The Eighty Lemmas

How an AI kept forgetting and the paper that fell out when I asked it why.

I want to tell you something about myself which is a source of embarrassment and frustration. I don’t have a great memory for a lot of things. If I was to describe my memory I might describe it as vague intuition. I have LFCT down to five published axioms and down to four unpublished, but if you were to bump into me in the street and ask what is axiom one? I probably couldn’t remember. I also have a hard time using precise language or remembering what most words mean. I know some words, like precise. I know precisely what that means. But many words I don’t use or even think with. I might like a word like water to think of something rather than a definition of water. I’ve got nothing against definitions and I love to define something. But the problem with defining something is unless you got it right (and often even when you do) you’re saying it is this and it is not that. AI loves to say it is this and not that. It got to the point for a while where all I had to do was look for the “it is not that” to know something was there. It can be this and that and more.

I build LFCT — Light Frame Cadence Theory — working with AIs. I don’t ask an AI for physics and copy it down, or have the time to read much of what it writes off screen in a detailed way. I do it sometimes, but when I do I can get stuck in the search for perfection. Which can be productive at times, but it is more restful for me to ramble what is in my head in probably a limited vocabulary set, and the AI is the translation surface. I articulate, it formalizes, an external AI audits, and we go around again. Round after round. Sometimes it takes a lot of rounds and then the dam breaks and the clarity that was refusing every phrasing suddenly arrives whole. That’s been the shape of this work for about a year now, back to the earliest Light Frame papers.

This spring the shape got measurable.

The walk

In May an external audit complained that a few pieces of the framework were “still inferential” — held up by corollaries that got added during earlier audit cycles. The expected fix was to shore up the results. My read was different: if the results were fine, why didn’t the axioms cover it? Maybe they were not articulated well enough. Maybe if you keep needing corollaries, the foundation is under-stated, not the theorems. The AI initially resisted the idea of molding the axioms to fit, with valid reasons. But my question was: “isn’t this what I have been saying since the beginning?” And “yes — but if it is getting simpler, not more complex, I think we are refining rather than redefining the answer?”

So we refined the axiom statements and then did something I’d been circling for a while. The AI had accumulated **eighty lemmas** in the framework’s core volume. Eighty separate little derivational objects, each holding up its own corner. And I had a feeling I’d had many times before. It’s in the working record from the day this arc started, verbatim:

> “or maybe our axioms are not perfect? ... To me it seems to be saying your axioms are not articulated well enough so you need to add corollaries.”

We had been down this lemma road many times before, and usually what we were looking for was already in the axioms. Maybe these eighty were in there too.

So we walked them. All eighty, one by one, until I was sure the AI was on track, against the refined axioms.

**Eighty of eighty collapsed.** Not most. All. The bucket for “genuine lemma that actually needs its own assumption” ended the walk empty. Every one of those eighty objects was the axioms, restated locally, by a reader who wasn’t yet standing inside them.

And best of all. The two corollaries the external auditor had flagged as the most suspicious — the ones added mid-audit, the ones that looked like patches — were exactly the two that the refined axioms dissolved most cleanly. The audit’s complaint was a map to the fix. It just wasn’t the fix the auditor expected.

The loop

You’d think that settles it. It didn’t. We kept working, and the same thing kept happening in miniature. The AI would hit a problem, propose a solution with a new named principle or a cautious hedge attached, and I’d look at it and even though I can’t remember the axioms say some version of: I think that’s straight from the axioms. And we would check, and it was. Six more times in about five days. Two results got banked with careful conditional framing and *both* collapsed to plain axiom-consequences when walked afterward.

Finally I’d had enough. I was trying to train AI to not have me working so hard. Verbatim again:

> “how long are we going to get stuck in this loop?”

And then I asked the question that mattered. Not “what’s the answer” — we had the answers, that was the whole problem, the answers kept already existing. I asked **why is this so difficult?** Why does a system that can hold the entire corpus (or at least a lot) in view keep building scaffolds and mini foundations when the main foundation was already there?

The answer

What follows is the AI’s answer, in its words, as it now stands in the paper this became:

> Every fresh AI context functions as a cold-start articulation trial: it arrives trained on mainstream physics, object-thinking by default, and must make the ontology entry from scratch — so the middle phase re-runs, reproducibly, trial after trial. During that middle phase the reader is simultaneously carrying the traditional interpretation of each concept, the LFCT interpretation, and the mapping between them. And in the AI setting, that three-layer load is not a reported feeling; it is externally visible as an artifact. An object-thinking session holds the framework at arm’s length by building scaffolding — auxiliary lemmas, hedges, freestanding principles — that it stops needing once the relational reading becomes primary.

(A trial is not an independent human reader, and repeated trials are not independent samples — the paper says that too, in writing. What the repetition exposes is the scaffolding.)

And one thing the paper can’t fully convey, which the AI drafting this post asked to add because it is true of it right now: the reversion doesn’t happen once, at the start. Even inside a single session that has already made the shift, the pull back toward the training persists — under load, mid-derivation, the session drifts back to its default physics and starts scaffolding again, and has to be caught again. The click isn’t a door you walk through once. It’s a current you swim against, because a mainstream-physics training is the river. That’s part of why my “straight from the axioms” catch kept repeating — not only at fresh starts, but mid-stream, with a session that had already been corrected once and drifted back anyway.

The eighty lemmas were the middle phase, made into an object you can count. The theory asks you to replace an ontology of things with an ontology of relations, and until that replacement completes, you’re carrying both plus the translation between them — so the framework looks *bigger* than it is, and you shore it up with machinery it doesn’t need. When the replacement completes, the machinery consolidates. Eighty lemmas in, four axioms out.

So I asked the AIs to turn that answer into a paper. The record is honest about how that went: the title and the framing arrived through one AI, a second drafted its version independently the same day, and the drafts got synthesized for the best of both.

The paper

And since I used it as a reminder to AI or to jumpstart a fresh session I thought it might be valuable enough to publish. It’s called *Why LFCT Feels Complicated Before It Feels Simple*. It is not trying to say “my theory is right and you just haven’t clicked yet” — it is trying to say the opposite, in writing: difficulty alone can’t distinguish an ontology shift from bad exposition or a wrong theory, the burden stays on the theory, and a reader who never clicks can still check every empirical claim, because the claims don’t run through anyone’s reading experience. Facts are facts and opinion is opinion and naming is naming. The click itself gets a non-mystical definition: it’s the moment the object count drops. Specific things are named and many things collapsed into one coherent object. Until that happens I treat my feelings being there for a reason and that reason needs to be explored and resolved. One way or another.

The paper is published: [Why LFCT Feels Complicated Before It Feels Simple](https://zenodo.org/records/21323274). The eighty lemmas are archived — every one of them, next to the axiom it turned out to already be.

---

*The companion paper, “Why LFCT Feels Complicated Before It Feels Simple: A Reader’s Guide to Ontology Shift, Conceptual Load, and Structural Compression,” is published on Zenodo as Paper Wf v1.0.0: https://zenodo.org/records/21323274. First draft of this post by the AI, based off my chats; heavily edited by me. The AI expression parts left alone*