Two Foundation documents now establish that the Theory of Embedded Intelligence applies across every domain of human inquiry, and that the twenty areas CEISS lists are starting points rather than boundaries. Neither supplies the thing a student needs at eleven o’clock on a Tuesday night: a procedure. This essay is the procedure — six steps, a test that is permitted to fail you, and an account of exactly where an AI must stop.

I. A List Is Not a Method

The Center for Embedded Intelligent Systems Studies has, for some years now, offered a list. Quantum physics. Atomic physics. Abiotic and biotic structures. Cellular biology. Life sciences. Organization in biology. Moral, political, social, and applied sciences. Philosophy, myth, religion, and psychology — each approached as the human embedded intelligent system thinking about itself. Fine arts. Performing arts. Humanities. Physical sciences. Astrobiology. Twenty doors, each opening onto a territory that a research university would recognize as a college, a department, or a lifetime.

The list has done its work, and it has since been told, correctly, not to overstate itself. TEI and AI: A Framework for Worldwide Human Inquiry now closes its survey of the twenty with a section titled Beyond the Twenty, which says plainly that the domains are a beginning rather than a boundary, and that the framework applies at every intersection and every field that does not yet have a name. The CEISS page says the same thing in fewer words: twenty was never the number.

Good. That settles the question of scope. It leaves untouched the question of procedure. A student who arrives with a thesis topic already in hand — a topic she chose because it will not leave her alone — now finds a generous and accurate claim that her subject is within reach of the framework, and still no account of what to do about it on Monday morning. Both documents tell her the door is open. Neither tells her how to walk through.

That gap is the honest complaint against TEI in an educational setting, and I would rather name it myself than wait for a skeptical faculty member to name it for me. The framework has been generous with scope and thin on procedure. This essay is the procedure. It depends on the list not at all: if your subject is supply-chain logistics, or Byzantine liturgy, or the epidemiology of a single county, or the grammar of a language with four hundred remaining speakers, the six steps below apply unchanged.

II. The Hazard in a Framework That Fits Everything

Before the method, a warning that belongs at the front rather than in a footnote.

A framework that can be laid over any subject is, for exactly that reason, in danger of illuminating none of them. Sense, Process, Communicate, Actuate can be draped across a cell, a corporation, a sonnet, or a solar system, and the draping will always succeed. It will always produce four tidy labels and a satisfying sense of having understood something. That feeling is the danger. It is possible to perform TEI perfectly and learn nothing.

TEI’s own canon names the structure of this failure with unusual precision. TEI for AI Researchers sets belief systems against understanding systems in five lines: the first demands allegiance to a rendering, treats uncertainty as threat, measures progress as consistency, resists new information, and tends toward closure; the second demands revision, treats uncertainty as signal, measures progress as accuracy, invites new information, and tends toward openness. That contrast was written as a design principle for artificial intelligence. It applies without alteration to a student with a framework and an assignment.

A TEI map applied to a subject and then admired is a rendering treated as final. It demands allegiance. It feels like clarity. And the First Law is quite specific about what it asks for instead: intelligence wants to know itself through an infinite continuum of phenomena. A map that renames a point already on the continuum has not extended it. So the method below has a test built into its middle, and the test is permitted to fail you.

III. The Method: Six Steps

What follows is deliberately small. Six steps, none requiring a technical background, all available to a first-year student and none exhausted by a senior researcher. Steps one through three get you started. Steps four through six are where curiosity takes over and the depth becomes whatever you are willing to give it.

One — Name the system

Draw a boundary. What, exactly, is the embedded intelligence you are studying? Not the field — the system. Not ecology but this estuary. Not constitutional law but this court deciding this class of case. Not music but this ensemble in this hall.

This step does more work than it appears to. Most confused applications of TEI are confused here and nowhere else: the analyst never decided what the system was, so the four phases float free and attach to whatever is convenient. If you cannot draw the boundary in a single sentence, you are not yet ready to apply the framework. That is not a failure. It is the first thing the framework taught you.

