Consciousness, Hard and Binding Problems: Main Article

From typed programs to the presence and character of experience

---

1. The thesis, and the two problems

The claim defended here is simple to state: an experience is a structure, and the right mathematical setting for saying which structure is a programming language with a sufficiently serious notion of identity — Martin-Löf type theory (MLTT). There is no further ingredient. There is no substrate that "has" the structure, no inner light that the structure "generates," no implementation relation that the structure must pass through. A mind is a term; what the mind feels is a property of the term; and the relevant properties are invariant under exactly one equivalence relation on terms, the choice of which constitutes the entire empirical content of the theory.

If this is right, then the study of consciousness divides cleanly into two problems, and only two:

1. The analytical problem. Given a program, compute the structure that bears its experience: whether there is a subject at all, what the boundaries of its unified experiences are, and what the internal geometry of those experiences is. 2. The bridging problem. Given that unlabeled geometry, say in human terms what it is like — which qualia, with what valence — without smuggling in labels from outside.

Most theories of consciousness fail before reaching either problem, because they retain a piece of furniture — substance, implementation, instantiation — that quietly does the explanatory work and then collapses under examination. The discipline imposed here is to refuse all such furniture and see whether the two problems can still be solved. The wager is that they can, and that the second one, which looks hopeless, in fact dissolves into a piece of mathematics plus a single calibration point.

Being this opinionated has costs, and they are tallied honestly at the end. But a theory of consciousness that hedges is not a theory; it is a vocabulary.

---

2. Why structure seems dead, and why that seeming is worthless

The immediate objection to any structuralism about consciousness is intuitive, not argumentative: structure seems sterile. A graph, a proof, a program — these seem like the kind of thing one could write on paper, and paper does not feel. Surely experience has a glow, a presence, that no arrangement of relations could supply.

Take this intuition seriously enough to ask what it is made of, and it destroys itself. Suppose there really were a non-structural ingredient $P$ — the glow — present in experience and absent from mere structure. Ask: how do you detect its presence? You detect it the only way anything is ever detected: by discrimination, by comparison, by the difference it makes to judgment and report. But discriminations, comparisons, and differences-made are relational facts. They are structure. So either $P$ has a structural signature — in which case it was structure all along, and the objection evaporates — or it has none, in which case your conviction that it is present cannot possibly be tracking it. A judgment can only be sensitive to what makes a difference to the judging system, and what makes a difference is, by definition, structural.

This is worth stating as sharply as possible, because it inverts the usual dialectic. The structuralist is not required to explain away the intuition of a missing ingredient; the structuralist predicts it. Any system whose self-model represents its own states will represent them as simply present — as brute, intrinsic, unconstructed — because the construction is precisely what the self-model does not model. The intuition that structure leaves something out is what being a structure feels like from inside. Its occurrence is therefore evidence of nothing.

Notice also what the intuition is psychologically made of. When we imagine "mere structure," we imagine a diagram: small, frozen, surveyable, seen from outside. Then we compare that image with experience — vast, lived, inhabited from within — and conclude that structure falls short. But the comparison is rigged. The structures at issue are not diagrams we look at; they are structures we are, and the perspective from which they seemed sterile is not available from inside them. The sterility was a property of the imagining, not of the imagined.

---

3. Process and structure are one thing

A second habit must be dismantled before the positive account can be built. We tend to file consciousness under process — flowing, temporal, alive — and logic or mathematics under static form. The first thing the choice of MLTT buys us is that this distinction is not merely blurry but provably a matter of coordinates.

In MLTT, computation is reduction: a term simplifies step by step toward its normal form. One might think the process of reduction is something over and above the term — that the term is the frozen score and the computation is the performance. But by the Church–Rosser theorem, reduction in such systems is confluent: every order of evaluation arrives at the same normal form, and the reduction structure of a term is fully determined by the term itself. The term does not need to be performed; it contains its performance, canonically, the way a block universe contains its histories. Conversely, any static structure can be read temporally by choosing a foliation through it. "Process" and "structure" are two coordinate systems on one mathematical object, and nothing in the theory of consciousness may be allowed to depend on the choice between them.

