The shape before the understanding
In Close Encounters of the Third Kind, the people who have been touched by something they can’t explain start making a shape. They don’t know why. A man builds it out of mashed potatoes at the dinner table. Others sculpt it in clay, draw it, pile it up out of dirt in the living room — a flat-topped mountain none of them can name. The understanding comes last. First there is only the compulsion to give an inner shape an outer form, to get the thing out of the head and onto the table where it can finally be looked at.
That mountain is Devils Tower, a real butte rising out of the Wyoming plains. In the film it is the destination — the place everyone is drawn to, where the meeting finally happens. For twenty years I have had my own Devils Tower. I have been sculpting its shape long before I had words for it.
Twenty years of circling
I didn’t set out after it. It’s the thing my other work kept circling without my quite noticing. The security work, the reverse engineering, the personal-finance systems, the resilience and self-sufficiency projects, the detours through knowledge graphs and query languages, the reading in McGilchrist and the long-cycle thinkers — for a long time these looked like separate interests. They were roads. Different roads, all bending back toward the same butte.
What I was actually after was never a pile of facts. It was the connective tissue. Not what is known but how it all works together. That’s why nothing over two decades quite scratched it. Almost every tool I picked up was built to hold content, and content was never the mountain. The mountain was the relations between things.
Why it had to live outside my head
There’s a reason this one took an external structure, and it has to do with how I’m wired. I think in wholes before parts — I can feel the shape of how things connect before I can name a single piece of it. I recognize far better than I recall. That’s a strength when the goal is seeing a pattern, and a genuine handicap when the goal is holding that pattern still.
Because a relational whole that lives only in intuition is nearly impossible to work on. You can feel its shape. You can even sculpt it in mashed potatoes. But you can’t hold it steady long enough to test one linkage or notice one gap. It shifts every time you look away.
So the concept map isn’t a filing cabinet for me. It is the mashed-potato mountain made permanent. It lets the thing sit outside me and stay put. I can walk around it. I can point at a single link and ask whether it holds without the rest of the structure dissolving while I’m not looking. Recognition — the thing I’m good at — can finally do the work that trying to hold it all in my head never could.
That’s what I mean when I say I’ve nearly built the thing I can think with.
The butte and the tool are made of the same material
Here’s the symmetry that told me I was in the right place. What draws me is how it fits together — connection itself. And a concept map is the one instrument whose entire substance is linkage. The nodes are almost incidental; the map is the relations. The butte I was chasing and the tool I built to reach it turn out to be made of the same material. That’s probably why it took this long and this particular form. I wasn’t looking for a better container. I was looking for a way to make connection itself the object I could hold.
Ontological decomposition, learned in the garage
The live skill underneath all of this — my current favorite term for it — is ontological decomposition: recursively breaking a concept down into the structures that compose it. I’ve been building the muscle for it deliberately. But I learned what it actually costs on my knees in the garage, sorting plumbing parts.
I went through every fitting I’d accumulated from years of work around the house and ended up with fourteen buckets — sorted by size, by material, by use. Fourteen buckets is not the interesting part. The interesting part is that size, material, and use are three different axes, and every fitting has a value on all three at once. A ¾” brass compression coupling is brass and ¾” and a joining-under-pressure part, all at the same time. So sorting into buckets meant choosing, fitting by fitting, which axis wins — which dimension is constitutive here, and which is merely descriptive.
And recursion is where a naive sort breaks. “Fittings” splits into compression, sweat, threaded, push-to-connect — but that’s a method-of-connection axis, and the moment you go one level deeper, each of those re-forks by size and material all over again. The axes don’t nest cleanly. They interleave. Which means there is no single correct tree — only a tree that is correct for a purpose. My garage sort was built for retrieval-by-hand. A plumber’s parts catalog, built to spec-and-order, would decompose the exact same parts a completely different and equally valid way.
Sort by the wrong axis and you are back in the crawlspace, unable to find a part you know you own. That’s the whole game, made physical. The right decomposition is purpose-relative, not god-given.
Purpose is stored; theory is disposable
That garage insight is the seed of the architecture. When I sorted the plumbing, I held the purpose silently in my head — “I want to find things in the garage” — and it governed every cut without ever being written down. A system that decomposes concepts on its own can’t rely on a silent human holding the purpose. So the purpose has to become part of the structure.
