The Presuppositional LLM Wiki: A Zettelkasten That Knows What It Believes

Niklas Luhmann, the German sociologist, published around seventy books and hundreds of articles. Asked how, he pointed at a wooden cabinet of paper slips, about ninety thousand of them, one idea per slip, each linked to others by a numbering scheme. He called it his Zettelkasten, his slip-box, and he described it not as an archive but as a communication partner. He did not sit down to write books. He sat down to converse with a box that had been thinking alongside him for decades, and books came out.

Most people who imitate Luhmann copy the box and miss the system. The box is not a filing cabinet. It is a supply chain.

I write the series on this site (theology of psychology, of technology, of economics, and of law) out of a descendant of Luhmann’s box: a plain-Markdown wiki that a large language model reads, files, links, and retrieves for me. But with one addition that changes everything, and that this article exists to explain. Before my wiki accepts a single claim from any source, it checks that source against a written, ordered hierarchy of my presuppositions. And it knows which of my convictions outrank which.

I call the result a presuppositional LLM wiki, a PLLMW, if you can forgive one more acronym in a world drowning in them. Here is what it is, why “just-in-time” is the right way to understand it, how it produces the articles you will read here, and how to build the foundation of your own. That last part comes with a free tool.

Just-in-time, for writing

Start with the supply chain, because the analogy is load-bearing.

Toyota’s production system, built out by Taiichi Ohno across the mid-twentieth century, rests on a distinction between push and pull. A push system manufactures parts against a forecast and warehouses them. Warehouses feel safe and are in fact waste. Ohno’s word is muda: capital sleeping on shelves, parts rusting, obsolete inventory nobody will ever install. A pull system inverts the flow. The assembly line signals for a part (the signal card is called a kanban) and the part arrives at the station at the moment of use, already inspected. Inventory stops being a hoard and becomes a flow.

Now look at how most of us research a writing project. We push. We highlight books we will never reopen. We save PDFs into folders named to-read. We keep four hundred browser tabs as a kind of prayer. Then the writing day arrives and we discover that the warehouse has spoiled: the highlight has lost the thought that made it worth highlighting, the quote has lost its context, and we cannot remember why we saved any of it. Batch research is push manufacturing, and it produces what push always produces: a warehouse of waste and a line waiting for parts.

Luhmann’s system is pull. Each slip is atomic: one claim, stated in your own words, with its source. Each slip is addressed: linked to the slips it answers, extends, or contradicts. Nothing enters the box as raw material; everything enters as a finished part, machined at the moment of reading, when the context is still hot. Then writing changes character. You do not face a blank page; you face a bin of parts your past self already made. The outline is a kanban card. It pulls. Drafting becomes final assembly.

That is the sense in which a zettelkasten is just-in-time supply chain management for writing. The thinking is distributed across years of reading; the writing pulls it to the line on demand.

The model runs the warehouse

The honest objection to Luhmann’s method has always been labor. He filed slips for forty years. Most of us have jobs.

This is the part that has changed. Andrej Karpathy sketched the idea of an LLM wiki: a knowledge base in plain Markdown files that a language model compiles and maintains. You hand it sources, it writes and cross-links the pages, keeps an index, and sweeps for contradictions, stale claims, and broken links. No exotic database. The Markdown is the system; the model is the librarian. My version runs three standing workflows: ingest (a source becomes atomic, linked pages), query (a question becomes a cited answer drawn from the pages), and lint (the box is audited for rot). Four more carry a source the rest of the way to an article: CAQR, the written source review that decides whether anything may be extracted at all (its four parts get their own section below); spec, where the standing instructions a series is written against are drafted and revised; draft, where an outline pulls vetted notes into prose; and revision, where the marked-up draft comes back and gets fixed. One workflow comes before all seven, and it is not the model’s: I read the source myself and mark it up by hand, and what I marked is what goes to review or to ingest. Knowledge compounds in files I own and can read with my eyes.

In supply-chain terms: the model is the materials-handling system. It receives, machines, bins, and delivers the parts. I run the line.

And there the trouble starts.