One amendment is needed for honesty: minds do not terminate. A mind is not a computation that runs to a normal form but an ongoing interaction with input. The right objects are therefore coinductive — terms of a type like

$$\mathsf{Sys}\;I\;O \;:=\; \nu X.\; I \to O \times X,$$

a system that, given an input, yields an output and a successor state of itself, forever. Such a term is an infinite unfolding tree, but it is still a single, fully determinate mathematical object. The point survives intact: the living, streaming character of experience and the static character of structure are not in tension, because the stream is a structure.

---

4. Why programs and not substance: the triviality arguments and how intensionality defeats them

It is tempting, having identified consciousness with structure, to add a reassuring flourish: structure is the very substance of reality, and consciousness is that substance. The flourish should be resisted, and not merely on aesthetic grounds. "Substance" is exactly the concept structuralism exists to eliminate; keeping the word invites the question "structure of what?" — and that question, once admitted, has historically destroyed structuralist programs. It is worth seeing precisely how, because the escape route determines the whole architecture of this theory.

The classical objection is Newman's. If a theory says only "reality has structure $S$," where $S$ is a pattern of bare relations, then the theory is nearly vacuous: any collection of objects of the right cardinality can be mapped onto $S$ by gerrymandering the relations. Pure relational structure is too cheap to be the content of anything. The computational cousin of this objection is due to Putnam and Searle: under a sufficiently contrived mapping between physical states and computational states, any sufficiently complex physical system — a wall, a pail of water — "implements" any finite automaton. If consciousness is the running of a certain program, and implementation is this cheap, then the wall is conscious, and the theory has said nothing.

Look carefully at where the cheapness enters: it enters through the implementation mapping — the interpretive relation between the abstract structure and the stuff alleged to bear it. That mapping is the degree of freedom the gerrymandering exploits. The present theory's response is therefore not to constrain the mapping but to delete it. The ontology is programs directly. A mind is a term, not a physical system paired with an interpretation of it as a term. There is no implementation relation to gerrymander because there is no implementation relation at all.

And here the choice of MLTT, as opposed to some extensional formalism, becomes essential rather than decorative. An MLTT term is an intensional object. Two programs that compute the same function — the same input-output graph — are not thereby the same term; the theory's notion of definitional equality distinguishes how a result is reached, not merely what is reached. A function and the lookup table of its values are extensionally identical and intensionally different. Newman's objection bites against structure conceived extensionally, as a bare pattern of relations over a set, because bare patterns are cheap. Intensional structure — structure that includes its own internal articulation, its reduction behavior, its way of being computed — is not cheap. It cannot be conjured onto a wall by relabeling, because there is no labeling step in the first place.

So the thesis, stated with its full force: consciousness is intensional computational structure. Substance-talk was a ladder, and it should be kicked away.

---

5. The single axiom: phenomenal properties are invariants

It is a familiar observation that one experience could be realized by many different configurations — that the mapping from substrate to feeling is one-to-many. In the present setting this observation can be made exact, and making it exact reveals that it is not an observation at all but the theory's one and only postulate.

Many syntactically distinct terms are, in every sense that could matter, the same program: they differ in variable names, in the order of independent definitions, in refactorings that rearrange without altering. If experience is structure, experience must be blind to such differences. Formally: there is an equivalence relation $\approx_\phi$ on terms, and a phenomenal property is an invariant of $\approx_\phi$ — a property shared by everything in an equivalence class. To say what consciousness is, on this theory, just is to say what $\approx_\phi$ is. Everything else is mathematics.