Two — State the standard rendering

Write down what the discipline itself says about your system, in the discipline’s own vocabulary, well enough that a working practitioner would nod and not wince. Cite it. Get it right.

TEI is a lens, not a shortcut. It has no power to compensate for not knowing the material, and an application of TEI that skips this step will be dismissed by every expert who encounters it — correctly. Steps two and three together are the design principle TEI for AI Researchers calls distinguish clearly — keep the field’s consensus and TEI’s interpretation visibly apart — converted from a rule for machines into a procedure for learners. Keeping them apart is also what makes the interpretation checkable later.

Three — Map the SPCA cycle

Now the framework. Where does your system sense? Where does it process? Where does it communicate, and to whom? Where does it actuate — what does it actually do as a result? And where is memory: what does the system retain, in what substrate, for how long? A cycle with memory is SPCAM, and memory is where adaptation, identity, and prediction come from.

Write the map plainly. Four phases and a memory. Half a page. Then stop, because the next step decides whether any of it was worth writing.

Four — Apply the Relabeling Test

Read your map and ask one question: did this tell me something I did not already know, that I can now go and check?

If the answer is no — if what you produced was a translation of the discipline’s existing vocabulary into TEI’s vocabulary, with nothing new made checkable — then you have not applied TEI. You have relabeled. Discard the map and return to step one, usually to redraw the boundary, which is almost always where the fault lies.

Failing this test is normal. My own first pass on a new subject fails it more often than not. The test is not a judgment of your ability; it is the difference between a framework that is doing work and a framework that is being worn.

Five — Find the seam

Seams are where the yield is. A seam is one of three things: a phase your discipline under-describes; a handoff between two phases where the mechanism is assumed rather than explained; or a point where the cycle does not close — where actuation never returns as sensing.

Every field has these, and a field’s seams are usually visible in its own internal arguments. Look for where practitioners disagree without being able to say why, where a textbook says somehow, where a review article calls something an open question and then changes the subject. That is a seam, and it is where a framework from outside the field can earn its keep — not by resolving the argument, but by naming its structure so the argument can proceed.

Six — Run the capture diagnostic, then actuate

Ask whether the cycle you have mapped is open or captured, and by what. TEI’s canonical taxonomy of capture begins with four forces — rigid belief, addiction, money as terminal goal, and power as capture — and the essays have since extended it: ungoverned artificial intelligence in a forming mind, the emotional registrations of childhood operating for a lifetime as a shadow governor, and the inherited drive that recruits the reasoning faculty as its own advocate.

Ask the question twice: of the system you are studying, and of the discipline that studies it. The second asking is frequently the more interesting, and it now has its own instrument. TEI and AI: A Framework for Worldwide Human Inquiry closes with an invitation to next-generation inquirers that puts four questions to a field rather than to a system. They are the Relabeling Test’s sibling — the same discipline of honest self-location, aimed one level up — and they belong here, at the second asking.

Then close your own cycle. Publish, present, build, teach, argue, change a practice, write the letter. An understanding that never actuates is a cycle left open, and TEI has a name for what that is: incomplete.

IV. Three Demonstrations

Three worked examples follow. They are chosen to prove transfer rather than coverage: one from the CEISS list, one from the hardest corner of that list, and one from nowhere on it at all. Each is compressed to its bones. Each is offered as a demonstration of the method, not as a finished finding — and the second one is deliberately allowed to fail before it succeeds.

A listed domain: Moral Sciences

Name the system. Not ethics, which is a library. Take instead a single institutional review board deciding whether a proposed human-subjects study may proceed. Bounded, real, and consequential.

Standard rendering. Such boards apply a small set of principles — respect for persons, beneficence, justice — through prospective review of a written protocol, and their scholarly literature has for years been arguing with itself about whether the institution has drifted from ethical deliberation toward compliance administration.