The credulity problem

A language model, left to its defaults, is the most credulous librarian ever hired. It will ingest Moses and Freud into adjacent bins with identical confidence. It was trained on the average of the internet, and its native posture is the average of the internet, a posture that presents itself as neutrality.

But neutrality is not the absence of presuppositions. It is a particular set of them, unexamined. Cornelius Van Til spent a career on this point: there is no neutral ground; every chain of reasoning hangs from commitments that were in place before the reasoning began. The only question is whether your commitments are stated or smuggled. A wiki that pretends to have no doctrine of authority has one: it is whatever the training distribution happened to believe, weighted by volume. For a writer working from conviction, that is not a neutral tool. That is an unaudited supplier with a majority stake in your conclusions.

Manufacturing solved this problem generations ago and did not agonize over it. No serious factory installs a part because it showed up on a truck. Suppliers are qualified before their first shipment. Incoming inventory is inspected against a written standard. Parts carry traceability: you can walk any component in the finished car back to a lot, a line, a supplier audit.

Ideas deserve the same discipline. That is the whole insight. The rest is implementation.

The presuppositional LLM wiki

A presuppositional LLM wiki is an LLM wiki with a quality system bolted to the loading dock. Five components:

1. The stack, written down. My presuppositions live in the wiki as first-class documents, pillar pages the model must read before it evaluates anything else. Not vibes in my head. Text in the repository, ordered from load-bearing to revisable. (The order matters so much that it gets its own section below.)

2. The gate. No source is extracted from until it has received a structured, written review against the stack. I call mine a CAQR: Critique (where does this author stand against my pillars: orthodoxy, anthropology, epistemology, purpose), Appraisal (what here can be safely taken, and what is the source’s vetted quality), Questioning (what would I ask the author; what needs independent verification), Recapitulation (the honest summary of what this source is). I read and approve every review before a single claim enters the box. The gate is an approval gate, not a book-burning: a source that fails review is quarantined with a written record of why, not erased. Sometimes the review is the most useful document the source ever produces.

3. Confidence ceilings. Every note extracted from a source inherits a ceiling from that source’s review: no claim in the wiki can ever be more confident than the vetted quality of the place it came from. This is traceability for ideas. Ask my wiki why it believes something and it can walk you back to the lot number.

4. Egyptian gold. Augustine, in On Christian Teaching, reads Israel’s exodus as a figure: the people left Egypt carrying Egyptian gold, because the gold was God’s wherever it sat, and it could be melted down for the tabernacle. His point: truth found among unbelievers belongs to the Lord and may be taken and rightly used. Operationally, the gate distinguishes a source’s observations from its teleology: what it saw from what it says the seeing is for. Freud observed real things about repression and self-deception; his account of what a human being is fails my first pillar. The data comes in with a ceiling. The anthropology stays at the border, with paperwork. The gold is plundered; the idols are not.

5. The human stays editor-in-chief. The model drafts, files, links, and retrieves. It approves nothing. Every source review, every revision to the stack, every article that ships crosses my desk. An LLM wiki without a human gate is not a supply chain; it is a conveyor belt into your own mind.

For the manufacturing-minded, the mapping is exact:

Toyota	Presuppositional LLM wiki
Supplier qualification	The written stack; author standing assessed at the gate
Incoming inspection	The source review (CAQR) before any extraction
Lot traceability	Confidence ceilings + citation back to the reviewed source
Kanban pull	The outline’s wikilinks pulling vetted notes into the draft
Standard work	The series and article specs (more below)
Andon cord	Lint: contradictions and stale claims stop the line
Kaizen	The wiki compounds; every article leaves the box smarter

The stack itself: levels of theological triage

Everything above depends on the stack being ordered, and this is the part most people skip. A flat list of beliefs cannot run a gate, because the gate’s first question about any challenge is not “do I agree?” but “where does this disagreement live?”

