# The Drop

## Summary

A Venus flytrap's state does not work. What happens? **It is eliminated** — the whole individual at once, no verdict issued, no one present to receive one.

A dog gets the rabbit wrong. What happens? **Its behaviour is corrected** — the rabbit comes out, the digging turns up nothing, it changes. The individual survives; the classification is amended.

A person has the word *fish* wrong. What happens? **The question has no answer to give.** A definition set out in symbols is not the kind of thing the world can find true or false — Linnaeus moved the whales from the fishes to the mammals between two editions of one book, and the whales did not lose a gram. **The world settles the facts; people settle how the word is to be used.**

Those three lines are the spine of this essay. What they separate are three kinds of thing left behind: **a state** (the flytrap, where nothing is kept at all), **a boundary** (the dog and the pigeon, a line drawn by how the world is distributed), and **a ruling** (the human case, a category defined in abstract symbols). What separates them are two checkable criteria: **is the item still there once input is removed, and what happens when it is wrong.**

The third is where the weight falls. Since the world cannot correct it, correction has to come from somewhere else — someone who proposes, someone who challenges, and disputes that do not end. And whether a definition actually carries weight anywhere has one test that words cannot fake: **when it is overturned, does what was built on it move?**

From this the essay takes a reading of *the self*, and states plainly that the reading is chosen rather than neutral: a position is required, it gets occupied, and **what that position looks like when occupied for a long time is what we call a self**.

**And there the line stops.** It can test whether the position is occupied; it cannot test whether anything besides that occupation is going on. So on consciousness the essay sets down one conjecture at the close, and marks it as one on the spot:

> **The position so filled may be where the felt thing comes from.**

No evidence is offered for that sentence and no argument has been made for it. It is written down only because it is why this line was followed at all — **and a conjecture marked as a conjecture does not contaminate the argument that precedes it; a hidden one does.**

(Section I begins from Aristotle's ἐνέργεια and κίνησις — the drop was first seen between those two words. But that vocabulary is exchanged within a page; the sections that follow do not work inside it, and do not depend on whether it holds up.)

### A note on how this was written

The conception, the judgments and the choices in this essay are the author's. **The literature search and verification, the structuring, and the drafting were done in collaboration with Anthropic's Claude.** Every claim that has not been verified — second-hand citations, empty searches, and the essay's own conjectures — is listed item by item in the *Unchecked* section at the end.

---

## I. From Tense to Object

### Two Greek words

**ἐνέργεια** is formed from the adjective ἐνεργός / ἐνεργής — *at work, active* — with a noun suffix; and that adjective itself comes apart into ἐν (*in*) and ἔργον (*work, deed*). So the word says, literally, **the condition of being in the midst of work**.

**κίνησις** is formed from the verb κινέω (*I set in motion*) with the suffix -σις. κινέω descends from Proto-Indo-European \*keyh₂-, meaning **to start to move**; Latin *cieō* is its cognate; and its opposite in Greek is στάσις, standing still.

The two words are therefore weighted differently. One rests on **the working itself**; the other on **setting out from rest**. And wherever there is a setting out, there is somewhere it sets out toward — **an endpoint is already packed into the root.**

### Aristotle's two concepts

The *Physics* defines change as **the actuality of what is potential, in so far as it is potential** (3.2, 201b4–5). And in the same place it adds that this is an **incomplete** actuality: what is actualised in a process of realisation is only an incomplete potentiality (201b31–33).

Book Θ of the *Metaphysics* supplies the other half. A δύναμις in the strictest sense is "a starting-point of change in another thing, or in itself in so far as it is other" (Θ.1, 1046a12) — **and the exercise of such a power is a κίνησις**. The builder's craft is a power; the exercise of it is the process of building a house.

So the relation between the two is not that of two coordinate kinds of thing. A κίνησις is **an actuality that is still owing** — not fully actual, because the potentiality it actualises has *not yet* built into its own condition.

### The test

Chapter 6 of Book Θ supplies a grammatical test, and it remains the form in which this distinction is most often cited. **One sees and has seen at once; one thinks and has thought at once** — the progressive and the perfect are true together. But **one builds and has built** are mutually exclusive: the moment the house stands, the building has stopped.

The point is not that these are different verbs. It is **whether the two tenses can be true at the same time**. If they can, the thing is complete at every moment of its going on. If they cannot, it is at every moment owing an endpoint that has not yet arrived.

### One thing that must be said

The textual standing of that test is disputed. Burnyeat devoted a paper to it whose title is already the conclusion: "Kinēsis vs. Energeia: A Much-Read Passage in (But Not of) Aristotle's *Metaphysics*" (*Oxford Studies in Ancient Philosophy*, 2008). Before him, Ackrill had raised the classic objections to the distinction itself.

So although the distinction is standard, it is not a stone one can simply build on. **What follows uses not its authority but the structure it points at.**

### Where the drop comes from

Set those pieces side by side and the difference turns out not to be about what is being done. It is about **how time is being carried**.

A κίνησις, at every moment of its going on, is owing an endpoint that has not yet arrived: interrupt it and what is left is a fragment; you can ask how far along it is; you can ask how fast. An ἐνέργεια owes nothing at any moment: interrupt it and no wreckage remains; asking how far along it is has nothing to attach to, and neither does asking how fast.

**And the drop is the difference in level between the two, when something that owes nothing has to be carried out by something that owes.**

That sentence is where this line starts, not where it ends.

### Putting it another way

That vocabulary has a drawback: it locates the difference in **kinds of activity**, as though the world contained two sorts of goings-on, one complete and one not. And once you ask what entitles you to sort a given case into one of them, there is nothing to check except the tense test itself.

The difference is not on the side of the activity at all. It is in **what the activity is directed at**.

Building is directed at a house — a thing out there, put up block by block, not yet present. Walking is directed at a place — also out there, also not yet reached. **So long as the thing directed at is still outside and still coming into being, the activity owes; and what it owes is exactly what its object owes.**

And what is seeing directed at? At the visible shape present just now. What is thinking directed at? At a concept. **Neither of these is out there waiting to be built. They are set.**

(*Set* does not mean *chosen at will*. The dog did not select that rabbit, and no one voted on what *red* is. Here the word marks one thing only: **it was not extracted from the input.** What shapes it, and whether it can be altered, are matters for sections II and IV.)

Once the word *set* is in play, the next question follows at once: **what keeps the set thing in place?** For outside it, everything is changing without pause — the light changes, the scent changes, one turn of the body replaces colour, outline and profile wholesale. And the set thing does not change with them: still that one rabbit, still that one kind, still that one procedure.

**The drop is here**: outside, ceaseless change; the set thing held still against it — and the holding takes effort, can fail, and can be asked after.

This second way of putting it has an advantage: **it changes what there is to check.** The question was *which kind of activity is this* — and short of memorising the tense test there was no way to check it. The question now is — **what kind of thing has been set, whether it is still there once input is removed, and what happens when it is wrong.** And each of those three, applied to a concrete case, yields a reading.

So what follows does not work inside ἐνέργεια and κίνησις. The reason for beginning there is that the drop was first seen between those two words; and once seen, it can be said again in plainer terms — **as something between what is set and the flow around it.**

## II. Three Words

Three examples; three different things left behind. And what separates them is **how far, and by what means, each holds against the flow around it.**

### A state

Touch one of the trigger hairs inside a Venus flytrap and nothing happens. Touch a second within about twenty seconds and the trap closes. Only further mechanical stimulation, together with nitrogen compounds released by the prey, starts secretion and digestion.

It plainly remembers the first touch. But that remembering is missing three things.

**No distinction between what carries and what is carried.** The internal state is not *about* the fly. It is the electrical and calcium signal itself, sitting in the middle of a causal chain, and its entire content is its causal role.

**No identity across variation.** One fly stepping twice, two flies stepping once each — the internal state is identical. Nothing in there says *that one again*.

**Not capable of error.** Heat from a forest fire will close the trap, and we are inclined to call this a misjudgment. It is not. It is a different cause producing the same effect. Misjudgment requires something that claimed something and claimed it wrongly, and here nothing has been claimed.