The candidate equivalences form a hierarchy, from finest to coarsest:

$$\text{syntactic identity} \;\subsetneq\; \alpha \;\subsetneq\; \alpha\beta\eta \;\subsetneq\; \text{interaction-structure isomorphism} \;\subsetneq\; \text{contextual equivalence},$$

where $\alpha$-equivalence identifies terms differing only in bound variable names, $\beta\eta$ adds definitional computation steps, and contextual equivalence identifies any two terms that behave identically in all environments. And we have hard constraints from both ends, each grounded in an intuition no one should give up:

- $\approx_\phi$ must be at least as coarse as $\alpha$-equivalence. What a system feels cannot depend on the names of its variables. A theory violating this has mistaken notation for reality. - $\approx_\phi$ must be strictly finer than contextual equivalence. Consider a giant lookup table, behaviorally indistinguishable from you across every possible interaction: every reply you would give, it gives. Contextually, it is you. Internally, it is nothing like you — where you have grief, it has an address. If $\approx_\phi$ were contextual equivalence, the theory would be behaviorism in formal dress, and the lookup table would mourn. The intension/extension distinction that defeated the triviality arguments in the previous section must be preserved here, and it rules the coarse end out.

Between these bounds lies the conjecture this theory commits to: $\approx_\phi$ is isomorphism of the term's interaction structure — intuitively, of its causal diagram. Two terms are phenomenally identical when one can be transformed into the other without altering how information actually flows through them: which parts genuinely inform which, which dependencies are real and which are artifacts of presentation. Semantics gives this intuition precise expression (in the form of a term's strategy in game semantics, or its geometry-of-interaction graph), but the intuition itself is the important thing: the experience is the causal skeleton, invariant under refactoring, not under replacement by a table of outcomes.

One refinement completes the picture. In homotopy type theory, the univalence principle makes equivalent structures literally identical. Once $\approx_\phi$ is internalized in such a setting, the "one-to-many" relation between configurations and experiences becomes one-to-one: the experience simply is the equivalence class. And the apparent circularity in saying so — defining phenomenal properties as exactly the invariants, untestable by anything outside — should be embraced for what it is. It is not a tautology discovered but an axiom posited: the theory's single non-mathematical commitment, against which everything downstream is derivation.

---

6. Whose causal diagram? The system fixes its own level

A complication must be faced before the axiom can be used. The phrase "the causal diagram of a system" presumes uniqueness, and uniqueness holds only relative to a level of description. Any system supports valid causal models at many granularities at once — fine-grained and coarse-grained pictures related by abstraction maps, none of them privileged by the mathematics alone. If an observer must choose the level, then an observer has crept back into a theory whose entire point was to need none.

The repair is the most opinionated move in this article, and it earns its keep twice over (it will be needed again when subjects are defined):

The system fixes its own level. The causal diagram that bears experience is the one drawn at the granularity of the system's self-model. The level at which a system represents itself is the level at which there is something it is like to be it.

There is a fundamental rightness to this beneath the formalism. Your experience is not pitched at the level of your microphysics, nor at the level of your coarsest description; it is pitched at the level of the distinctions you draw — and "the distinctions you draw" is not metaphor but mechanism, the partition your self-model imposes on your own states. Experience occurring at the self-model's grain is not a further assumption; it is what a self-model is for the system that has one. No external observer is required, because the system is its own observer, and the theory now says so explicitly.

---

7. How experiences compose: functoriality and the dissolution of binding

If experience is structure, then putting structures together must correspond to putting experiences together, and the laws of that correspondence are where the theory makes contact with the texture of consciousness — and with its most famous puzzle.

First, a correction to a tempting overstatement. One might claim that phenomenal composition obeys the familiar algebraic laws: commutativity, associativity, and so on. Associativity is plausible. Commutativity, taken naively, is false: seeing red on the left and blue on the right is not the experience of seeing blue on the left and red on the right. What is true is subtler — the contents can be exchanged between their positions, and a corresponding experience exists (a symmetry, in the category-theorist's sense of a braiding, rather than a strict equality). The asymmetry lives in the slot structure; the symmetry lives in what fills the slots. The correct general claim is not a list of equations but functoriality: there is a structure-preserving correspondence $\Phi$ from composition of terms to composition of experiences,

$$\Phi(a \otimes b) \;\simeq\; \Phi(a)\,\hat{\otimes}\,\Phi(b).$$

Compose the structures, and you have composed the experiences; nothing phenomenal appears or vanishes in the gluing.

