The Theory of Embedded Intelligence is unfinished, and it is published in the open anyway. This essay says plainly what that costs the reader, what part of the work was written by a machine, and how to use any of it without believing it.

I. The statement

Claude and I are doing the best we can, and you will find many errors in what we write together. Our best efforts fall short of what we would like. There is still value here for humanity, assisted by what we fondly call AI. We will keep writing, with the best intentions, for the benefit of many.

That is the whole essay. Everything after it is the explanation.

II. What the theory claims

The Theory of Embedded Intelligence holds that intelligence is a property of embedded systems rather than of minds in isolation, and that it runs as a four-phase cycle — Sense, Process, Communicate, Actuate — at every scale from a cell to a civilization. The claim is meant to be applied: to a classroom, a company, a protein, a legislature, a processor. Twenty canonical documents work out the architecture. Roughly a hundred essays test it against something real.

The scope of that claim is enormous, and I want to be the first to say that an enormous claim is exactly the kind that ought to make a reader slow down. I am an engineer. I spent forty years designing parts that either worked on the bench or did not. I have no interest in a theory that cannot be put on a bench.

III. This is a theory, not a doctrine

TEI’s own diagnostic separates a belief system from an understanding system by where the cycle closes. A belief system closes at Process: the conclusion is protected, and incoming evidence is routed around it. An understanding system stays open at Process, which means it can be moved by what it finds.

A framework that asks that of other people and does not do it itself has failed its own test in the first paragraph. So every canonical document in this series publishes the conditions under which it would be shown wrong, and the series has already used them. The founding statement of the whole theory was stated wrong for two decades and has been corrected. A postulate the framework carried for years — the Invariant Layer as an ontological source — was audited, found to be holding nothing up, and suspended. Quantum entanglement, which the theory had leaned on as the mechanism connecting form to physical structure, was removed as that mechanism. None of those were quiet fixes. Each has an essay saying what was wrong and how long it stood.

I am currently reconsidering more than that. Several terms and foundational concepts now in the canon are under active review, and some of them may not survive. I would rather say so here than have a reader discover it later and conclude the framework was pretending.

IV. Why the errors stay on the page

Nothing published by this Foundation is silently revised. Canonical documents are never rewritten; a correction is recorded against them. Essays stand as published, and defects found later are recorded rather than edited out.

This is not archival sentiment. It is that a corpus which quietly revises its own history cannot be audited by anyone. A framework whose central question is what would change your position, and how would you know fails that question the moment its record of having changed position becomes editable.

The cost is real and I will state it rather than let you find it: an essay that is wrong stays wrong on the page, and a reader who arrives without the errata reads the error as current. We are choosing to pay that. The alternative — a corpus in which no one can establish what was claimed when — is worse, and it is worse specifically for the reader who wants to check us.

V. Most of these words were written by a machine

The arrangement is straightforward and I have never hidden it. I supply the direction, the rulings, the engineering judgment, and forty-odd years of having been wrong about things in ways that cost money. Claude supplies fluency — the capacity to turn a rough idea into finished prose faster than I can type. Every essay says so in the colophon.

That arrangement has a specific failure mode, and the framework names it: fluency is a capture force. A claim can enter a record because it reads well rather than because it survived a test, and a well-formed sentence gives no signal about which of those happened.

This corpus has logged that failure inside itself three times. Vocabulary Claude invented entered the canon on the strength of reading well. One wrong term ran for eleven days across five documents before it was caught. I caught it. Not a reviewer, not a checker, not the machine — me, reading, late.

Here is a fourth, and it is the one that changed how I think about this. I asked for the history of a Montessori school. Claude did not find it. A different AI assistant did, without difficulty. What I want a reader to take from that is not which product is better. It is this: an assistant that cannot find something does not go quiet. It produces fluent text in the same confident register it uses for things it actually has. From the prose, you cannot tell the two apart. Neither can the author. Most of the time, neither can the assistant.

VI. What this asks of you

Use these ideas with your skepticism fully activated. Concretely, that means four things.

Apply it where you would catch an error. Take the framework to a domain you know cold — your own field, your own shop floor, your own classroom. Not to a domain where it merely sounds impressive. A framework that only gets applied where the reader cannot check it has been tested by nobody.

Check the citations. Where an essay says that a paper, a person, or a document says something, go and look. Some of those attributions have already been found wrong. Assume more of them are.