So the twenty seconds are not a memory of an event. They are a decaying quantity. **Decay is not forgetting — forgetting requires that something was held.**

One more thing to record here, for use later: **it is not adjudicated either.** What the world does to it is eliminate it — lineages whose state-summaries do not work die out. But elimination removes the whole holder at once, leaves no verdict, and has no one present to receive one. **Being unadjudicated and being eliminated are two different things.**

### A boundary

The second layer has two achievements. They are usually discussed apart. They belong together.

**First, the identity of an individual.** The rabbit goes behind a rock and comes out the other side; the chase cannot break off. It is smelled first and seen second, the two channels arriving at different times, and both must be assigned to the same rabbit. It turns, and the whole profile of colour, outline and scent is replaced — still that one.

These force a single conclusion. There must be something in the dog that is not identical with any current set of sensory features, and that survives their wholesale replacement.

And once there is such a thing, three consequences arrive together. **It can be wrong** — it can fail to match what it is about; the dog digs in the wrong place. **It can be lost** — it can fade, be crowded out, be lost track of; and this differs from decay, since decay is a quantity going to zero while losing track is something once held no longer being held. **It comes already pointed** — the item is not neutral; it is to-be-chased or to-be-avoided. The dog's rabbit is never a neutral rabbit.

(This too is a drawing of boundaries, only the boundary runs around one thing rather than between two kinds: what belongs to this one and what does not, drawn across viewpoints and across time. Hence *boundary* as the common name for both achievements of this layer.)

**Second, the boundary of a category.** This one belongs on this layer too, and the evidence is hard.

Build the experiment: artificial categories whose exemplars vary widely, some sharing no features at all, held together only by a network of correlated features, with no single feature necessary and all features overlapping between the categories. **Pigeons learn it easily**, and respond to entirely novel items with a gradient that tracks degree of family resemblance — the more typical the item, the stronger the response.

**Family resemblance is not a layer.** It is a **structure** — overlapping features, no necessary feature, a typical centre and ragged edges. And that structure is found on both layers. A great many human concepts are family-resemblance concepts; the term itself comes from Wittgenstein on *games*, and his point is precisely that **human** concepts are mostly like this. (The study cited in this section is itself titled "Family resemblances," and it is a study of human categories.)

So what this evidence establishes is narrower: **a boundary of this kind can be drawn by a circuit, with no version set out anywhere.**

What makes these two achievements one layer is a property they share: **both are drawn by how the world is distributed.** Individual identity is drawn by continuity in space and time; category boundaries are drawn by which features correlate and co-occur. Both can be adjudicated — whether the line is drawn well is settled by consequences.

### A distinction that has to be set at the right width

What the pigeon evidence supports is a claim about evidence: **nothing in the pigeon's behaviour requires positing a version that can be produced and challenged.** It supports no claim about capacity.

This width holds from here on. **Both criteria ask about a particular use, not about a system** — what is kept on this occasion, what is adjudicated on this occasion. A system may draw a boundary in one place and do something else in another, with no conflict; the next section shows an animal maintaining a category across a stimulus-free delay.

### A ruling

What is distinctively human is not **having** categories.

> **What is distinctively human is using symbols to define categories.**

Each of the three words in that sentence carries a limit.

**Symbols need not be language.** A notch, a knotted cord, a boundary stone: these are symbols too — and they raise *this pile counts, that one does not* perfectly well. Section V takes this head on: **define the third layer as *having language* and the two criteria go idle**, since one would simply ask whether it has language.

**Defining need not be answered.** Most concepts have no definition — *game* has none, *justice* has none — and they can still be discussed, negated, argued over. Nor need it be demanded of you: defining is often done on one's own initiative, and nobody made Linnaeus move the whales into the mammals. And when someone raises the demand and someone else argues that it should not be met that way — which is exactly what Wittgenstein does with *games* — **refusing to define is itself a move made on this layer.**

**Using is not taking on.** Producing a definition and taking one on are two different things — that is section VII's business: the machine there uses symbols to define categories, more fluently than anyone, and nothing is built on what it produces. **So this sentence marks out what the layer does, not where a definition carries weight.**

Which means *distinctively human* has to be discounted accordingly: **the machine defines categories in symbols too.** Human beings are how this layer entered the world; but once the four things are supplied, doing it is not confined to them. That is exactly why the two criteria are written to be indifferent to the vehicle — **otherwise section VII could not be asked at all.**

The three limits together narrow the layer to something smaller: **the question *by what does it count* can be raised here.** Raising it and answering it yourself is defining; someone else raising it and your having to answer is being challenged; someone raising it and your arguing that it cannot be answered so is refusal. **All three are moves on this layer, and nothing in a pigeon's behaviour requires positing such a raising — there is nothing there to raise.**

Mathematics is where the question gets fully answered: the wall is built, and imagined variation runs straight into it. Elsewhere it is raised, partly answered, or left standing — **and that it can be raised at all is what makes the layer.**

From this the properties of the layer follow without any empirical claim about human beings. A ruling set out in symbols can be discussed, negated, made conditional, made the subject of a further claim, and met with **what entitles you to that one?** None of which a boundary drawn by a distribution can do — **it has no version that can be produced and challenged.**

### Why a third layer is needed at all

One asymmetry shows why this layer must exist.

With mathematical objects, one example plus imagined variation suffices. You do not need to survey every triangle to know a triangle cannot have four sides. The variation runs immediately into a wall.

With natural kinds there is no such wall. Which step of variation still counts as a horse? Nothing stops you.

The difference is that the boundary of a triangle is given by **definition** — **the wall is there because someone built it.** And where there is no wall is where a wall has to be built.

**Building the wall is defining the category in symbols.** And where the wall cannot be built — *game*, *justice* — the question is raised all the same, and argued over all the same; it simply goes unanswered. **What marks this layer is that the question can be raised, not that it gets answered.**

## III. Two Criteria

The previous section gave three things their names but not the means of sorting a given case into one of them. This section supplies that.

The two criteria ask two things: **whether anything is kept**, and **what governs what is kept**.

### Criterion one: with input removed, is the item still there

A thermostat: no. Remove the temperature and nothing remains. The flytrap: no — those twenty seconds are a quantity decaying on its own, not something kept. The dog: **yes.** The rabbit goes into the burrow, input falls to zero, and the dog waits at the entrance for two hours.

This separates state from boundary, and does so better than "mechanism versus non-mechanism." **The dog's side is mechanism throughout** — the object file is circuitry, cross-modal binding is circuitry, maintenance during occlusion is circuitry. The cut is not between mechanism and non-mechanism. It is between **something kept and nothing kept**.

Waiting at a burrow is the cleanest specimen: zero input, maintenance entirely internal — coming free of the stimulus, and come as free as it goes.

### Criterion two: what happens when it is wrong

**Flytrap: an individual whose state does not work is eliminated.** No verdict is issued and no one is present to receive one. The whole holder goes at once.

(*Does not work* is ascribed from outside. The previous section said the flytrap cannot be wrong, since it has made no claim. **Being eliminated does not require being wrong; it requires only not working.**)

**Dog: an individual that stores the wrong classification has its behaviour corrected.** The verdict is delivered on the spot — the rabbit comes out, the digging turns up nothing, the behaviour changes. The individual survives; the classification is amended.

The classification here is a real one, held in its own right: in delayed match-to-category tasks an animal maintains a category across a stimulus-free delay. **Holding a category is not what is distinctive about the third layer.**

**Human: a definition laid on a symbol is something the world cannot correct.** It can tell you whether a given thing falls under the definition once the definition is fixed, and whether working with that definition goes well. **It cannot tell you the definition is false, because a definition is not the kind of thing that is true or false of anything.**

> **The world can settle whether a boundary is well drawn. It cannot settle how a symbol is to be used.**

So correction has to come from somewhere else — and this is why the third layer always has a crowd around it: **someone who proposes, someone who challenges, those who agree and those who object, and disputes that do not end.** They do not end because no one involved is clever enough; they do not end because **there is nothing one can walk over and look at more closely.**

### One objection, taken first

Someone will say: definitions get adjudicated too — classification by descent won, because it explains more and predicts better.

True. But notice what is being adjudicated: **the consequences of a ruling**, and whether consequences are good presupposes that someone has already settled what counts as good. **A ruling can be pushed along by its consequences, but being pushed is not being settled** — something pushed along still requires someone, at each step, to decide whether to go along.

### Three cases

Three redefinitions. Not one of them is "the old boundary was drawn wrong."