The ordering discipline is what Albert Mohler named theological triage and Gavin Ortlund’s Finding the Right Hills to Die On developed at book length: doctrines are not all load-bearing in the same way, and maturity is knowing which are which. My stack, from the top:

Level 1: Orthodox Christianity. Defined, for me, as affirmation of the Apostles’ Creed, the Nicene Creed, and the Athanasian Creed; the unique authority, trustworthiness, and inspiration of Scripture; and the biblical definition of marriage. This is the boundary of the faith itself. It is not held pending better evidence, and the gate treats a challenge here differently from every other kind.

Level 2: Conservative Neocalvinism. Where I stand within orthodoxy: the Reformed tradition in its Kuyperian stream, God sovereign over every sphere of life, grace restoring nature rather than escaping it, the cultural mandate still in force, held in its confessional, conservative form.

Level 3: Long Postmillennialism. My eschatology: the conviction that the gospel wins in history, slowly (leaven, not lightning); a horizon measured in centuries. This is an interpretive framework, and it earns its keep daily: it governs how I read providence, progress, and every utopian and doomer claim about technology, and especially about artificial intelligence, that arrives in the news and in my social feeds.

Level 4: Missional Theology. Posture: the church is sent into culture, neither a fortress that abandons the world nor a mirror that flatters it.

Level 5: Transformational Christian Psychology / Transformational Pneumatology. The working layer: my operating theory of how the Holy Spirit actually changes persons, and therefore of what counseling is. This is where the psychology series does its work, held with conviction and open hands at once.

Three rules make this a triage rather than a trophy case:

Each level interprets under the levels above it. My psychology answers to my pneumatology; my pneumatology answers to the creeds. Nothing at level five is permitted to quietly renegotiate level one, which is precisely the move secular psychology asks of you, and precisely what the gate exists to catch.

Revisability increases as you descend. Level five revises on a strong journal article and better exegesis; I expect it to keep moving as long as I keep reading. Level three would take a generation of exegetical argument to move. Level one is not on the table, and saying so out loud is more honest than pretending everything I believe is equally up for auction.

Fellowship widens as you ascend. I share level one with every Christian who has ever lived. Level two, with a tradition. Level five, with perhaps a few dozen colleagues. So the stakes of any disagreement are priced before the argument starts: a writer who denies the Nicene divinity of Christ has failed level one. His theology is quarantined even while his observations may be plundered. A Reformed brother who reads Revelation differently disagrees with me at level three. He is a conversation partner I will cite for the rest of my life. A fellow counselor with a different model of change differs at level five. That is a journal club, not a schism.

Without the ordering, one of two diseases takes hold. Either every disagreement feels like an attack on the faith (everything becomes level one, and discernment dies of hypertension), or nothing feels worth defending, everything becomes level five, and conviction dies in its sleep. The triage is the treatment for both.

Clinical preferences: the second ladder

The stack above is theology, and theology governs. But I do not sit with a client as a theologian. I sit with them as a counselor, and counselors carry a second set of commitments: working accounts of what a person is made of, what goes wrong, and what makes change happen. Those are not doctrines. They are also not interchangeable, and holding them flat breeds the same two diseases a flat theology breeds. So I ordered them too. Four levels, not five; the clinical list is shorter because it is doing less work.

Level 1: Interpersonal neurobiology (IPNB). Daniel Siegel’s synthesis: the mind is embodied and relational, it forms in the space between people, and the brain keeps reorganizing itself in response to relationship for as long as a person is alive. This is the floor. It decides what I am looking at before I have looked at anything: a nervous system in its relationships, the history that shaped it, the regulation actually available in the room. A level-one clinical commitment is one I bring to every client and every method without choosing to. If it turned out to be wrong, most of what I do would have to be rebuilt from the slab.

Level 2: Emotionally focused therapy (EFT). Sue Johnson’s attachment-based model, developed first for couples: the argument is rarely about the dishes, it is about whether you will be there when I reach for you. Level two is my default map for relational work. I will use a different map for a particular case, but I need a reason, and I notice myself needing one.

Level 3: The Carnes task model. Patrick Carnes’s task-based approach to recovery from sexual addiction, the model I am trained and certified in as a Certified Sex Addiction Therapist (CSAT). Level three is a framework I reach for when the presenting problem calls for it. It is domain-specific by design, and it does not pretend to be a theory of everything.

