# How This Corpus Was Built

## A Public Manifesto for Open, Agentic Theological Research

**By Elijah Faviel**

**Portfolio version:** V1

**Publication status:** public V1 preview before the first stable release

People are right to ask how one author could publish a connected body of books,
research records, study tools, and multilingual editions in such a concentrated
period.

The short answer is that the work did not begin when the books appeared. It
began years earlier with *Rethinking Reality*. It grew through a large and
expensive AI research system. And despite the speed of the models, it moved
through a slow human process of reading, redirection, judgment, and
cross-verification.

AI made the scale possible. It did not make the conclusions true. This is the
longer and more honest account of what happened.

## Before the Corpus, There Was One Book

I began planning *Rethinking Reality* in 2019. The book took years to develop
through research, theological and scientific questions, pastoral
conversations, repeated restructuring, and the ordinary difficulty of trying
to say something true about a complicated world.

It was the first book in this corpus, and it received pastoral and theological
feedback during its development. But its notes kept producing questions that
one accessible book could not responsibly carry. How should biblical,
historical, empirical, philosophical, and pastoral claims relate without
borrowing authority from one another? How could original-language Scripture
remain the norm for Christian confession while empirical claims remained
answerable to evidence? How should early Christian witnesses be used without
turning the fathers into a second canon? How could a framework mark the
difference between direct reading, synthesis, analogy, judgment, and genuine
uncertainty?

Those questions became the seedbed of *Divine Design Framework*, or DDF.

The use of AI did not begin with DDF. I was already building AI research
pipelines while developing *Rethinking Reality*, and I estimate that roughly
half of the research that eventually fed the wider corpus began there. DDF
inherited that material and experience, then made the method far more formal,
wide, deep, and demanding about cross-verification.

## Why DDF Had to Be Built

In conversations with pastors, engineers, thinkers, and people trying to make
sense of Scripture and modern life, I kept finding the same need. They needed
more than another set of answers. They needed a disciplined way to research
biblical concepts, test connections, preserve distinctions, and turn what
survived into books and practical tools that could still be corrected.

The scope was larger than any one specialty. A person can devote a lifetime to
biblical studies, patristics, philosophy, physics, biology, psychology,
clinical care, history, artificial intelligence, or pastoral practice and
still remain finite. Serious work across all of them ordinarily requires a
multidisciplinary community and generations of accumulated expertise.

This project does not pretend that earlier scholarship failed or that the
research came from nowhere. It depends on Scripture, ancient witnesses,
primary documents, generations of scholarship, modern research, and the
Church's long memory. To research “from scratch” meant something narrower: I
did not want to adopt one modern theological system, denominational handbook,
philosophical school, or popular summary and make every source fit it. I wanted
to return as close as possible to original-language Scripture, ancient textual
witnesses, patristic documents, primary historical materials, technical
papers, and observed reality, then rebuild the connections under pressure.

AI made that scale attainable.

## Why AI Did Not Make the Work Fast

My own background made this method conceivable. I am a principal engineer and
an author of fiction and nonfiction. My career has included building
large-scale production systems and working in AI across multiple companies. I
am comfortable with agents, pipelines, versioning, failure modes, and very long
documents.

That background explains why I could build and inhabit this process. It does
not make me a theologian, historian, scientist, clinician, or biblical-languages
specialist by title. It does not validate a single conclusion. Credentials are
context, not warrant.

Over many months, I used a changing environment of models and many specialized
agents. The work drew from three model families: xAI's Grok, OpenAI's GPT, and
Google's frontier models. I always used the frontier release from each family
and the highest reasoning level available to me at the time. When a new
generation arrived in any family, I reran important questions and recertified
the relevant conclusions.

This mattered because each family showed recurring emphases and appeared drawn
toward certain kinds of ideas. Moving work between them helped expose when a
conclusion might be an echo of one model lineage, a higher-order loop of model
assumptions, rather than something the sources warranted. Cross-family
comparison did not create independent peer review, but it added another check
against correlated reasoning.