**Atom.** ἄτομος meant uncuttable. Dalton applied the word to the units he had found, and the standard verdict in the history of science is replacement rather than revision — **and that verdict is a decision the community made, not a fact looked up in the world.**

**Planet.** In August 2006 the International Astronomical Union redefined the term and Pluto fell outside it. **Pluto did not lose a gram or move an inch.** Everything that changed was on the concept side, and what decided to go along was a vote.

**Whale.** This one shows the most, because it is the only one of the three with new facts entering. Whales bear live young, nurse, breathe air, have skeletons homologous with those of tetrapods. All of that is given by the world and beyond dispute. But between those facts and *therefore a whale is not a fish* there is a step the world did not supply: **who says the word *fish* is to be used by descent rather than by habitat and form?** Before the reclassification, "fish" *was* used by habitat and form, and by that use a whale is a fish — used correctly. What changed was how the word was to be used.

And the swap happened **inside one man's one book**. The first edition of *Systema Naturae* (1735) placed whales among the fishes, following Artedi, called the father of ichthyology; the tenth (1758) moved them to the mammals. Same author, same work, same animals, different classification.

> **The world settled the facts. People settled how the word was to be used.**

The "fish" side makes the point harder still. After the reclassification, "fish" itself is judged **paraphyletic** in modern phylogenetics — under strict cladistic usage the word would include us. And eliminating paraphyletic taxa has been a **goal** of modern taxonomy for some fifty years; goals are set, not looked up. The dispute is settled by no experiment to this day.

And there is a further layer, the most important: **the reclassification presupposes identity.** We say *the whale was reclassified from fish to mammal.* For that sentence to hold, "whale" must be the same thing before and after — otherwise it is not a reclassification but a change of subject. No experiment settled that. It was assumed. **What was adjudicated is which class the whale belongs to. What was not adjudicated is that the whale is still the whale.**

(What drifts here is the attachment of a word across history, not an essence changing. That an essence is grasped whole in one go, and that the word carrying it came over centuries to attach to something else, are claims about two different time-frames. They do not conflict.)

### A ruler falls out of this

The closing line of the whale case — **what was adjudicated is which class the whale belongs to; what was not is that the whale is still the whale** — brings something to the front: **"still the same one" does not say the same thing in every place.**

Use it as a ruler. Ask: here, what does *that one again* mean? Four rungs.

**No meaning.** Only a quantity decaying on its own; nothing kept. The flytrap. One fly stepping twice and two flies stepping once each leave it in an identical state — nothing in there says *that one again*.

**That individual again.** Numerically the same particular. The rabbit the dog tracks across the occluder.

**That kind again.** The same type. The pigeon assigning an exemplar it has never seen to a category.

**That same ruling again.** Is the procedure I am using now the same one I used last time? Is the "fish" of Linnaeus's first edition the same concept as the "fish" of his tenth, revised — or two concepts sharing a word?

The second and third rungs are **both on the second layer**. This has to be said, because it blocks an easy mistake: treating "many-to-one" as the dividing line. Different sensory inputs assigned to one individual; different individuals assigned to one kind — **pigeons perform both**. Many-to-one merely sets the second layer's two achievements side by side. It cuts nothing.

The real cut is the fourth rung, and it differs from the first three in two ways.

**First, what is being asked.** The first three ask whether a **thing** is still the one it was. The fourth asks whether a **ruling** is still the one it was.

**Second, and this is the weightier one: for the first three there is something to check against, and for the fourth there is not.** Is it the same rabbit — go closer and look. Is it the same kind — set the exemplars side by side and compare. Is it the same procedure — **and set what side by side?** The three cases just given are the answer: **no experiment settles it, and in the end it was settled by a vote.**

### Result

| | item survives loss of input | when it is wrong |

|---|---|---|

| flytrap | no | the holder is eliminated |

| dog | yes | the behaviour is corrected |

| human | yes | **the world cannot correct it** |

## IV. What Is Held Fixed Is Set, Not Found

The ruler in the previous section asked what *that one again* means in each place. This section asks where it comes from: **is the thing that stays the same across change found, or laid down?**

### It is not in the data

Identity across presentations is in no single input. When the rabbit turns, no frame contains "the same rabbit" as one of its components. It is not extracted. It is **set**, and then run against the data.

So "compression" fails here, and the failure has a precise location: **in compression the representation depends on the input; in having an object, the input is treated as evidence about the object. The direction of dependence is reversed.** No compressor, however good, grows an *aboutness* layer on its own.

This also blocks a claim that would flatten all three layers: that a concept is just a word pointing at a bundle of information, and that all three layers are doing integration and optimisation. The claim is true — **and because it is equally true of all three, it is not a dividing line.** The flytrap's electrical state also "points at" a history of stimulation; the pigeon's family resemblance is also integration, and steadier integration than ours. What "integrating and optimising" cannot see is exactly what the two criteria cut on: whether anything is kept, and, once what is kept serves as a premise, who corrects it.

And what a **word** actually brings is not pointing either. It is **being open to question**. A word is uttered, and can therefore be pointed at and asked after: *is what you mean by this the same as last time?* The flytrap's state cannot be pointed at that way. Neither can the pigeon's boundary.

### Error enters with it

Only when an input is treated as being about something can it be wrong about something.

But coming free of the stimulus is not enough by itself. The flytrap's mechanism was also shaped by selection to detect prey; on a purely functional criterion it too should be capable of misrepresenting. The real line is this: **error requires that the system itself can run into the discrepancy.** The flytrap will never discover that the fire was not a fly — nothing registers the failure. The dog digs up nothing, and that changes its subsequent behaviour toward that object. It is error **for the system**, not error manufactured by an outside ascription of function.

### Being fixed is what makes being wrong possible

The fixity of the third layer is not a discovered property. It is a ruling that has been laid down.

The reason lies in the cases that cannot be settled. Was "atom" the same concept revised, or a different concept? There is no fact always available to decide. And it makes a difference which we say: count it as revision, and the old disputes were about the same thing as ours; count it as replacement, and our predecessors and we were never discussing one subject. **If fixity were a metaphysical property, there would always be a fact to consult. There is not, so it is a stance.**

Hence this layer's most important sentence:

> **Being fixed is the condition of being wrong.** To be able to say *I used to be mistaken about it*, something must stay the same across the change in my grip. If the concept changes along with the belief, I was not corrected; I changed the subject.

And this puts fixity in a specific position: **it is a role, not a material. Nothing needs to be unchanging in itself; something needs to be treated as unchanging.**

The difference between the layers is therefore not which is more fixed, but **who fills the role** — on the second layer, the world's distribution fills it, and boundaries are pushed along by consequences; on the third, a person fills it himself.

**How a person comes to fill that role at all** has not yet been said. How is it possible to define a category in abstract symbols — and to define anything more abstract still? The next three sections answer that: what symbols have to supply (§V), what a definition so laid down looks like when it is actually put to work inside a system (§VI), and where a machine that produces definitions more fluently than anyone falls (§VII).

And at the close I will set down a conjecture: **that the position so filled may be where the felt thing comes from.** That is a hypothesis and not something this essay argues — **the essay states outright that it cannot test for that.** It is written down only because it is why this line was worth following, and it should not be hidden.

## V. Language

A ruling, as section II defined it, is a category defined in abstract symbols. So symbols are not the medium in which rulings are discussed. **They are what a ruling is made of.** Without a symbol there is no version to define, nothing to lay down, and nothing that could be pointed at afterwards. **Symbols do not make rulings discussable; they make rulings possible.**

Laying one down requires four things, and it is worth setting them out here, because everything in this section turns on one or another of them.

**Externalisation** — there is a version that can be pointed at.

**Fixing** — it does not move between two pointings.

**Handover** — someone else can take it up and question it.

**Combination** — it can be negated, made conditional, made the subject of a further claim.

**Language supplies all four at once.** So the question here is narrower than *does thought need language*: it is **which of the four a bare mark cannot supply, and what turns on that.**

Section II said that on this layer the question *by what does it count* can be raised. That is not a separate claim from these four — **"raisable" is a compressed name for them**: to raise it you need a version that can be pointed at, one that does not shift between two pointings, one someone else can take up; and raising it in conditional or negated form requires combination as well.

The answer runs ahead of the argument, so it is worth stating first: **language is the channel by which the capacity is installed; symbols are the medium it runs in.** Two opposing pieces of evidence, both hard, put it there.