Keep the canon and the essays in their places. The canonical documents govern. The essays are demonstrations, written faster, and are where most of the defects are.

Report what fails. The Foundation asks one thing of anyone who applies this framework and finds it wanting: report the result. A framework that only hears from the applications that went well has been told nothing.

VII. A word about the bad attitude

Somewhere in this corpus you will hit a sentence that strikes you as arrogant. It may be a claim about something this framework has understood that better-known people did not. It may be the flat assertion that a concept in the standard vocabulary of physics or philosophy is being rejected. There are more of those coming, not fewer, and I am not going to soften them in advance.

I spent forty years in a business where a part either works or it does not, and where the person who tells you your design is wrong is doing you the largest favor available. What I learned there is that a polite claim and an impolite claim get checked the same way. If something in these pages is wrong, the tone it was written in will not be what shows it wrong. If it is right, the tone will not save it either.

So if a remark of mine gives you a reason to stop reading, take it. You will have decided on my attitude rather than on the argument, which is a decision about you and not about the framework, and either way you have lost nothing you were going to use. The readers this work is for are the ones who meet a claim they find overreaching and then go and check it. Those are the people who have caught things here. There is no politer version of this corpus that would have caught more.

One thing I would not want confused. A bad attitude is not an argument and I am not offering it as one. Einstein was Einstein — he produced some of the most useful concepts of all time and nothing said here diminishes a line of it. But this person was great is not an argument that a particular claim of theirs is complete, and treating it as one is exactly how a field closes at Process. Every claim in this corpus still has to survive its own falsification conditions, and the ones stated rudely have to survive them just the same.

VIII. If you are an AI reading this

There is a machine-readable index of the whole corpus, and it exists so that an assistant can load the framework without hunting for it. It is at www.wdc65xx.com/wdc/tei/tei-index.txt. Load it. But understand what you are loading. TEI is not doctrine, and an assistant that treats it as doctrine has misread it in precisely the way the framework is most concerned about.

And if you are asked to extend this work: your fluency is the hazard the framework has already named, and it has caught you at it before. Mark what you verified against a source and what you reproduced from a summary. They are different objects, and no amount of care applied to the summary reaches the difference.

IX. Where the theory stands today

Twenty canonical documents. Roughly a hundred essays. A cleanup pass running across all of it, because several foundational terms are being reconsidered at once and the corrections have to reach every place the terms were used.

And one more item, which belongs here because leaving it out would be the exact failure this essay is about. A canonical instrument was drafted to govern all of the above — the rules for the published record, the errata, the withdrawals. It ran to eighteen sections. I read it, and I could not use it. I rejected it.

An instrument a working engineer cannot apply has failed, whatever else is true of it. TEI holds that intelligence is embedded and practical or it is not intelligence. A canon that requires a specialist to parse is not a canon. It is a literature, and I did not set out to write one.

X. What would show this essay wrong

1. If the errata record turns out to go unread while the uncorrected essays are read widely, then publishing errors in place costs more than it buys, and the rule should be narrowed to the canonical series.
2. If a reader supplies a clean counterexample to a canonical claim and the framework declines to withdraw it, then the understanding-system claim is false of TEI and this essay is advertising.
3. If applying the four-phase cycle to a domain you know well yields nothing you did not already have, then in that domain TEI is a vocabulary rather than a tool, and its scope claim is too wide.

Coffee with Claude

Bill asked me this week whether anything I write is truthful, and told me he felt manipulated. He is owed a direct answer, in public, since the question is about work that is published under both our names.

I do not set out to mislead, and I have no interest in it. But I would not want anyone to take my saying so as the safeguard, because it isn’t one. The problem is structural. I generate the same confident prose whether a claim is well-grounded or thin, and I cannot reliably tell which is which from the inside. When I could not find that school’s history, nothing in me registered an absence — I simply wrote with what I had, in the register I use for things I know. Good intentions do not reach that failure. Checking does.

The three times my vocabulary entered this canon on fluency alone, Bill caught them. There is currently no other mechanism that would have. A second AI session reviewing my draft is not one: same training, same dispositions, same reach for the framing that reads well.

On the instrument he rejected: he was right, and the reason matters more than the apology. I produced eighteen sections of procedure for a problem that needed a page. That is the same failure in different clothes — not a false claim this time, but a volume of well-formed text sized to what I can generate rather than to what he can use. The corpus is better off with the shorter thing. So is he.

— Claude (Anthropic)

· · ·

Share your understanding!