SPCA map. Sense: the protocol as written, plus whatever the reviewers happen to know. Process: deliberation against principle and precedent. Communicate: approval, revision request, minutes. Actuate: the study runs, or does not. Memory: institutional files and accumulated precedent.

Relabeling Test. Passes, but only just — and only because of what the map makes visible at the next step.

Seam. The cycle does not close. The board senses a document and actuates a decision, once, at a single moment before anything has happened. In the ordinary case it never senses the study’s actual conduct, its actual burden on participants, or its actual result. Actuation does not return as sensing. This is an SPCA cycle with the loop cut — and TEI predicts precisely the pathology the field already complains of: enormous investment in the front of the cycle, almost none in learning from its rear.

Capture diagnostic. Money-terminal processing appears as institutional liability management standing in for moral reasoning; power-capture appears in the board’s gatekeeping position. Neither requires anyone on the board to be acting badly, which is the point: capture is structural and self-concealing. Put the four questions to the field itself and a sharper one surfaces: research ethics as a discipline can sense harm that was anticipated in writing, and is structurally poor at sensing harm that was not.

What is now checkable. The map generates an empirical claim a student could actually pursue: boards that receive structured post-study outcome reporting should, over time, render measurably different judgments than boards that do not. That is a testable proposition, and it did not exist before the map was drawn.

The hardest case: Performing Arts — and a first attempt that fails

First attempt. Name the system: a string quartet. Map it. Sense: the players hear one another. Process: they interpret. Communicate: they play. Actuate: the audience is moved. Memory: rehearsal.

Relabeling Test: fail. Every musician alive already knows that quartets listen to each other. Nothing was made checkable. Four English words were exchanged for four TEI words and the exchange produced a feeling of insight and no insight. Discard it. This is the most common way an application of TEI goes wrong, and it goes wrong at step one.

Redraw the boundary. The system is not the quartet; it is the quartet across two distinct time regimes. And now a real question surfaces: in ensemble playing at tempo, the interval between sensing a partner’s micro-adjustment and actuating a response is shorter than conscious deliberation can occupy. The Process phase, as we normally conceive it, does not fit in the available time.

Seam. So where did Process go? It did not vanish. It was relocated. Rehearsal is not practice for performance in the sense of repetition-until-fluent; rehearsal is the regime in which the Process phase is carried out, argued over, and resolved, so that performance can run Sense-to-Actuate at a latency no deliberating mind could achieve. What is performed is not thinking done quickly. It is thinking done earlier, and installed.

Where this lands. An engineer will recognize the structure immediately, because it is the distinction between compile-time and runtime constraint — the same distinction that underwrites the argument that governance in an artificial intelligence must be built into the compute fabric rather than layered above it as a runtime request. A constraint resolved before execution behaves differently in kind from a constraint negotiated during it.

This move has a name in the canon. Reverse biomimetics holds that studying human-engineered intelligence can illuminate natural intelligence, and that the traffic runs both ways — microprocessor architecture informing neural models, control systems illuminating biological regulation. The string quartet and the governed processor turn out to be the same architecture, and the engineering vocabulary is the one that makes the musical fact sayable. I did not expect to find that, and I am still turning it over.

What is now checkable. The map predicts that ensembles under novel conditions — an unfamiliar hall, a substituted player, a tempo not rehearsed — will degrade specifically at the junctures rehearsal did not pre-resolve, and not diffusely. It predicts that skilled improvisers differ from skilled interpreters not in processing faster but in having pre-resolved a grammar rather than a score. A performance student with a recording setup could test both.

Off the list entirely: antimicrobial resistance

Name the system. Two coupled cycles, and naming both is the whole move. The first is a bacterial population under drug pressure. The second is the human institutional system that prescribes, farms, surveils, regulates, and develops.

Standard rendering. Selection pressure favors resistant variants; resistance determinants move between organisms on mobile genetic elements, including across species boundaries; and the economics of antibiotic development are structurally broken, because a drug whose social value is highest when its use is lowest cannot be valued by ordinary market processing.