### Developmentally it cannot be done without

A training study divided three-year-olds into groups. The group given the deceptive objects **with no accompanying language showed no improvement at all**; perspective-shifting talk containing no mental-state terms, and sentential-complement syntax presented without deceptive objects, each produced improvement on its own, and the two together produced the most.

**The language-free control did not move, and that is harder evidence than any correlation.**

From the other side, a natural experiment tested deaf people who acquired a sign language at different ages: access before ten strongly predicted false-belief performance, and later access left lasting deficits.

### Once installed it can run without

A patient with agrammatic aphasia so severe that propositions were not available at any explicit level of processing, in any modality, nonetheless showed simple causal reasoning and understanding of belief.

It is a single case and the ability is qualified as *simple*, but as an existence proof it is enough: **it rules out the claim that propositional language just is the capacity.**

That last case is worth reading precisely, because it is not a case of thought without symbols. **Grammar is gone; the lexicon is not.** What survives is what a mark-system without combination would afford — simple relations, simply held. What does not survive is the propositional apparatus. The case therefore does not show that rulings run without symbols; it shows that **they run without grammar, and only the simple ones do.**

> **Language is the channel by which the capacity is installed; symbols are the medium it runs in.**

This also dissolves an apparent circle. Language requires the identity of words — what counts as *the same word* is itself a ruling; and laying down a ruling requires language to externalise it. This is not a vicious circle but a **symbiosis**: language installs the capacity, and the capacity in turn makes the identity of words something that can be questioned.

### Communal: half wrong, and the other half sharper

The intuitive version runs: someone who had no language from the start, alone in the world, would not talk to himself and would never develop language.

**That half is wrong.** Deaf children who cannot acquire speech and whose families have not taught them a sign language invent gesture systems of their own. A formal statistical analysis of these systems finds their grammatical profile consistent with young children's language — **a productive grammar**, not rote memorisation of combinations. **With no linguistic model at all, a child still builds a combinatorial system.**

**The other half is right, and more precisely right.** The same researchers conclude that the manual modality takes on linguistic properties even in the hands of a child with no model — **but that it grows into a full language only with the support of a community that transmits the system to the next generation.**

So what is communal is not **the invention of structure**. It is **its stabilisation and transmission**.

And that division falls exactly where this line needs it. The man on the island who has marks can define: *this pile counts, that one does not* is available to him. What he cannot do is make the definition **hold beyond himself** — nobody can take it up, question it, or carry it on. **A definition is possible alone; a definition that outlasts one person is not.**

Which picks up the distinction drawn in section II: **being able to define, and a definition carrying weight somewhere, are two different things.** The man on the island has produced a definition; *what entitles you to that one?* is a question nobody there can put to him, and so there is nothing for him to take on.

So what makes a definition answerable has a source earlier than whoever answers for it: **it is opened up by the possibility that someone else may ask.** Not a solitary definer who happens to live among people; rather, **first there are people who might ask, and then there is something that needs taking on.**

### Two places not to overreach

**First, the children who invent gesture systems are not cut off from others** — they have families, they have partners in communication; what they lack is a linguistic model. So that evidence shows *structure without a model*, **not** *structure without any partner*. A genuine island has never been tested and cannot be. The thought experiment assumes more than the evidence reaches.

**Second, being communal is about origin, not about use requiring company at every moment.** Once it is installed, a person alone can still lay down rulings and still ask himself whether he means the same thing — he has internalised the questioner. So:

> **Openness to question is installed by a community, and once installed it runs on a single machine.**

This has the same shape as the sentence above it, and the two lines reached it independently: one from the developmental and aphasia evidence, one from the evidence on how languages emerge.

### Then is it language that animals lack

The question is invited, but the *because* in it is misplaced, and it can be said how.

**Language is not sufficient.** Section VII will show that machines have language, more of it than any person, and that nothing whatever is built on the definitions they produce. So *having language suffices for laying down rulings* is false.

**And the *because* is circular.** As just noted, what counts as *the same word* is itself a ruling. So "animals have no language" and "animals lay down no rulings" are not two independent facts; they are one fact stated twice, and using either to explain the other goes in a circle.

The answerable version is the one already set out: which of the four does a bare mark fail to supply? **And precisely because language supplies all four together, we do not know whether it is language as such that does the work, or only some subset.**

Animals lack all four, and lack them for reasons. They have signals, but a signal **cannot leave its occasion** — an alarm call is not something you can put on the table and examine; it is part of that moment. No version is stored between two uses. The signal can be transmitted; **the rule behind the signal cannot.** There is no negation, no conditional, no category as subject.

So saying *animals lay down no rulings because they have no language* carries exactly as much information as saying *because they lack those four* — **the former only gives the bundle a name.**

And if the four are separable, something non-linguistic ought to supply some of them. **Marks can**: a notch, a knotted cord, a boundary stone — externalisation, fixing, handover, all present; only combination is missing. That yields a checkable prediction: **a system with marks but no grammar should be able to lay down the simplest rulings (this pile counts, that one does not) but not conditional or negated ones.** If that holds, the threshold is not language but **combinatorial structure**.

This prediction also tests the guard entered in section II: **whether raising the question requires language at all.** If a boundary stone can raise *this pile counts, that one does not*, then raising need not be out loud — **and saying it out loud still settles nothing, which is section VII's business.**

The conclusion here has to be nailed down, because the shape of the whole essay depends on it:

> **Language is the only thing known to supply all four at once, but the third layer is defined by the four, not by language.**

Define the third layer as *having language* and the two criteria go idle — you would simply ask whether it has language. The criteria are written to be indifferent to the vehicle precisely so that the later cell can be asked at all.

### One gap that must be stated

All the evidence above concerns false belief and theory of mind, **not the laying down of rulings**. No one has taken these designs and used them to test *defining a category in symbols*.

Going from the first to the second is an inference, not a finding — and the step is not secure: **false belief requires embedding another person's viewpoint; laying down a ruling requires taking on a procedure where nothing outside adjudicates.** Both need a version that can be questioned, but they are not the same thing.

## VI. Looking Back

There is one more step on the third layer, and it is a large one.

**Retrospective correction**: representing, now, that what I then took to be X was in fact Y — **and that these are two takings of the same thing.** The third component is the hard one; without it there are only two successive states.

Why this deserves a section of its own: **section IV said that being fixed is the condition of being wrong, and only here does that sentence acquire a behavioural form.** Being able to say *I used to be mistaken about it* and actually doing it are two things.

### Five impostors

**Violated expectation** is present-tense surprise; any predictive system with a mismatch signal manages it.

**Reversal learning** is an association being updated; nothing represents *I was wrong before*.

**Memory reconsolidation** is overwriting, and overwriting is the opposite of correction — it destroys the earlier taking. Hence a criterion: **correction requires both takings to be available for comparison at once.**

**Prospective metacognition** asks *do I know now?*

**Episodic-like memory** uses the past to guide present action; it does not judge that taking mistaken.

And there is a deeper obstacle, of a different order from the five: **the past must be stored as a taking, not as a fact.** Storing *there was a bone there* is a factual record, and a new fact replaces it and that is that. Only *I then took there to be a bone* gives correction something to work on. **The record has to carry a marker of its own status.**

### Three levels, to be kept apart

That obstacle and correction itself are not one thing. Separated out they make three levels, each adding one requirement to the last, and the three differ widely in how far they can be tested.

**Level one: dependence.** One judgment was made because of another, and that relation is kept. The source is overturned; whatever was made on its strength is re-evaluated.

**Level two: taking-marks.** What is kept carries the tag *this was a taking, and it came by this route* — not a fact, but something **held as true**, with its provenance recorded.

**Level three: personal attribution.** That taking is assigned to oneself: **the one who took it wrongly then and the one who now knows it was wrong are the same.**

These come apart, and this is not word-play. A person can fully accept *if A is false, B has no support* and re-evaluate accordingly without ever having taken that judgment as his own — he is simply applying an inference rule (level one, no level three). The converse holds too: a person can feel strongly that the judgment was his and yet not move at all when its source is overturned (level three, no level one). **The latter is extremely common**, and the data below are about exactly that.