The larger V1 effort represents thousands of hours of agentic work and tens of
thousands of dollars in compute and research costs. Repeating difficult
research across model families and generations at their highest reasoning
levels contributed substantially to that cost. Models mapped unfamiliar
fields, divided large questions into researchable parts, inspected source
families, and compared interpretations. They also translated and reconstructed
documents, searched for counterevidence, built rival accounts, tested bridges
between disciplines, found contradictions, drafted prose, and returned to the
same claims under new constraints.

That sounds fast. It was not.

The models could place biblical exegesis, patristic theology, historical
reconstruction, philosophy, physics, biology, psychology, clinical research,
institutional analysis, and AI engineering in front of me within one project.
Those fields do not share one vocabulary, one standard of evidence, or one form
of argument. Before I could direct the next pass, I had to understand what a
research document was saying, why its sources mattered, what the strongest
competing argument was, and how much weight its conclusion could carry.

I read the research documents presented for retention and integration. I read
the patristic and theological material necessary to understand the historical
arguments. I worked through the competing positions instead of treating an AI
summary as self-validating. I had to learn enough of each local field to notice
when an agent was repeating itself, hiding a disputed premise, misclassifying a
source, or joining two domains too quickly.

Every research document in the durable record, and every research result that
materially shaped the manuscript, passed through that human process. The work
did not move directly from a model response into DDF or from DDF into another
book.

In engineering terms, human comprehension became the rate-limiting step. AI
could generate another research pass in minutes. Reading it, comparing it with
the underlying materials, understanding its rivals, and deciding what research
had to happen next could take days. Cross-verifying a biblical claim against
original-language work, early reception, doctrinal history, empirical
evidence, and possible human consequences took longer still.

The concentrated publication period is therefore not a story of instantaneous
authorship. AI made breadth and iteration possible. Human reading made
responsible synthesis slow.

## What the Research Process Actually Looked Like

There was no clean conveyor belt in which one agent researched, another
verified, and a final model wrote the book. The work behaved more like an
ecology: many loops operating at different scales, sometimes in parallel,
sometimes in conflict, and often reopening questions that had seemed settled.

The exact orchestration changed with the problem, but six kinds of work
recurred.

1. **Map the whole field.** Broad agents separated a question into its textual,
   doctrinal, historical, scientific, philosophical, clinical, technical, and
   practical surfaces. Other agents explored adjacent fields, rival schools,
   neglected periods, edge cases, and questions the original frame could not
   yet see.

2. **Go narrow and return to sources.** Deep-research passes focused on one
   word, text, claim, doctrine, experiment, historical episode, clinical risk,
   or disputed bridge. When summaries were not enough, the work returned to
   Hebrew, Aramaic, and Greek Scripture; ancient versions and witnesses;
   patristic and primary historical documents; technical papers; datasets;
   standards; and versioned model reports. Some materials were transcribed,
   translated, or compared directly across languages. Many questions needed a
   second, third, or later deep dive.

3. **Build the strongest opposition.** Agents searched for counterexamples,
   contrary evidence, category errors, hidden assumptions, unsafe applications,
   and stronger rival explanations. The purpose was not performative
   skepticism. It was to stop the preferred account from grading its own exam.

4. **Attempt synthesis, then test the bridge.** Separate passes tried to connect
   local findings without allowing one field to govern another improperly.
   Every proposed bridge had to return to exact definitions, negative cases,
   the proper source for each claim, and the best alternatives. Many elegant
   connections weakened into analogies. Some disappeared.

5. **Use human direction where the system stalled.** Models often entered a
   conceptual attractor: they repeated one family of ideas in new language,
   defended an inherited frame, or optimized for a balanced and persuasive
   answer when the evidence required a colder result. More autonomous
   iteration then made the circle deeper, not wider.