Level 4: Twelve Step work. The fellowship structure that came out of Alcoholics Anonymous and now runs in a hundred rooms for a hundred problems. I recommend it often and require it never. Level four is a tool: decisive for some clients, wrong for others, and not something I would argue about with a colleague past the end of a case consultation.

The same three rules that make the theological list a triage make this one a triage.

Each level is read under the ones above it. How I time a Step is shaped by where the client sits in the task model; the task model is shaped by attachment; attachment is shaped by what I believe a mind is. When a Step lands crosswise of a client’s attachment reality, attachment wins and the Step waits.

What it would take to move me grows as you climb. I will set level four aside for a client in a single session. Level three moves on outcome research and on what my own caseload keeps showing me. Level two would take a serious body of contrary evidence. Level one would take a decade of it, and I would know I was losing a foundation rather than changing a technique.

Agreement widens as you climb. Nearly every competent clinician I know already works as though relationships shape brains, whatever vocabulary they use for it. Fewer run EFT. Fewer still are CSATs. So the price of a disagreement is set before the conversation starts: a colleague who never uses Twelve Step work is a colleague, and a supervisor who thinks a person can be understood one skull at a time is doing a different job than I am.

Then the part that is easiest to fudge. This ladder sits underneath the other one, all four rungs of it. A clinical commitment does not get to renegotiate a theological level, however well it works in the room. Where the two collide, the clinical side yields, and it yields the way the gate already handles any outside source: the observations come in, the teleology stays at the border. What Siegel saw about integration is data I use daily; what he says integration is finally for answers to level one of the theology, not to itself. Twelve Step work names a higher power as you understand him, which is a level-one collision and not a small one; I keep the structure of the Steps and I do not leave the Person unnamed. That is the whole arrangement. My clinical preferences are real preferences, held with real conviction, and they are still the second ladder.

A colleague who shares none of my theology still has reason to care, because the ladder is portable. The ordering is a discipline, not a doctrine. A thoroughly secular clinician already has a stack whether or not it has ever been written down: something at the bottom deciding what a person is, something above it serving as the default map, something else that is a specialty framework, and a handful of tools. Write yours out. Rank it with the same three questions: what governs what, what evidence could actually move each one, and what it would cost you to disagree with a colleague there. The payoffs are the same ones I got. You stop defending a level-four preference with level-one heat, which is the quiet cause of a great many bad supervision hours. And you catch the reverse, which is worse and more common: a level-one commitment renegotiated by a good weekend workshop, one comfortable inch at a time, with nobody in the room aware that the floor moved.

Three series, one production line

Now the payoff: what this machinery actually builds.

Each domain I write in gets three series. (Yes, three series: English declines to give the word a plural, which is inconvenient for a man launching several sets of them.) The pattern is the same every time; only the domain changes:

Narrative Expository [domain]. Scripture itself, walked pericope by pericope, exposited on its own terms, and then the discipline’s concepts articulated from the text: the passage read as an engineer, an economist, a clinician would read it, so that practitioners find their own categories rising out of the exposition rather than being pasted onto it.
Historical Christian [domain]. The tradition’s own works, from the Fathers through the Reformers to the moderns, one portion of one work at a time. The church has been thinking about work, wealth, tools, and the soul for twenty centuries; most of that thought is simply unread.
Secular [domain] in Christian Perspective. The discipline’s major non-Christian works, exposited faithfully as data (steel, not straw) and then tested through a Christian anthropology. Egyptian gold, worked at series scale.

One question routes every source: what is the object of exposition, and who wrote it? Scripture routes to the first series. After that, the line is the author’s own profession of faith, never the subject matter. A professing Christian writing coolly technical media theory belongs to the historical Christian series; a post-Christian writing on the same subject belongs to the third. The three cross-reference constantly, because they are one argument wearing three kinds of source.