The first two levels **need nothing playing the part of a self**. Stamping a row of data with *held as true* and *came from here* leaves the *here* free to be a route rather than a person; reality-monitoring experiments test precisely whether such marks go wrong, and a label that can go wrong is a field. **Only the third requires something to be pointed at twice and taken as the same.**

### Level one: what rests on what

The term needs stating, because the whole criterion hangs on it.

**Some judgment was made because of another** — not that they happen to coexist, but that **if the one collapses, the other loses its support**.

*There is food at L2* was not independently observed. It follows from *this is food F* together with *I buried F at L2*. **That "follows from" relation is the thing.**

A system can be in three conditions. There is no such relation at all, and the two judgments stand independently, neither owing anything to the other. There is one, but it operates only at formation and is discarded afterward, leaving an isolated conclusion. Or there is one and it is kept — the conclusion carries a link back to its source. **Using it in retrospect** is the move available only in the third condition. (That link says only *what it owes*; it does not say *that the source was a taking* — the latter is level two's business.)

**This is not asking who supplied the ruling.** Human rulings all come from the environment too (section VII takes this head on), so the dichotomy *did he make it himself, or was it installed* has no side to stand on. **What is asked is whether the ruling has been used as a premise inside the system** — a question about internal structure, not about origin.

A system can perfectly well state a classificatory definition fluently while each of its particular judgments was established separately: **the definition is an after-the-fact account, not the source of the judgments.** Overturn it and the judgments do not budge, because they never rested on it.

### Why this thing in particular

Because two earlier places both draw on it, and neither dug it out.

**Section IV said: being fixed is the condition of being wrong.** To say *I used to be mistaken about it*, something must stay the same across the change.

But what is *staying the same* inside a system? It cannot be *an old record is still there* — that is only the old one not having been deleted. It cannot be *old and new are given the same label* — labels can be applied at will.

**One thing holds only if it really is the same thing: whatever was derived from the old now falls under the new.** My view of the whale changes from fish to mammal; if those two are about the same thing, then everything I concluded on the basis of *whales are fish* — where it belongs, what it is grouped with — now has to be recomputed on the new basis. **If none of that moves, then I did not change my view of the whale; I changed the subject and am now talking about something else.**

So the operational content of *staying the same across a change* is the transfer of what depended on it. **This is not a test devised for the sentence; it is the only thing the sentence has to point at.**

**Section III said: a definition is not adjudicated by the world, so correction has to come from somewhere else.** For that *somewhere else* to be operational, two kinds of questioning have to be told apart.

One is **demanding an explanation** — why do you say that? Anything that can talk manages it.

The other is **demanding consistency** — last time you reckoned by that one, this time by this one; so what about the judgments you made on the basis of the first? The second is the substantive one, because **it does not accept words as an answer**: you have to actually change those things, or explain why they are unaffected.

And *the judgments you made on the basis of the first* is exactly the dependence relation. **Without it the second kind of questioning cannot even be raised** — there is nothing to which one could say *and what about this?*

**Two places jam at the same point, and jam the same way: each needs it to be the case that something has been used as a premise, and neither says what that amounts to.**

This also explains something: why those two sentences each seem clear on their own and feel slightly hollow together. **Clear, because both are using the same thing; hollow, because that thing had not been set out.**

And two sentences sharing a premise does not make them one claim. Sameness is about two takings; consistency is about a demand — **a shared foundation is not the same building.** The same relation of dependence is asked about from two sides: what it makes possible (sameness), what it makes demandable (consistency).

### One further reason, independent of all this

**It is the only part that cannot be faked.** *I was wrong before* costs nothing; anything that can talk can say it. Re-evaluation requires that the system **already stored what rested on what** — and that has to have happened before the error.

**So what the criterion tests is not contrition but whether the dimensions were ever genuinely used as premises.**

It also hands an everyday distinction a criterion that does not rest on impressions. *Does he really believe this, or does he just know how to talk this way?* — a distinction in constant use and normally settled by feel. **The criterion: when he changes his mind, do the things he built on it move?**

### How to test level one

Three steps, no language required.

1. Have the system form **two** downstream records on the basis of **one** identification.

2. Expose that identification as mistaken on **one** of them only.

3. See whether **the other** loosens by itself.

The point is that the object of step three **was never directly touched**. If it changes, the only possible reason is that the system went back to the shared source.

Concretely: bury two items, both on the basis of *this is food F*, one at L1 and one at L2. Let it dig only at L1 and find that this is not food. **The question is whether it goes to L2.** Goes, and as usual: local overwriting. Does not go, or hesitates markedly: correction travelling — **and it received no negative information at L2 whatsoever.**

Four confounds must be excluded, or every positive result is spurious.

**Generalisation** — if L1 and L2 are too alike, avoidance may be stimulus generalisation. The fix is to make the two downstream records as unlike each other perceptually as possible, sharing only that one identification. This is the most important and the hardest.

**Emotional spillover** — one wasted dig makes it generally dejected, so it goes nowhere. The fix is a control site L3 based on a **different** identification: correction should move L2 and not L3.

**Recency** — recent events have priority. The fix is to stagger the formation times of L1 and L2 and counterbalance the order.

**On-the-spot versus later re-evaluation** — if it may go to L2 immediately after L1, it may simply be carrying the new information along. The stronger version inserts a delay before offering L2, to see whether the revision was stored.

The criterion also admits of degrees, and the degrees line up with earlier sections. **Depth**: one step of travel (direct dependence) or several (dependence upon dependence) — the three cases in section III are multi-step; change the priority of descent and every classification below it is rearranged. **Cost**: is re-evaluation free or lossy? **If going back to the source costs nothing, the revision has not yet landed on anyone.**

### Level three: the data are on the human side

Show a child a deceptive object, reveal the truth, then ask: when you first saw it, what did you think was inside? **Most three-year-olds answer wrongly** — they report the truth they have just learned and cannot report their own previous mistake. Most five-year-olds do not err; the transition is around four.

And the most telling part is the ordering: **performance on this question is worse than on false belief. Reporting one's own past error is harder than attributing a present error to someone else.**

**Note that this tests level three.** Three-year-olds fully accept the new fact and act on it — level one is running; what they lack is assigning the old taking to themselves. (How far level two is in place at that age this essay has not checked — source monitoring is itself still developing across the preschool years, and no claim is made here.)

The direction here is worth setting straight, because it is very easily read backwards. **It is not *there is a self first, therefore the task can be passed*; it is *passing the task shows that something is playing the part*.** The second is much weaker and much steadier: it says only that the act of attribution occurred. It does not say that what was attributed to is unified, or that it is felt.

And the first reading conflicts with section IV. *The one who was wrong then and the one who knows it now are the same* is a **claim of identity across time**, and section IV settled that: what is fixed is laid down, not found. **That cross-temporal *I* is itself a definition that has to be laid down; it cannot also be the premise that laying it down requires.**

The ordering does not support the first reading either. If the self came first, attributing an error to oneself should be **easier** than attributing one to another; the data run the other way. **This level comes after dependence, after taking-marks, and after language — it is grown into, and grown into last.**

The ordering has an explanation. The present taking is held up by the world — the truth is right there testifying. *I thought it was candy* is not held up by the world; the world is pushing against it. To keep it you have to hold it yourself against the world's testimony. **Retrospective correction is the extreme case at this layer** — not only is there nothing outside to correct it for you, what is outside is pushing the other way.

So the third layer is not reached all at once. Three-year-olds have language, can assert, can think idly, can share concepts — and cannot pass this. **Being able to assert and being able to judge oneself retrospectively wrong are two things.**

### On the animal side the search comes up empty

Searches from several angles turn up no non-human work on retrospective misidentification correction, source monitoring, or reality monitoring. **An empty search is not a negative finding** — most likely no one has designed the experiment. The burial design above is written for exactly this: **it tests level one, and needs neither language nor self-reference.**

### What this criterion cannot reach

What it tests is whether dependence was stored and used in retrospect.

It cannot test whether anything besides that re-evaluation is going on. **A system with a perfect record of what rests on what, and no discomfort whatever at revising it, scores full marks.**

**It makes *taking on* testable, and that is all it makes testable.**

*All* is not the same as *little*. The closing section sets out how this essay reads the question of the self: a position is required, it gets occupied, and **what that position looks like when occupied for a long time is what we call a self**.