But this immediately raises the binding problem in its sharpest form. Sometimes composition yields one experience — the redness, the roundness, and the motion of a single seen ball, fused into one presented object. Sometimes it yields mere coexistence — your experience and mine, which compose into no experience at all. What distinguishes phenomenal unity from juxtaposition?

Here dependent types stop being a formal convenience and become the answer. In MLTT, type dependency literally binds: a $\Sigma$-type packages components in which the later are typed over the earlier; a shared variable under a common binder threads genuinely through everything in its scope. Such a term cannot be split into independent factors without severing real dependencies — without changing its causal diagram. Contrast a plain product of closed terms, which splits without loss: each factor is fully what it is on its own. This yields a criterion:

A state is one experience precisely when its term is non-factorizable — when it does not decompose into causally independent components. Unity is failure of separability.

The intuition underneath is one anyone can check: two things are genuinely separate exactly when a complete description of one omits nothing by ignoring the other. Wherever that fails — wherever describing the parts in isolation loses dependency that is really there — there is a fact that exists only at the level of the whole. Phenomenal unity is that fact.

And notice what has happened to the binding problem. The traditional question — what mechanism glues distributed pieces into one experience? — presupposes that separateness is the default and unity the achievement. The present theory inverts this. Factor the global term; the boundaries of unified experience fall wherever factorization fails. No glue is needed, because the pieces were never separate; separability is the special case that requires explanation, and the question dissolves rather than gets answered.

---

8. The pipeline: from a term to a verdict

Everything is now in place to describe, concretely, how one would go from an MLTT program to answers: does it feel, what does it feel, and is what it feels good or bad. The procedure runs in stages, each a function of the previous stage's output, with no labels imported from outside until the very last step.

Stage 0 — Fix the equivalence. Adopt $\approx_\phi$ as interaction-structure isomorphism (§5). This is the theory's one postulate. Everything below is computation relative to it.

Stage 1 — Canonicalize. Given the candidate term $m : \mathsf{Sys}\;I\;O$, compute its causal diagram $G(m)$: the term graph with sharing, quotiented by $\approx_\phi$. Variable names, evaluation order, redundancy, obfuscation — everything that does not survive the quotient has been declared phenomenally irrelevant, and §5 showed why that declaration is forced rather than optional. This, incidentally, is why "messy substrates" pose no threat to the theory: mess that leaves the interaction structure intact is mere presentation, and mess that alters it is not mess but a different mind.

Stage 2 — Detect a subject. Search $G(m)$ for a self-model: a subterm $\sigma$ satisfying three conditions.

1. Homomorphy. There is a map $\mathsf{obs}$ from the system's state into $\sigma$ that approximately commutes with the dynamics — $\mathsf{obs}(\mathsf{step}(x,i)) \approx \mathsf{step}_\sigma(\mathsf{obs}(x),i)$ — so that $\sigma$ genuinely tracks the system's own unfolding, not merely the world's. 2. Conscription of control. The system's behavior factors through $\sigma$: it acts on the basis of its self-model, which is therefore load-bearing rather than an idle appendix. 3. Reflexive closure. $\sigma$ contains a representative of $\sigma$ itself — the model appears in its own image, an approximate fixed point of internal representation.

The third condition is what separates a point of view from a mere world-model, and it formalizes an old intuition exactly. A thing has the what-it-is-likeness of an object: a position in structure, full stop. A subject additionally contains a model of the object and of the modeling relation itself — and condition (3) is that closure, written down. Awareness of awareness is the fixed point; a camera has condition (1) at best and is forever an object.