The pattern began with psychology: Narrative Expository Psychology, Historical Christian Psychology, and Secular Psychology in Christian Perspective, the triad I developed in my counseling work, which will publish here alongside its younger siblings. Technology, economics, and law are scaffolded now. Four domains, twelve series, one method, and one wiki warehouse behind all of it, where a note vetted for the psychology series is already on the shelf when the technology series needs it. That is the just-in-time dividend: the domains share inventory.

The specs (a tease)

A production line needs standard work: the written instructions that make the thousandth part match the first. Every series here runs on specifications: a common spec per domain (voice, citation discipline, publication rules, and three coverage appendices) and an article spec per series (the fixed architecture every article must satisfy).

The specs stay unpublished. They are the shop’s quality system, and standard work hangs at the workstation, not in the showroom. But you deserve to know they exist and what they enforce, because you will see their fingerprints on every article:

A voice spec I privately call Academic Hemingway: scholarly rigor at plain-language directness. Short sentences carry the argument. No adverbs of emphasis. Every technical term defined on first use: the engineer gets “hamartiology” defined, the pastor gets “affordance” defined, and neither ever needs a secondary source.
A fixed eight-section architecture for every article: exposition of the text or work; the concepts named; their development traced across the Bible’s six-act storyline (Creation, Chaos, Covenant, Christ, Church, Consummation) once for the theology and again for the discipline; formation and practice; charitable engagement with opposition, steel-manned by rule; and prophetic confrontation, which is required to begin with an inward turn: the author’s own idols get named before the culture’s.
A credible-witness check in the historical and secular series: before a word of a work is exposited, the article answers on what basis, and to what degree, you should trust its author as a guide, calibrated, for Christian authors, against Scripture’s own qualifications for those who teach.
Three coverage appendices per domain: a disciplinary cross-walk (so the sweep covers what the discipline’s own curricula cover), a register of practice concerns, and a register of idolatries and ideologies, each swept against every text, bounded by what the text actually warrants. A checklist, never a quota.
A curated canon rule where a domain’s biblical warrant is uneven, with a twist I have grown fond of: a pericope can earn its slot by domain warrant (the text genuinely speaks to the subject) or by abuse warrant: texts routinely weaponized in the domain’s debates are on the list precisely so they can be rescued from their abusers, both flanks of them.

The specs run to many thousands of words each, and they are the real answer to a fair question: how can AI-assisted writing be trustworthy? The same way any manufacture is. Not by trusting the machine. By gating it.

The architect and the inspector

Which leaves the fairest question of all: if the model files, links, retrieves, and drafts, who exactly is writing these articles?

Here is the plain answer. The model does most of the typing. I do two other things, and the two other things are where authorship actually lives. I am the architect, and I am the inspector.

The architect draws before anyone builds. Every constraint these articles are written under existed as text before a single draft was generated: the presupposition stack, ordered and written down; the specs, with their fixed sections and their coverage appendices; the curated canon lists, a warrant recorded beside every entry; the source reviews, each one read and approved by hand. That is the blueprint set, and every line of it is my judgment, argued out slowly, revised, owned. When the model drafts, it drafts against those blueprints the way a framing crew builds against drawings. Nobody walks through a finished building and asks whether the architect swung the hammer; the building is still his, and if it falls, that is his too.

The inspector signs after the work is done, or refuses to. Nothing in this operation publishes without my approval. Every source review is approved before a claim is extracted. Every draft is read against the text of Scripture, against the cited sources, against the stack, and corrected until it says what I mean, sometimes lightly, sometimes down to the studs. Then it ships under my name, which means its errors are mine, and its corrections will carry my signature too. An unsigned draft is not an article; it is inventory.

Between the blueprints and the signature, the machine contributes something real: labor, fast, tireless, uncomplaining. What it cannot contribute is authority. It does not decide what I believe, which witnesses qualify, or what ships. Those decisions live in the drawings it builds against and at the desk where its work stops. That division, the machine’s labor under the man’s authority, is no apology and no workaround. It is how making has always scaled: the master’s drawings, many hands, the master’s inspection at the end, and the master’s name on the work, answerable for it.