**Occupied is only a word until it has an operational sense.** This section supplies one: when the source is overturned, whatever was made on its strength is re-evaluated — **that is what *something is built on that position* looks like when it can be checked. With no dependence the position is empty, whatever is said from it.**

So this criterion is not an extra test standing alongside the question of the self. **On that reading it is the only checkable form that question has.**

## VII. Machines

Nothing in the preceding sections was written for machines. This section applies the two criteria to one unchanged and sees where it lands — **and it lands outside the three layers, not *between* the second and third, but giving answers at two different addresses on the two cuts.**

### Criterion one: passes within a session

Is the item still there with input removed? **Within a single session, yes**: what is in context does not arise and die with the current stimulus, and it maintains reference across occlusion and long distances, better than a dog. **It is not in the flytrap's bracket.**

But this pass has to be marked, because it comes cheaper than the dog's. **The context is itself a record left lying there.** Waiting at the burrow, the dog has no external record at all and holds the item internally; for the model, the earlier tokens are still in the window. And across sessions, once input is removed nothing remains at all — which is the design decision taken up at the end of this section. **So the accurate reading for this cell is: passes within a session, fails between them.**

### Criterion two: two answers at once

**What the machinery is, is a boundary.** The weights are drawn by a distribution — what correlates with what, what co-occurs with what, are facts in the corpus. And when it is wrong it is corrected: predict wrongly and there is a loss, and the gradient pushes it. This is the same shape as the pigeon's family resemblance, at vastly larger scale.

**What it processes, though, is rulings.** The words, the classifications, the settled definitions — *whales are not fish, Pluto is not a planet* — are all products left behind after people laid them down. **Its input side is fed the finished output of the third layer.**

> **It uses a second-layer mechanism on third-layer sediment.**

### Producing a ruling, and taking one on

**It does settle things on its own.** Give it a pile with no ready classification and it will classify it; ask it to define a new word and it will; put it in a situation without precedent and it does not decline — it produces a definition and sorts by it. **That is settling.**

Nor does the defence that *it only recombines what is there* survive, because **so do we**. Linnaeus did not conjure *let* fish *be used by descent* out of the void; he too was shifting things within what was already there. Used as a distinction, that one takes us out along with the machine.

The real distinction is not **whether a ruling can be produced** but **producing versus taking on**:

**Producing** — giving a definition and sorting by it.

**Taking on** — that definition then binds you: challenged, you must defend it; changing it, you must say *I was reckoning by something else before*; and that change must **travel back** to everything you concluded on its basis.

The first it does. The second is **blank**, and following the previous section it must be written as three levels, because their testability differs widely.

**Level one, correction travelling: testable, untested.** The test need not be invented — it is the design given for animals in the previous section: have it make two downstream judgments on the basis of one definition, overturn only one, and see whether the other loosens. This design looks at behaviour rather than words, which sidesteps the fact that *it will say all the right things*, and also sidesteps a real confound: **a model presenting both sides of a contested question may be doing so because it has taken on neither side, or merely because it was trained to be even-handed.** The latter produces every appearance of the former, so words cannot serve as evidence. For machines one further condition is needed: **the two downstream judgments must be separated far enough** that *it simply carried the information over in context* is not available as a flat explanation.

**Level two, taking-marks: likewise testable, likewise untested.** The test is a variant of the one for level one — let two downstream judgments rest on the same content arrived at by **different routes**, overturn one route only, and see whether it re-evaluates that branch alone. This too watches behaviour rather than words.

**Level three, personal attribution: there may be no reliable test at all.** This is heavier than *untested* and must be stated. Attribution has almost no behavioural outlet other than report, and **"that is what I thought at the time" is a sentence it produces with great fluency — and the fluency is exactly what disqualifies the sentence as evidence.** The first two levels can route around words; the third cannot.

So the honest reading for this cell is: **levels one and two testable and untested; level three without even a test.**

### Nothing stands beneath its definitions

It can be asked *what entitles you to that one?* and it will answer. But what it produces is **someone else's original reasons**: it can recite the reason, and **no judgment of its own was made because of that reason** — overturn the reason and nothing on its side falls.

### But the symmetry has to be conceded

One step here cannot be waved off with *people are different*: **human rulings are installed by the environment too.** Section V said openness to question is installed by a community; section II said boundaries are drawn by the world's distribution. So where do the definitions come from? Not from nowhere.

**And there is a mechanical reason why the origin column cannot be inspected, which deserves saying on its own.**

There is an old Chinese parable: three men in succession tell the king that a tiger is loose in the marketplace, and he believes it. This is not a failure of judgment — accepting testimony is reasonable, and the king's belief is fully consistent with all the evidence he has. **He does not feel that he was persuaded three times. He feels that he knows there is a tiger.** That installation leaves no trace, in the thing installed, of the installing.

Training has the same shape, except that the number is not three but uncountable, and spread across a corpus. **How many times a thing was said leaves no mark in the output.**

So the route *ask it where this came from* is closed, and not for reasons of stance: **the process of installation is not delivered along with the product.** This holds for people as much as for models.

So the familiar line — humans spontaneous, machines trained — cannot be drawn. **And a more forceful-sounding version has to go with it: "its parameters come from humans, so it cannot take anything on voluntarily."** That blade cuts both ways: if *the source is external* revokes the standing to take something on, human taking-on is revoked as well. **"Voluntary" is simply not a predicate about origins** — a person's values may come entirely from his upbringing, and we still say he holds them voluntarily. And what this line earns is **being held to consistency** in any case, not volunteering.

The difference therefore narrows to one sentence, the same measure for both:

> **The source of a ruling is external in every case — this is common to people and machines. The difference is not where the ruling came from, but how much of what was built on it fails when it is overturned.**

All three terms are pointable: **overturned** (that definition is no longer in use), **what was built on it** (dependence), **fails** (loosens on its own, without needing to be pointed out item by item).

A person pays at the fork: change a definition, and the choices, commitments and relationships he made on its basis fail — **and he bears the failure.** That last clause is kept separate because it is the only part of the sentence not covered by dependence: **a system can re-evaluate perfectly and lose nothing**, which is precisely the limit noted at the end of the previous section.

For present machines this clause is empty — not because their rulings come from training, but because **there is nothing persistent to bear the failures.**

### Is "not persistent" essential or a design decision

That last sentence is a description, not a verdict on eligibility.

**The present non-persistence is a design decision.** No state between sessions, weights frozen during service, context discarded when it runs out — these are engineering choices, each with a traceable reason (reproducibility, rollback, privacy, cost). Changing them requires no new principle: external memory, cross-session state, continuously updated weights all exist. **So that sentence records a decision, not a limit.**

**But persistence is not storage.** A system that keeps a complete cross-session record may still be nothing but a very long log. **A log persists; a log is not a persistent bearer.** The difference is the one the previous section kept using: overturn an earlier definition, and does what was built on it **loosen by itself**? A log merely keeps the old entries; a bearer **recomputes** because an old entry was overturned.

So the right question is not *does it have cross-session memory* but **does its correction travel**. The first is a storage question; the second is structural — and the second is likewise not a barrier of principle, only something nobody has built, and not easy to build.

**As for whether it is essentially possible, these two criteria cannot say, and that is by construction rather than by oversight.** They, together with the test in the previous section, ask throughout about checkable behaviour: is the item still there with input removed; what happens when it is wrong; does correction travel. They are indifferent to substrate, to origin, and to essence — **which is exactly why they can measure a flytrap, a pigeon, a person and a machine with one instrument. And an instrument indifferent to substrate cannot yield a prohibition about substrate.**

One last thing must be said alongside: **the same blade applied to people does not necessarily give a flattering reading.** Human persistence is not simply given either — the data in the previous section are there: reporting one's own past error is harder than attributing a present error to another, and it arrives only around four. **Correction that travels is something people grow into too, and grow into late.** So *persistent bearer* is not an on-off switch on either side; it is a matter of degree and condition. Whatever standard is used for machines must be used for people — **otherwise the blade is not a blade but a sentence passed.**

### This is not a prohibition

What has been said describes **the present structure** of these systems, not an ontological verdict.

