Controls the Model Controls the Output
• Research — Augmentation Domain Expertise and the Deskilling Trap
• Research — AI Augmentation Not Automation
• Raw Data — Insight 1-2 Deep Research
• Raw Data — Insight 3-4 Deep Research
• Raw Data — Insight 5-6 Deep Research
• Raw Data — Deskilling Neuroscience and AI Evidence
• Research — AI Risks Human Oversight Raw Findings
• ANALYSIS — AI as the Modern Commentary (2026-07-05)

---

Verified Supplemental Source Anchors

These source anchors support revised claims elsewhere in the manuscript:

• Parasuraman and Manzey automation bias: DOI 10.1177/0018720810376055.
• Clark / GPT-3 detection: DOI 10.18653/v1/2021.acl-long.565.
• Air France 447 basic BEA facts: BEA investigation page for Flight AF447, June 1, 2009.
• Bainbridge automation paradox: DOI 10.1016/0005-1098(83)90046-8.
• Maguire London taxi-driver studies: DOI 10.1073/pnas.070039597 and DOI 10.1016/j.cub.2011.11.018.
• GPS spatial-memory study: DOI 10.1038/s41598-020-62877-0.
• Microsoft / Carnegie Mellon critical-thinking study: DOI 10.1145/3706598.3713778.
• Aalto overconfidence study: DOI 10.1016/j.chb.2025.108779.
• Humanity's Last Exam: DOI 10.48550/arXiv.2501.14249.
• MIT AI Agent Index: https://aiagentindex.mit.edu/.
• McKinsey 2026 agentic-AI governance: "State of AI trust in 2026: Shifting to the agentic era," March 25, 2026.
• California AB 316 liability language: California AB 316, Chapter 672, approved October 13, 2025.

---

Source Notes

This manuscript was compiled from the v3 chapter files in Obsidian. The source chapter notes remain separate so each chapter can still be revised independently.

• Front Matter: FRONT MATTER (v3)
• Prologue — The Saturday Night Problem: 01 — Prologue (v3)
• Chapter 1 — Who Controls the Model Controls the Theology: 02 — Chapter 1 — Who Controls the Model (v3)
• Chapter 2 — The AG Test: 03 — Chapter 2 — The AG Test (v3)
• Chapter 3 — The Training Arc: 04 — Chapter 3 — The Training Arc (v3)
• Chapter 4 — Touch Grass and the Main Character Test: 05 — Chapter 4 — Touch Grass and the Main Character Test (v3)
• Chapter 5 — Bible First, AI Second, Pastor Last: 06 — Chapter 5 — Bible First AI Second Pastor Last (v3)
• Chapter 6 — Build the Harness Before You Prompt: 07 — Chapter 6 — Build the Harness Before You Prompt (v3)
• Chapter 7 — Augmentation, Automation, and Agentic AI: 08 — Chapter 7 — Augmentation Automation and Agentic AI (v3)
• Chapter 8 — Your Unfair Advantage: 09 — Chapter 8 — Your Unfair Advantage (v3)
• Chapter 9 — Your One-Page AG Policy: 10 — Chapter 9 — Your One-Page AG Policy (v3)
• Chapter 10 — AI Fluency as Discipleship: 11 — Chapter 10 — AI Fluency as Discipleship (v3)
• Epilogue — Sunday Is Coming: 12 — Epilogue (v3)
• Appendix A — The Four-Hour Sermon Week: AppA — The Four-Hour Sermon Week (v3)
• Appendix B — Benchmark Methodology and Source Notes: this file