   At those points I changed the objective or the search space. I required
   truth before fit, separated direct biblical claims from canonical synthesis,
   later doctrine, empirical findings, analogy, and authorial judgment, and
   rejected warmth or doctrinal familiarity as substitutes for warrant. I
   redirected stuck inquiries into areas such as mathematical discovery,
   scientific model formation, sleep paralysis and anomalous experience,
   clinical differentials, institutional abuse, and negative cases. Sometimes
   I demanded a better rival. Sometimes I narrowed an overloaded term.
   Sometimes I reset the work from primary sources.

6. **Recertify at several scales.** Mature claims were revisited at sentence,
   section, chapter, whole-book, and portfolio levels. Different passes checked
   biblical exactness, source roles, claim strength, contradictions, current
   evidence, embodied risk, bibliography, dependencies, and revision
   conditions. Downstream books were checked again so that simplification for a
   church, household, catechetical, or public audience did not quietly make a
   claim stronger.

These were loops, not sealed phases. A pastoral danger could reveal a false
doctrinal compression. A rival account could force a new translation. A source
check could dissolve an attractive analogy. A production audit could send a
finished-looking chapter back into research.

Multi-agent scale mattered because one model in one continuing context tends to
inherit its own framing. Separate agents and fresh contexts allowed breadth to
be distinguished from depth, proposal from opposition, extraction from
synthesis, and drafting from recertification. But these agents were not
independent peers. Models trained on overlapping data can share the same error,
defer to one another, or reinforce the first framing they receive. Five agents
agreeing is not the same as five scholars independently examining a source.

Human direction was not an independent warrant either. I could anchor the
models, import a bias, or ask a question that favored the conclusion I wanted.
That is why the record preserves major methodological corrections and why each
claim still has to survive the relevant sources, evidence, rivals, risks, and
revision conditions. My direction determined where to look. It did not
determine what was true.

## What AI Could and Could Not Contribute

Large language models are trained statistical sequence models, not witnesses
with access to truth. Attention-based architecture, large-scale pretraining,
instruction tuning, human feedback, retrieval, search, code execution, and file
inspection give them remarkable abilities. They can reorganize a difficult
question, compare many texts, expose a missing distinction, generate candidate
explanations, and draft useful prose.

They can also make a wrong answer arrive faster, cleaner, and with footnotes.

The research system was designed around several recurring failures:

- **Fluent invention and stale knowledge.** Plausible quotations, sources,
  facts, and current claims were checked against the underlying document,
  provenance, date, model version, or empirical record.
- **Sycophancy and premature harmony.** Agents were asked to contradict the
  preferred conclusion, construct serious rivals, name negative cases, and
  allow “underdetermined,” “analogy only,” “in tension,” or “contradicted” as
  valid results.
- **Correlated agreement.** Different models, roles, prompts, source routes,
  and fresh contexts were used to diversify the search. Agreement among models
  was never counted as independent evidence. Work on correlated errors across
  large model populations, including
  [Kim et al.](https://arxiv.org/abs/2506.07962), reinforces why apparent
  multi-model consensus cannot carry that weight.
- **Context loss, circularity, and false self-correction.** Work was divided
  into bounded claim families with stable identifiers, ledgers, source maps,
  and later whole-book audits. A stalled loop was given new evidence or stopped
  for human redirection instead of being trusted to think itself free.
- **Automation bias.** Generated reasoning was not treated as a transparent
  record of model cognition, polish was not treated as proof, and final
  authorship and responsibility remained human.

These controls reduce risk; they do not eliminate it. They are also consistent
with the broader AI literature:

- The basic model architecture and the limits of next-token prediction are
  described by [Vaswani et al.](https://arxiv.org/abs/1706.03762) and
  [Kalai et al.](https://doi.org/10.1038/s41586-026-10549-w).
- Research on human-feedback training and sycophancy includes
  [Ouyang et al.](https://arxiv.org/abs/2203.02155) and
  [Sharma et al.](https://proceedings.iclr.cc/paper_files/paper/2024/hash/0105f7972202c1d4fb817da9f21a9663-Abstract-Conference.html).
- Confabulation, long-context failure, unreliable self-correction, and
  unfaithful rationales are documented by the
  [NIST Generative AI Profile](https://doi.org/10.6028/NIST.AI.600-1),
  [Farquhar et al.](https://doi.org/10.1038/s41586-024-07421-0),
  [Liu et al.](https://aclanthology.org/2024.tacl-1.9/),
  [Huang et al.](https://proceedings.iclr.cc/paper_files/paper/2024/hash/8b4add8b0aa8749d80a34ca5d941c355-Abstract-Conference.html), and
  [Turpin et al.](https://proceedings.neurips.cc/paper_files/paper/2023/hash/ed3fea9033a80fea1376299fa7863f4a-Abstract-Conference.html).
- Multi-agent research reports both gains and persistent failure modes:
  [Du et al.](https://proceedings.mlr.press/v235/du24e.html),
  [Liang et al.](https://aclanthology.org/2024.emnlp-main.992/),
  [Oh et al.](https://arxiv.org/abs/2503.16814),
  [Choi et al.](https://aclanthology.org/2026.acl-long.650/), and
  [Becker et al.](https://aclanthology.org/2026.findings-eacl.268/).

AI was therefore a research instrument and generative partner. It was not a
source of revelation, a theological authority, an independent witness, or a
replacement for the people and realities being studied. Model output created
work to be tested. It did not end the test.

## What DDF Became

DDF is an argument, a field guide, and an audit trail. It states its Christian
proposal openly while leaving empirical, historical, philosophical, clinical,
and technical claims answerable to the evidence and methods of the fields
competent to judge them.

Its major claims are governed through claim cards and shared profiles. These
record the assertion, domain and scale, claim type, governing source, warrant,
confidence, strongest rival, limits, risks, dependencies, and the conditions
that would require revision. The V1 register contains eighty-nine stable claim
IDs. The method also permits a bridge to fail. Putting two true sentences next
to each other does not establish a relationship between them.

The public research library shows how this discipline developed. It preserves
claim extraction, source checks, contradiction maps, chapter research,
high-risk findings, calibration, bibliography matching, gap analysis, and
repeated whole-book stress tests. It also preserves important corrections and
the current durable method. Earlier development states remain recoverable
through version history rather than being presented as current guidance.

At the V1 snapshot, the English DDF manuscript and durable research library
contain 59,317 source lines. The ledger moves through thousands of numbered
conclusions and repeated textual, historical, scientific, clinical, source,
rival, safety, and currentness audits.

That volume is evidence of process, not evidence of truth. The public record is
also not a replay of every temporary prompt or discarded branch. It preserves
the durable state needed for examination and revision: source trails, claims,
audits, decisions, manuscript history, and exact repository snapshots.

## What the Framework Concluded

The research repeatedly found that early Christian confession was more
embodied, relational, sacramental, intellectually serious, and resistant to
both material reduction and spiritual escape than many modern summaries
suggest. It also found that Christian theology does not need science to pretend
to prove doctrine, or theology to control empirical results in advance.

DDF's central proposal is that reality is one created order, given through the
personal Logos and received by creatures through distinct but integrated forms
of embodied, rational, relational, moral, and spiritual life. The framework
preserves distinctions because truth requires them, then asks how those
realities belong together.

That proposal aligns in important ways with Scripture's canonical movement and
central early Christian emphases. Creation is good. The Word truly became
flesh. Salvation is not escape from embodiment. Truth and love cannot be
severed. Resurrection and new creation are the Christian horizon.

Alignment does not mean that every father agreed, every historical problem is
settled, or DDF is identical to the early Church. The records preserve
disagreement, development, uncertainty, and places where later doctrine is a
disciplined synthesis rather than direct biblical wording.

## From One Framework to a Library

The chronology matters. DDF did not generate *Rethinking Reality* from nothing.
*Rethinking Reality* came first; its questions and research helped generate
DDF. Once DDF became explicit, it returned to the first book as an audit and
alignment architecture.