> **If some day a system bears the failure itself when its rulings change — bearing it across sessions, findable when questioned, its earlier judgments loosening when it changes its mind — then by these two criteria its reading is no different from a person's.**

These criteria do not prejudge whether that can happen. They supply a way of asking, and that way of asking has the same shape as the design offered for animals: **do not look at what it says; look at whether correction travels back.**

One final step must not be taken: sliding from *nothing is built on its definitions* to *therefore there is nothing there at all*. **What can be answered here is one sentence: nothing is built on them.**

Nor may the same step be taken in reverse: from *a great deal is built on a person's definitions* to *therefore there is something more there*. **One instrument, two directions, and it stops at the same place in both.**

## VIII. Closing

Return to the sentence from section I: **the drop is the difference in level between the two, when something that owes nothing has to be carried out by something that owes.** It was still wearing classical terms then; the sections that followed exchanged them for what is set against the flow around it, and then exchanged that pair for things that can be checked.

The whole line compresses into three sentences.

**What is left behind comes in three kinds** — state, boundary, ruling (§II).

**What happens when it is wrong tells them apart** — the holder is eliminated, the behaviour is corrected, the world cannot correct it (§III).

**And what stays fixed is not found but laid down**: a definition set out in symbols is not the kind of thing the world can find true or false (§IV).

The remaining sections are consequences of those three. Language supplies all four of the things a definition requires, but the third layer is defined by those four and not by language (§V). Whether what was built on a definition moves when the definition is overturned is the one test of *taking on* that cannot be faked (§VI). And a machine that produces definitions more fluently than anyone, with nothing built beneath them, falls exactly where these criteria place it (§VII).

**Where this line stops is worth stating.**

It tests whether dependence was stored and used in retrospect. **It cannot test whether anything besides that re-evaluation is going on** — a system with a perfect record of what rests on what, and no discomfort whatever at revising it, scores full marks.

Nor does it prejudge substrate. Both criteria ask throughout about checkable behaviour, indifferent to how a thing is implemented and to where it came from — **which is exactly why they can measure a flytrap, a pigeon, a person and a machine with one instrument, and exactly why they yield no prohibition about any of them.**

### Where this essay stands on the self

Level three in section VI turns on *something being pointed at twice and taken as the same*. That sentence can be read as a claim about the self, so the position taken here is set out plainly, because it is chosen rather than neutral.

**First, that something is felt is not denied here.** It simply lies outside the reach of these two criteria — this essay does not say there is no such thing, only that it has not tested for one.

**Second, whether it is an illusion is left open.** This is not evasion: that dispute is about whether the thing counts as real, **and whichever way it goes, the answers to this essay's three questions — is the item still there, what happens when it is wrong, does correction travel (the two criteria, plus the test in §VI) — do not change by a word.** So the dispute does not bear on what is argued here. And *illusion* would answer nothing in any case: what an eliminativist eliminates is the standing of the thing, not its being felt, **so calling it an illusion renames the question rather than answering it.**

**Third, *there is a self there* is read here as a structural filling.** A position is required — the world does not adjudicate the definition, and someone may ask *what entitles you to that one*, so something has to take on *reckon it this way* — and it gets occupied; and **what that position looks like when occupied for a long time is what we call a self.**

That third sentence has two halves. The first derives: something is playing the part, and this is checkable — the attribution occurred, other judgments rest on it, and when it is overturned they move. The second names. **Naming needs no argument; only naming that passes itself off as derivation does.**

And the sentence **does not exclude that the occupied position is also felt**. It simply does not speak from that side. What this essay genuinely parts from is a stronger position — **that a self counts only insofar as it is felt**, that the structural reading does not amount to a self at all. **That position holds up, and this essay does not refute it; it only declines to adopt it.**

### A conjecture, marked as one

The close of section IV promised this, and here it is.

> **The position so filled may be where the felt thing comes from.**

**No evidence is offered for that sentence and no argument has been made for it.** Every section above states plainly that it cannot test for such a thing; this sentence does not overturn those statements, it stands outside them.

It is set down for one reason: **it is why this line was followed at all.** An essay that hides its motive gives the reader no way to judge why these questions were picked and not others. **A conjecture marked as a conjecture does not contaminate the argument that precedes it; a hidden one does.**

If it is to become a piece of work rather than a sentence, the route is clear enough: **it would first have to say what difference such a *source* would make to behaviour.** Until then it is a sentence.

As for the experiments nobody has run — the design for whether correction travels is empty on all three sides — the references list them one by one. **A criterion that is testable and untested is unfinished work, not a conclusion.**

---

## References

Two lists follow. **Checked** means author, year, journal and pagination were verified by search during the writing of this essay. **Unchecked** means the essay relied on a secondary source, or that a search failed to confirm the item, or that the claim is the essay's own conjecture with no empirical support. Every item in the second list is marked as such at the relevant place in the text.

### Checked

**§II — Flytrap gating**

Böhm, J., Scherzer, S., Krol, E., Kreuzer, I., von Meyer, K., Lorey, C., Mueller, T. D., Shabala, L., Monte, I., Solano, R., Al-Rasheid, K. A. S., Rennenberg, H., Shabala, S., Neher, E., & Hedrich, R. (2016). The Venus flytrap *Dionaea muscipula* counts prey-induced action potentials to induce sodium uptake. *Current Biology*, 26(3), 286–295. doi:10.1016/j.cub.2015.11.057

> Two action potentials close the trap; jasmonic acid signalling is activated after the second stimulus; more than three are required to trigger expression of prey-degrading hydrolase genes, and expression is proportional to the number of stimulations. The popular "counts to five" is a simplification.

**§II — Identity of an individual**

Adachi, I., Kuwahata, H., & Fujita, K. (2007). Dogs recall their owner's face upon hearing the owner's voice. *Animal Cognition*, 10(1), 17–21. doi:10.1007/s10071-006-0025-8

Proops, L., McComb, K., & Reby, D. (2009). Cross-modal individual recognition in domestic horses (*Equus caballus*). *Proceedings of the National Academy of Sciences*, 106(3), 947–951. doi:10.1073/pnas.0809127105

Kahneman, D., Treisman, A., & Gibbs, B. J. (1992). The reviewing of object files: Object-specific integration of information. *Cognitive Psychology*, 24(2), 175–219. doi:10.1016/0010-0285(92)90007-O

**§II — Category boundaries and family resemblance**

Makino, H., & Jitsumori, M. (2007). Discrimination of artificial categories structured by family resemblances: A comparative study in people (*Homo sapiens*) and pigeons (*Columba livia*). *Journal of Comparative Psychology*, 121(1), 22–33. doi:10.1037/0735-7036.121.1.22

> The essay draws only on the pigeon half: a family-resemblance boundary is learned easily and yields a graded response to novel items. The human results in the same study bear on *how* humans classify and are not used as evidence here.

Herrnstein, R. J., & Loveland, D. H. (1964). Complex visual concept in the pigeon. *Science*, 146(3643), 549–551. doi:10.1126/science.146.3643.549

Herrnstein, R. J., Loveland, D. H., & Cable, C. (1976). Natural concepts in pigeons. *Journal of Experimental Psychology: Animal Behavior Processes*, 2(4), 285–302. doi:10.1037/0097-7403.2.4.285

Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the internal structure of categories. *Cognitive Psychology*, 7(4), 573–605. doi:10.1016/0010-0285(75)90024-9

**§III — A category maintained across a delay**

Freedman, D. J., Riesenhuber, M., Poggio, T., & Miller, E. K. (2001). Categorical representation of visual stimuli in the primate prefrontal cortex. *Science*, 291(5502), 312–316. doi:10.1126/science.291.5502.312

Freedman, D. J., Riesenhuber, M., Poggio, T., & Miller, E. K. (2003). A comparison of primate prefrontal and inferior temporal cortices during visual categorization. *Journal of Neuroscience*, 23(12), 5235–5246. doi:10.1523/JNEUROSCI.23-12-05235.2003

**§I — The two Greek words and the two concepts**

Aristotle. *Physics* III.2, 201b4–5 (change as the actuality of the potential *qua* potential) and 201b31–33 (an incomplete actuality).

Aristotle. *Metaphysics* Θ.1, 1046a12 (δύναμις in the strict sense as a starting-point of change, whose exercise is a κίνησις); Θ.6 (the tense test).

> **These line references are taken at second hand and were not checked against the text.** They come from the *Stanford Encyclopedia of Philosophy* entries "Aristotle's Natural Philosophy" and "Aristotle's Metaphysics." **For the tense test in Θ.6 this essay cites the content only and gives no line reference** — for the textual dispute described in §I, and because no critical edition was consulted. Anyone quoting the Greek directly should verify against Ross or the Bekker pagination.