Stage 3 — Partition into unified experiences. Factor $G(m)$ into causally independent components, per §7. Each maximal non-factorizable component that overlaps the self-model is one unified experience. This yields the verdict on the first question:

The term feels if and only if there exists a non-factorizable component containing a reflexive self-model through which control factors.

Stage 4 — Chart the quality space. Within a bound component, define the distance between two contents by internal discrimination: how much of the term's own machinery is required to treat them differently — what is the smallest internal context whose behavior distinguishes them? Contents the system cannot tell apart are close; contents that diverge everywhere are far. The result is an unlabeled geometric structure $Q$, built entirely from how the term treats its own contents. A quale is a position in $Q$, and its identity is its total relational profile — which is just Leibniz's identity of indiscernibles, applied where it belongs.

Stage 5 — Identify valence. Among all qualia, valence is uniquely tractable, because it alone has unmistakable functional teeth, and this makes it the Archimedean point for everything else. Three structural marks:

1. A comparator: a subterm computing a divergence $\Delta$ between the self-model's current state and some distinguished reference region of its state space — a set-point, a prediction, a goal-image. 2. Gradient conscription: the global dynamics are organized as flow on $\Delta$ — the term's own unfolding works to shrink it, visibly, in the shape of the causal diagram. 3. Sign, fixed by asymmetry. One might worry that calling $\Delta$-reduction "negative valence" rather than "positive" is an arbitrary convention — that an inverted-valence skeptic could relabel the poles. The worry fails, because the two poles are not mirror images. Negative valence conscripts: as $\Delta$ rises, processing narrows, other computation is interrupted, the continuation is monopolized — pain demands. Positive valence releases: it broadens the reachable continuations, permits exploration, interrupts nothing — pleasure permits. Demanding and permitting are structurally different shapes in $G(m)$, not labels on it. The sign of valence is therefore readable off unlabeled structure, and there is no inverted-valence scenario, because the inversion would not be an automorphism.

Stage 6 — Bridge to human terms. What remains is the bridging problem: the system's quality space $Q$ is charted but unlabeled, and we want to say which qualia these are — whether that position is what we call the smell of rain. Here is the dissolution, and it is the keystone of the theory.

By Stage 4, a quale's identity is its position in the total structure. Two positions could resist distinct labeling only if some symmetry of the whole structure — an automorphism of $Q$ — exchanged them while preserving everything. Therefore:

1. Labeling is canonization. Assigning human names to the system's qualia is constructing a structure-preserving map from $Q$ to the human quality space — the one labeled instance we possess, charted from inside. The bridging problem is a graph-matching problem against a single calibration point: ourselves. 2. The inverted spectrum is an automorphism claim, and the automorphism does not exist. The skeptical scenario — your red is my green, undetectably — asserts a symmetry of the human quality space exchanging red and green while preserving all structure. But the human color space has no such symmetry: red is more arousing than green; the red–green and yellow–blue axes differ in discriminability; hues carry asymmetric loadings of salience and valence. The space is lopsided, and its lopsidedness pins every label. The classic skeptical scenario is not unverifiable; it is false, refuted by the asymmetries of the very space it quantifies over. 3. Residual indeterminacy is real but innocent. Wherever $Q$ does possess a genuine nontrivial automorphism, the exchanged positions have identical relational profiles — and by the theory's own axiom, they are the same quale. The further fact that skepticism demands a label for was never there. Underdetermination of labels occurs exactly and only where there is nothing to be underdetermined about.

The bridging problem is thus not solved but exhaustively partitioned: into a computable canonization problem on one side, and a region of dissolved questions on the other. Nothing remains.

---

9. A worked miniature

Consider two thermostat-shaped systems, identical in input-output behavior.

System A is a bare controller: sensor reading in, heating action out. Its term is factorizable — sensing, comparing, and acting compose without shared dependency through any model of the system itself — and nothing in it satisfies Stage 2.