The later books were developed through the DDF core, but they are not DDF
chapters wearing new covers:

- *Cognitive Resonance Model* presents a correctable model of how people
  receive, resist, and integrate claims.
- *Truthful Communion* applies the architecture to church life, authority,
  care, protection, repentance, and repair.
- *The Faith That Holds* carries it through an ordinary-language catechism.
- *Households of Formation* applies it to children, caregivers, households,
  church support, safety, and responsibility.
- *DDF Church Blueprint* turns it toward planting, governance, worship,
  operations, accountability, crisis, succession, and closure.
- **DDF in Action** demonstrates the framework through shorter public essays.

The three-language boundary began with *Rethinking Reality*. I personally
revised its English and Spanish editions because I am a native speaker of both
languages. I did not perform its Indonesian-language revision. Gabriela
Tifany, a native Indonesian speaker who has helped me edit books in the past,
revised that Indonesian edition and contributed Indonesian linguistic and
cultural context. The expanded portfolio is available in English, Spanish,
and Indonesian because these are the three languages for which I have direct
access to native speakers. The newer companion localizations remain V1 preview
work and still require direct native-speaker review before stable release.

If you are a native speaker of another language and can honestly commit the
hundreds of hours required to verify and localize this work in your language,
please contact me. I would be happy to work with you.

Agents generated, expanded, criticized, and revised drafts, while the source
rules, DDF coverage gate, ledgers, safety boundaries, and cross-book audits
controlled what could move downstream. The corpus took its concentrated shape
over roughly one to two years. It stands on the longer development of
*Rethinking Reality*, the earlier AI pipelines, the human reading of their
research, and the accumulated sources on which the entire project depends.

## Publication Opens the Circle

Except for the earlier pastoral and theological feedback on *Rethinking
Reality*, V1 was not co-developed through a formal external reviewer program.
Internal agents are not independent peers. Automated audits are not pastors,
textual scholars, historians, scientists, clinicians, safeguarding
professionals, engineers, translators, or readers living with the consequences.

The scale is unusual. It is not a substitute for reception by the Church or
examination by qualified people.

The website publishes the DDF documents, research method, source trails, and
available portfolio materials for download. Readers can inspect the documents
directly, place them into the AI system of their choice, rerun questions,
compare claims with primary sources, find contradictions, or build stronger
counterarguments. The [research hub](https://systemstheology.com/research)
provides the downloads, version information, and public feedback path.

That invitation includes biblical scholars, original-language readers, and
patristic, historical, and systematic theologians. It includes pastors and
church members; scientists and clinicians; safeguarding, legal, privacy,
financial, and governance professionals; engineers and AI researchers; Spanish
and Indonesian reviewers; parents, skeptics, and readers formed in traditions
different from mine.

Feedback is recorded against an exact V1 snapshot. Accepted, modified,
deferred, and declined findings receive a rationale, and accepted changes
remain traceable to their implementation. Participation is not represented as
endorsement. A criticism is not rejected because it came from outside the
framework.

The portfolio is marked V1 and remains V1 throughout this first public
examination. It does not advance to a stable version until it passes the
published stable-release gate.

## Read It Against Reality

Do not trust this work because it used AI. Do not dismiss it because it used
AI. Test it.

Download DDF and ask:

1. Which claims come directly from primary sources, and which are later
   syntheses?
2. Where does the language outrun the evidence?
3. What is the strongest rival account or contrary case?
4. Does a cross-domain metaphor preserve its meaning, or hide a category
   error?
5. Could this conclusion become false, unsafe, or distorted when another book
   applies it?

Then send what you find through the public feedback path.

The aim is not to make AI sound theological or theology sound scientific. It
is to pursue one reality with enough discipline for different forms of truth to
meet without being confused. It also requires enough humility for every part
of the system to be corrected when reality presses back.

The process is large because the questions are large. The records are public
because scale without transparency should not be trusted. Publication begins
examination; it does not end it.

Read it. Test it. Break what is false. Strengthen what is true. Help build what
the Church and the world can actually use.