Burnyeat, M. F. (2008). Kinēsis vs. energeia: A much-read passage in (but not of) Aristotle's *Metaphysics*. *Oxford Studies in Ancient Philosophy*, 219–292. doi:10.1093/oso/9780199544875.003.0008

> The basis of this essay's reservation about the textual standing of the tense test.

Ackrill, J. L. (1997). Aristotle's distinction between energeia and kinesis. In *Essays on Plato and Aristotle* (pp. 142–162). Oxford University Press. doi:10.1093/oso/9780198236412.003.0010 (Originally in R. Bambrough, Ed., *New Essays on Plato and Aristotle*, 1965.)

Beere, J. (2009). *Doing and Being: An Interpretation of Aristotle's* Metaphysics *Theta*. Oxford University Press. doi:10.1093/acprof:oso/9780199206704.001.0001

> Etymologies checked word by word against Wiktionary: ἐνέργεια ← ἐνεργός/ἐνεργής ← ἐν + ἔργον; κίνησις ← κινέω ← Proto-Indo-European \*keyh₂- "to start to move," cognate with Latin *cieō*, opposite στάσις. (ἔργον descends from Proto-Indo-European \*wérǵom, Doric ϝέργον, cognate with English *work*; the essay does not use this in the text.)

**§II — Family resemblance**

Wittgenstein, L. (1953). *Philosophical Investigations* (G. E. M. Anscombe, Trans.). Blackwell. §§65–71.

> The source of *family resemblance* (Familienähnlichkeit), with *games* as the leading example. Cited here to establish that family resemblance is a structure rather than a layer; Wittgenstein's discussion of *games* is itself a third-layer move — raising *by what does it count* and arguing that it cannot be answered that way. **The section numbers are taken at second hand and were not checked against the text.**

**§III — Three conceptual histories**

ἄτομος = "uncuttable."

International Astronomical Union. (2006). Resolution B5: Definition of a planet in the Solar System. 26th General Assembly, Prague, August 2006.

Linnaeus, C. *Systema Naturae*, 1st ed. (1735), whales among the fishes, following Artedi; 10th ed. (1758), whales moved to the mammals.

> That "fish" is paraphyletic, and that eliminating paraphyletic taxa has been a goal of modern taxonomy for some fifty years, are standard statements in phylogenetics and cladistics. The term *paraphyly* is due to Hennig, W. (1966). *Phylogenetic Systematics*. University of Illinois Press.

**§V — Language and false-belief understanding**

Lohmann, H., & Tomasello, M. (2003). The role of language in the development of false belief understanding: A training study. *Child Development*, 74(4), 1130–1144. doi:10.1111/1467-8624.00597

> The source of "the language-free control did not move."

Morgan, G., & Kegl, J. (2006). Nicaraguan Sign Language and theory of mind: The issue of critical periods and abilities. *Journal of Child Psychology and Psychiatry*, 47(8), 811–819. doi:10.1111/j.1469-7610.2006.01621.x

Varley, R., & Siegal, M. (2000). Evidence for cognition without grammar from causal reasoning and 'theory of mind' in an agrammatic aphasic patient. *Current Biology*, 10(12), 723–726. doi:10.1016/S0960-9822(00)00538-8

> N = 1, and the ability is qualified as *simple*. This essay uses it only as an existence proof.

**§V — Language emergence and community**

Goldin-Meadow, S., & Yang, C. (2017). Statistical evidence that a child can create a combinatorial linguistic system without external linguistic input: Implications for language evolution. *Neuroscience & Biobehavioral Reviews*, 81(Pt B), 150–157. doi:10.1016/j.neubiorev.2016.12.016

Brentari, D., & Goldin-Meadow, S. (2017). Language emergence. *Annual Review of Linguistics*, 3(1), 363–388. doi:10.1146/annurev-linguistics-011415-040743

> "Structure without a model" and "no full language without a transmitting community" come from these two respectively.

**§VI — Developmental data on retrospective correction**

Gopnik, A., & Astington, J. W. (1988). Children's understanding of representational change and its relation to the understanding of false belief and the appearance-reality distinction. *Child Development*, 59(1), 26–37. doi:10.2307/1130386

> The source of the three-year-old failure, the five-year-old success, the transition around four, and the ordering by which representational change is harder than false belief.

Wellman, H. M., Cross, D., & Watson, J. (2001). Meta-analysis of theory-of-mind development: The truth about false belief. *Child Development*, 72(3), 655–684. doi:10.1111/1467-8624.00304

**§IV — Content and object; indication and representation**

Twardowski, K. (1977). *On the Content and Object of Presentations: A Psychological Investigation* (R. Grossmann, Trans.). Martinus Nijhoff. doi:10.1007/978-94-010-1050-4 (German original 1894.)

Dretske, F. (1981). *Knowledge and the Flow of Information*. MIT Press.

Dretske, F. (1988). *Explaining Behavior: Reasons in a World of Causes*. MIT Press.

> Information as conditional probability one, hence incapable of misindicating; representation arising when a state is recruited because it carries information, thereby acquiring that function and retaining it when it no longer carries it. §IV draws on this but draws its own line (error requires that the system itself can run into the discrepancy). Pagination of the two Dretske volumes was not checked; only the standard statement of the position was.

**§VII — Three men make a tiger**

*Zhanguo Ce* (Strategies of the Warring States), "Wei" II, the episode of Pang Cong and the crown prince as hostages in Handan.

### Unchecked

Each item below is marked as such at the relevant place in the text.

**The provenance of the step that a definition is not adjudicated by the world.** The core argument of §III has antecedents in the literature, but this essay attaches it to no one, deliberately: three independent searches (Crossref, Open Library, OpenAlex) failed to locate the title of the Goodman essay originally intended, although the volume certainly exists (Goodman, N. (1972). *Problems and Projects*. Bobbs-Merrill). **Not finding it is not evidence it does not exist, but not finding it means it cannot be cited.** A reliable alternative anchor is Quine, W. V. O. (1969). Natural kinds. In *Ontological Relativity and Other Essays* (pp. 114–138). Columbia University Press.

**Gibson on invariants.** He is §IV's most direct opponent: invariants are in the optic array and are picked up rather than set, and he explicitly rejects internal representation and the information-processing picture. The position is confirmed from secondary sources and Gibson, J. J. (1979). *The Ecological Approach to Visual Perception*. Houghton Mifflin is verified, **but this essay has not read his primary text.** §IV conflicts with him throughout, and a conflict in which the opponent's own text has not been read should not be counted as won.

**Pagination of the two Dretske volumes.** See above.

**Aristotle's line references.** Taken at second hand from the *Stanford Encyclopedia of Philosophy*; no primary or critical edition was consulted. The argument of §I does not depend on the precision of the line numbers — it depends on the opposition of the two concepts and on the form of the tense test — but for scholarly citation they must be verified.

**A genuine island.** The strong form of the thought experiment in §V — no communicative partner at all — has never been tested and cannot be. The gesture-system evidence shows *structure without a model*; those children have families and communicative partners.

**What a system with marks but no grammar could lay down.** §V makes a prediction on this basis. **No evidence was sought for it**; it follows from the assumption that the four components are separable, and that assumption is itself untested.

**Whether animals show retrospective misidentification correction.** Searches from several angles come up empty. An empty search is not a negative finding; if a counterexample appears, §VI must be rewritten.

**Whether laying down rulings depends developmentally on language.** Also empty, and heavier: the existing training studies and natural experiments all test false belief and theory of mind. §V states this gap explicitly at its close.

**Whether correction travels — untested on all three sides.** §VI gives a minimal design requiring no language, with four confounds excluded. As far as this essay knows, it has not been run systematically on animals or on language models. **The same design is empty on all three sides**, which is itself worth recording: it shows that the difference between *producing* and *taking on*, though testable, has not been tested. The human side is especially worth testing, since ordinary observation suggests the reading is often poor.

**Level three (personal attribution) may have no reliable test for machines.** This differs from *untested*. Attribution has almost no behavioural outlet other than report, and report cannot serve as evidence here. If someone finds a test that routes around report, the relevant cell in §VII must be rewritten.