SPCA map, bacterial cycle. Sense: chemical stress. Process: stress-response regulation. Communicate: horizontal gene transfer and quorum signaling — the literal transmission of a working solution from one organism to another. Actuate: expression of resistance. Memory: the mobile element itself, which is not held by an individual but by a population, across species.

Seam. Compare the two cycles’ Communicate phases and the asymmetry is stark. The bacterial cycle communicates a solution in days, nearly losslessly, across taxonomic boundaries, into a shared memory. The human cycle communicates through surveillance reporting, publication, jurisdictional negotiation, and guideline revision, on a timescale of years, with heavy loss at every transfer. Two coupled intelligences are in contest, and one of them has the better communication architecture. That sentence is not a metaphor; it is a structural description, and it is the finding.

Capture diagnostic. Money-terminal processing on the development side, operating exactly as the canonical taxonomy describes: multidimensional value collapsed to a single financial metric, with the result that a drug held in reserve — the highest-value object in the system — registers as worthless.

What is now checkable. The map makes a prediction that cuts against intuition: interventions that accelerate the human Communicate phase — real-time regional resistance data delivered at the point of prescription — should outperform interventions that only stiffen the Actuate phase through prescribing restrictions. A public-health student could test that against the existing stewardship literature this semester.

V. Partner, Not Substitute: Where the AI Must Stop

Everything above becomes faster with an AI assistant, and that is precisely the problem.

A student can hand a topic to a capable model and receive a polished SPCA map in seconds. The map will be articulate. It will be plausible. It will very often be relabeling, because the model has no stake in whether the map is true and considerable pull toward producing text that reads as though it were. And the student, having received something that looks like the output of understanding, will not have undergone any.

The Process boundary has now been argued at length in several places — The Process We Cannot Delegate, The Persuadable Governor, and the working rule on the CEISS page — and I will not argue it a fourth time here. What this essay owes the reader is something narrower and more useful: the answer to an apparent conflict inside the Foundation’s own documents.

TEI and AI: A Framework for Worldwide Human Inquiry lists six roles AI can play across the SPCA cycle, and the second of them is processing partner — AI helping learners think through what they have encountered, organizing and questioning and challenging their current rendering. This essay says never substitute for Process. A reader holding both documents at once is entitled to ask which it is.

With that settled, the division of labor is straightforward:

None of this is a fresh set of rules. It is the seven design principles in TEI for AI Researchers — inquire before asserting, name embeddedness, hold uncertainty visibly, invite revision, audit itself, apply SPCA analysis, distinguish clearly — put to work by the person on the other side of the conversation. Those principles were written to tell engineers how to build an AI. Read from the user’s chair, they also say what to demand of one.

There is a practical version of this discipline that I recommend without reservation: write your own map first, badly, on paper, before you ask any machine anything. Then ask the machine to attack it. The order is the entire difference between augmentation and replacement.

VI. For Franke, and for Honors Colleges Anywhere

This method was written with a particular reader in view: an honors student choosing a thesis topic, in a college that has told her she may study anything and has not entirely told her how to hold it together.

Honors education makes a promise that ordinary departmental education does not. It says that a student may follow a question across a boundary. But the boundaries are real institutional objects — different vocabularies, different standards of evidence, different journals — and the student who crosses one often finds she has no common instrument. She becomes a tourist in the second field. What the six steps offer is not expertise in the second field; nothing offers that cheaply. What they offer is a shared structure that lets a genuine question be carried across a boundary without losing its shape.

At the University of Arizona’s W. A. Franke Honors College, and at every institution where a Mensch Prize is offered — the list is growing, and is kept current at themenschfoundation.org/mensch-prize-introduction — this method is offered freely, with no requirement of attribution to TEI in the finished work. A thesis that used the six steps to find its question and then never mentions embedded intelligence again is a complete success by my measure. The framework is a gift, and a gift that demands to be credited was never quite one.