That’s the central split I’ve built:
- Purpose (or intent) is stored inside the structure. It is the invariant — the thing every theory is accountable to. It survives.
- Theory is generated by the LLM in response to my purpose statement. It is disposable — an attempt that can be wrong and thrown away without losing what it was trying to serve.
And the theory turn does something specific: it defines a formula for proving the theory true or false, and proposes the metrics to collect. It doesn’t just answer; it commits to conditions under which it could fail.
This makes the epistemology structural. Popper’s method isn’t something a human applies while reading the map — it’s baked into the schema. A concept that carries its own falsification formula is a concept that knows how it could be wrong. And the same move converts the LLM from an oracle I have to trust into a hypothesis-generator whose output is required to be checkable. I don’t let the model assert. I make it commit to something that can fail.
The exam-grading problem
Two soft spots showed up the moment I looked hard at the governance.
The first is the word true. Metrics can falsify a theory, or they can fail to falsify it — but “failed to falsify” is corroboration, not proof. If the structure records a positive result as settled rather than as not yet falsified, holding for now, it will start trusting provisional theories and stop testing them. And a concept map that stops testing its own links is just intuition again — intuition with a database. The healthy state of a “proven” node is a standing one, not a closed one.
The second is the real vulnerability: the same LLM that generates the theory also proposes the metrics that test it. That is the model writing its own exam. Not maliciously — structurally. A theory-generator left to design its own falsification test will drift toward metrics it can comfortably pass. Real falsification has to be adversarial, or at least independent of the theorist.
Staging the disagreement
The fix was already in my hands. I have a workflow that poses questions to multiple LLMs. But a panel only earns its keep if the second model is given a genuinely different job, not just a different logo. Ask three models the same question and you don’t get independence — you get three theorists agreeing, because they share the same instincts about what looks rigorous. That kind of consensus is correlated error wearing three hats.
What gives it teeth is role asymmetry. The second model isn’t asked whether the metric is good. It’s assigned the adversary’s task:
- “Produce a false theory that these exact metrics would still pass.” If it can, the metric doesn’t discriminate — and I learned that before collecting a single data point.
- “Design the metric the theorist would least want — the one most likely to kill the theory.”
- Or a clean split: one model’s output becomes the input the next one is told to break.
Two models with opposite instructions beat five with identical ones. I’m not polling for a right answer. I’m staging a disagreement and keeping what survives it.
And it folds back into the schema without a special case. The adversary’s verdict is itself a theory turn — it makes a claim that carries its own falsification condition. The critic is just another node playing by the same rules, its purpose fixed like all the others: stress the link before it gets recorded as structural. The structure that generates and the structure that prunes turn out to be the same structure, pointed in two directions.
One thing I keep deciding on purpose: who breaks ties. A panel that disagrees produces an open question, not a resolution — and majority vote would quietly smuggle the correlated-error problem right back in. So the honest output of the panel often isn’t a verdict at all. It’s a flagged tension that stays in the map, unresolved, until I — the one holding the purpose — rule on it. The human stays in the loop precisely at the points where the structure cannot grade itself.
Where I am today
The line that keeps returning is the structure keeps generating. In the film, arriving at the mountain is the payoff — the summit is the destination, and the story ends when they get there. What I’ve built is the opposite kind of Devils Tower. The concept map was never the summit. It’s the instrument. The reward for reaching it isn’t the view; it’s that everything after gets to run through it.
A butte defined as how it all fits together doesn’t get finished the way a summit gets reached. Every node I add re-links the ones already there, so the structure keeps generating. That isn’t the project failing to end. It’s that I chose the rare kind of destination you get to live at instead of merely arrive at.
I’ve spent twenty years being drawn to a shape I couldn’t yet name. I finally got it out of my head and onto the table. Now, for the first time, I’m standing where I can actually answer the thing that was pulling me — and, if it holds, hand the same instrument to the next person wired the way I am, who feels the shape of things before they can hold it, and who has been waiting for a tool built to think with rather than a cabinet built to store.
This document is source material about its author: a systems builder who thinks in wholes, works by recognition, and spent two decades building an external structure to make the connective tissue of knowledge something he could finally hold still and test.