Build your foundation first (free tool)

Everything above rests on one prior act, and it is the one act no model can do for you: you have to articulate what you believe, in order. Not a statement of faith you signed once, but a working stack: what is load-bearing, what is confessional, what is framework, what is posture, what is theory held with open hands. Most Christians I know have never written it down. Most have never ranked it. And an unranked faith is exactly the kind that either hardens into brittleness or erodes into mush the first time a brilliant unbeliever makes a true observation.

So the companion tool to this article is the first thing my method would hand you: the Presuppositional Triage skill, a free, portable instruction set you can load into whatever model you already use (Claude, ChatGPT, Gemini, or a local model; it installs as a skill or pastes as a prompt). It will not tell you what to believe. It interviews you, Socratically, one question at a time, then helps you sort your convictions into five triage levels using three tests (what governs what; what evidence could move each; what each would cost in fellowship), stress-tests the result for mismatched grip (the level-five opinion you defend with level-one heat, we all have one) and then produces two artifacts you keep: a Presupposition Profile, the founding document of your own stack, and an LLM Gate Block, a short preamble that makes any AI you use evaluate sources through your hierarchy and tell you which level a challenge actually lives at.

That gate block is a presuppositional LLM wiki in seed form. Plant it in a chat today; grow it into pillar pages when you are ready.

→ Download the Presuppositional Triage skill (free).

The box only returns what you put in

Luhmann liked to deflect praise toward the cabinet, and the deflection was half true. The box made him prolific. It never made him Luhmann. It returned, ordered and juxtaposed, only what he had already thought: the slips were his, every one.

Same here. The model is a magnificent materials handler: it files without fatigue, links without forgetting, and retrieves at the exact moment of need. That part is just-in-time, and it is a genuine revolution in the writer’s supply chain. But the convictions on the shelves were reviewed onto them one source at a time, against a stack I wrote down and ordered and own. The just-in-time part is the machine. The presuppositional part is the man.

Which is as it should be. Of all people, Christians should be unafraid of powerful tools and unwilling to be discipled by them. We have known since Exodus what to do with Egyptian gold, and we have known since Eden that making is the assignment, not the Fall. The story we are writing toward does not end in a warehouse. It ends in a city, one whose builder already knows what he believes.

The four series families (psychology, technology, economics, and law) will publish here as articles clear review. If you want the method’s foundation for yourself, start with the Presuppositional Triage skill above.

References and further reading

One more honesty the method requires. My wiki refuses to blur what has actually been read with what is merely known about: every source in it carries an honest read-state, and a work known only by summary is never presented as digested. The same rule binds the author. Of the works below, I have read Mohler and Ortlund. The rest I know at working distance: excerpts, summaries, secondary discussion, reputation. They are cited as pointers for your reading, not as books I claim to have digested; as the series engage them properly, they will pass through the gate like everything else.

Read:

Mohler, R. A. (2005). “A call for theological triage and Christian maturity.”
Ortlund, G. (2020). Finding the right hills to die on: The case for theological triage. Crossway.

Known at working distance:

Ahrens, S. (2017). How to take smart notes. The popular treatment of Luhmann’s method.
Alcoholics Anonymous World Services. (2001). Alcoholics Anonymous (4th ed.). The founding text of Twelve Step work.
Augustine. On Christian teaching (De doctrina Christiana), Book II, ch. 40. The Egyptian gold figure.
Carnes, P. (2015). Facing the shadow: Starting sexual and relationship recovery (3rd ed.). Gentle Path Press. The task-based recovery model.
Johnson, S. M. (2019). The practice of emotionally focused couple therapy: Creating connection (3rd ed.). Routledge.
Karpathy, A. LLM wiki sketch. https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
Luhmann, N. (1981). “Kommunikation mit Zettelkästen” (“Communicating with slip boxes”).
Ohno, T. (1988). Toyota production system: Beyond large-scale production. Productivity Press.
Siegel, D. J. (2020). The developing mind: How relationships and the brain interact to shape who we are (3rd ed.). Guilford Press. The founding statement of interpersonal neurobiology.
Van Til, C. (1955). The defense of the faith. P&R.