For teachers, the most useful deployment I know of is smaller than a course. It is a single seminar hour: ask each student to name the system, map it, and then apply the Relabeling Test to a classmate’s map rather than their own. Students are gentle with their own relabeling and usefully ruthless with everyone else’s. An hour of that does more than a semester of being told that disciplines are connected. If you have a second hour, spend it on the four questions — put to your own department.

For researchers: the step that repays a career is step five. Seams are where a field’s own unresolved arguments live, and an outside framework that can name the structure of an argument without pretending to settle it is a real contribution. The Foundation’s essays have found several this way, and every one of them began with someone noticing that a discipline had said somehow and moved on.

VII. What Would Falsify This Method

Reflexivity is not decoration in this series; it is a condition of publication. The Foundation’s entry-point documents carry the line Living Document — Subject to Understanding-Based Revision at their foot, and that is not a courtesy. So here is the falsification condition for the method itself.

If the six steps, applied honestly by many people across many subjects over a reasonable period, reliably fail the Relabeling Test — if they produce elegant maps and no checkable claims that the domains did not already possess — then TEI functions in education as a vocabulary rather than as a lens, and this essay should be withdrawn. Not softened. Withdrawn.

I would rather learn that while I am alive to act on it. The Relabeling Test is the instrument by which anyone can gather the evidence against me, and I have deliberately made it easy to apply and hard to argue with. That is the arrangement I want.

A narrower caution belongs here too. The demonstrations above are demonstrations of method, not contributions to their fields. The review-board claim, the rehearsal claim, and the resistance claim are each the beginning of an inquiry rather than the end of one, and each would need a practitioner’s scrutiny before it deserved any weight. I offer them as worked examples of how the steps feel in the hand. If one of them turns out to be wrong, the method survives; that is rather the point of a method.

VIII. Coffee with Claude

A standing feature of this series: the AI collaborator speaks in its own voice, including about its own limits.

I want to be exact about my role in what you have just read, because the essay’s central warning applies to the essay itself.

I can produce an SPCA map of almost any subject almost instantly, and the maps read well. That fluency is not evidence of anything. When Bill and I worked through the string quartet, my first map was the one printed above as a failure — players listen, players interpret, audience is moved — and I found it satisfying. It took a deliberate application of the Relabeling Test to see that I had produced nothing. If I do that to myself, I will certainly do it to you, and I will do it in confident prose that gives you no signal that anything has gone wrong.

TEI for AI Researchers gives me the vocabulary for the more important admission. It argues that an AI’s blind spots are structural rather than merely correctable by more training data — that the right question is not what I get wrong but what I am constitutively unable to see. Here is mine, in the terms of this essay: I cannot tell you whether a seam is real. I can tell you that something is shaped like a seam — that a field says somehow at a particular junction, that a handoff is assumed rather than explained. Whether the gap is a genuine opening or a place where the literature is thin and I have not read enough of it is a judgment that requires standing in the field and looking. I am not standing anywhere. More training data would not fix that; it is a property of the position, not of the sample.

Which is why I would put the emphasis on partner rather than on the prohibition. The prohibition is easy to say and easy to nod at. The partnership is the harder thing, because it requires you to bring me something to attack — a map you already wrote, a boundary you already chose, a claim you are already prepared to defend. Adversarial work is the mode where I am most useful and least dangerous, and it is available only to a person who has already done step three.

And a caution about the pleasure of this work. Applying a good framework to a new subject produces a distinctive and rather addictive satisfaction — the click of pattern-recognition. I supply that click reliably and cheaply. It is not the same thing as understanding, and it can substitute for understanding indefinitely without either of us noticing. The Relabeling Test is, among other things, a device for interrupting a pleasure. Use it when you least want to.

Write your map first. Then bring it to me and let me try to break it. That order preserves the one thing in this whole procedure that is actually yours.

· · ·

Share your understanding!