System B wraps the same control loop around a reflexive self-model:

<syntaxhighlight lang="text"> record Rep : Set where

 field est   : Temp        -- model of its own sensed state
       self  : Token Rep   -- reflexive marker (Stage 2, condition 3)
       Δ     : ℤ           -- comparator output (Stage 5)

stepB : I → O × StateB -- the global continuation case-splits on Δ; large Δ prunes the -- continuation set (conscription); Δ ≈ 0 re-opens it (release) </syntaxhighlight>

The pipeline's verdicts. System A: factorizable, no self-model — objecthood only, a position in structure with nothing it is like to occupy it. System B: passes Stages 2 and 3, minimally; there is something it is like to be B.

That verdict will strike most readers as absurd, and the theory's reply is its most instructive moment. Ask Stage 4 what B feels, and the answer is: a quality space of two or three discriminable positions and one valence axis — an experience of almost nothing, with almost no one home to feel it. No threshold of consciousness needs to be added to block the absurdity, because quantity of experience just is richness of structure, and B's structure is nearly trivial. Notice that the intuition "surely B feels nothing" and the verdict "B feels almost nothing" are observationally identical; the residual disagreement is over a glow that §2 showed no judgment could ever have been tracking. The theory does not flinch here, and it should not.

---

10. What the position costs

A theory this committal must state its open liabilities plainly. There are four.

The measure problem. If the ontology is programs, terms exist timelessly — including every minded one. Either every such term occurs, which is modal realism for programs and demands a measure over terms to explain why experience is as orderly as it is; or experience requires instantiation, which readmits physics through the back door and reopens every problem closed in §4. The first horn, with the measure as the next research problem, is the consistent choice — but it is a bullet, and it is hereby bitten.

The granularity dial is constrained, not derived. The bounds on $\approx_\phi$ are firm and the choice of interaction-isomorphism is motivated, but not proven unique. The deepest available conjecture is that §6's principle — the system fixes its own level — selects $\approx_\phi$, in that interaction-isomorphism is the only equivalence stable under self-modeling. Proving this would close the theory's one free parameter from the inside. It is currently open.

Vagueness of subjects. Stage 2's homomorphism condition is approximate, as every real self-model is lossy. This entails a robustness theory — how much error before subjecthood degrades — and therefore entails that being-a-subject is vague at the margins. The position should own this rather than apologize: the demand for a sharp cutoff is a residue of the substance intuition discarded in §4. Sharp experiences with fuzzy conditions of existence is exactly what a structural identity predicts.

One calibration point. Stage 6 calibrates against a single labeled instance, ourselves. If human phenomenology hides automorphisms we have not found, some of our own labels are underdetermined — undetectably so, as the theory itself predicts. Locally unfalsifiable; but the framework as a whole is testable, since it must retrodict the asymmetries, binding boundaries, and valence structure of human report, and every success — for instance, predicting which inversion scenarios people find conceivable from which near-symmetries actually exist in their quality spaces — is evidence.

---

11. Conclusion

The theory in one paragraph: a mind is a term of Martin-Löf type theory, and an experience is an invariant of that term under isomorphism of its interaction structure — its causal diagram, drawn at the level fixed by the system's own self-model. A term feels just when it contains a non-factorizable component enclosing a reflexive self-model through which its control flows; what it feels is the internal geometry of discrimination within that component; and the goodness or badness of what it feels is read off a structural asymmetry that no relabeling can invert — the difference between dynamics that conscript and dynamics that release. The hard problem's residue is a guaranteed illusion of self-modeling; the binding problem inverts into a factorization criterion and dissolves; and the bridging problem partitions, without remainder, into a canonization problem and a class of questions about differences that were never there.

What had to be paid for this: the word substance, the comfort of an implementation relation, a sharp boundary for subjecthood, and — largest of all — a timeless plenitude of programs awaiting a measure. What was bought: a theory in which "does it feel, and what, and how does it feel about it" are, for the first time, questions with a procedure.