The Indexed Format: A Constitutive Theory of Consciousness for Organisms and Artifacts

1What Thirty Years Actually Bought

The premise that consciousness science is nearing explanation is correct, but the progress has not converged on one theory. It has converged on a factorization: empirically dissociable components with different mechanisms and instruments (Bayne et al., 2016; Seth & Bayne, 2022). Ginsburg and Jablonka (2019) had already catalogued nearly the same emergentist list: global access, binding, selection and learning, temporal thickness, intentionality, value, embodiment, agency, and self. Dehaene et al. (2017) then separated C0, C1, and C2; Northoff and Zilio (2022), state, structure, content, and access. Three decades of advance and disagreement are therefore compatible.

The 2025 Cogitate adversarial collaboration makes the point brutally. Integrated information theory (IIT) and global neuronal workspace theory (GNWT) preregistered divergent predictions and pass/fail criteria, and both lost central commitments. Posterior cortex carried decodable stimulus content throughout presentation while prefrontal decoding lacked GNWT’s onset-plus-offset ignition profile, and the preregistered phase-synchrony analyses failed IIT (Cogitate Consortium et al., 2025; Melloni et al., 2023; §8; Supplement S4). The standard conclusion is that the theories isolate different factors. Correct, but incomplete. Their mechanisms say how information is organized and recruited, not whose information it is or why any state matters to the system carrying it. Those possessives are IFT’s opening.

This paper states the problem and takes the derivation. It also answers Chalmers’s (2023) final challenge: what could remain missing after senses, world and self models, recurrence, memory, a global workspace, and unified agency have been built into one system. The missing condition is not another module but the relation that fixes who the modules are for. The theory can be given in one sentence, and the rest of the paper is the argument for it and the specification of what would refute it.

A content is conscious if and only if it belongs to a representational format whose relational structure is standingly recruitable, as structure, by the processes that constitute one system’s policy-integrated persistence regime: the processes that keep one active token in existence and resolve its competing viability demands into one commitment. What the content is like is the relational structure so recruitable. Whose experience it is, is fixed by the regime.

Each clause of that sentence answers one of the field’s open questions, and the four theses in Table 1 are the clauses stated as commitments, each with the finding that would end it. Table 2 fixes the terms, because the argument turns on distinctions that ordinary usage runs together: format and content, bearer and beneficiary, token and type, recruitable and recruited. Readers who want the theory’s logical skeleton before its evidence should read §4 first; it contains the derivation on which everything else depends.

T1–T3 are, I will argue, established in the sense that any competitor now owes an explanation of the same data at higher cost. T4 is a bet, and I will mark it as one. Section 12 states the protocols in full and, for each, what the principal rival theories predict instead, so that the tests are adversarial rather than confirmatory.

What the account buys: it preserves the functional achievements of workspace, higher-order, and metacognitive theories without identifying any with subjectivity; explains why phenomenal structure can outrun notice, report, memory, and use; predicts the Cogitate pattern; derives the cerebellar and enteric exclusions; counts subjects where integration theories miscount them; is immune to the unfolding argument; supplies the indicator program with a negative indicator no added module can counterfeit; and specifies in advance the experiments that would destroy it. The first conscious artifact will probably be small, embodied, fragile, and bad at almost everything, and we will not believe it, because it will not be able to tell us.

2Five Factors, One Design Problem

The minimal definition remains Nagel’s (1974): a system is conscious when there is something it is like to be that system. Block’s (1995) distinction does more than organize the terrain; it forbids transferring the functions of access to phenomenal consciousness without argument. IFT accepts that prohibition and rejects the inference that phenomenal character must therefore be nonfunctional. Held against the empirical literature, the criterion covers at least five separable questions.

Global state. Whether the system is in a condition permitting the relevant processing at all, and this is not a scalar. Bayne et al. (2016) show that ordering vegetative state, minimally conscious state, sedation, REM sleep, and alert wakefulness along one dimension is untenable: having a point of view is not gradable. The consequence for measurement is severe, since the perturbational complexity index (PCI) and Φ are advertised as measures of level (Casali et al., 2013; Farisco & Changeux, 2023).

Access. Why a representation is selectively recruited for shared control rather than remaining local or merely present in the phenomenal field: Block’s access and Dehaene et al.’s (2017) C1. Substantially solved in outline (Baars, 1997; Dehaene, 2014; Dehaene & Naccache, 2001).

Structure. Why the content has this phenomenal character rather than another (Seth & Bayne, 2022). Once barely begun, this now has a concrete research program: spatial extendedness can be derived from relational topology, and visual phenomenology appears vastly richer than what can be noticed or reported (Haun & Tononi, 2019, 2025).

Metacognitive typing. Whether and how the system represents its own states as seen, remembered, intended, known, reliable, or merely guessed: the family of functions Dehaene et al. (2017) call C2 and Lau (2008) formalizes as higher-order Bayesian inference. This is the one factor a major tradition treats as constitutive rather than as a service: higher-order theories hold that a state one is in no way aware of being in is not a state there is something it is like to be in (Brown et al., 2019). §7 treats that tradition as this account’s closest competitor and isolates the single point on which the two cannot both be right.

Valence. Whether the content matters. A representation can be recruitable, structured, and metacognitively typed while remaining affectively neutral, so this is a distinct question with a distinct mechanistic story, and it is where the machine case turns. The dissociation is not a thought experiment: pain sensation and pain affect come apart in humans by lesion, by drug, and by hypnotic suggestion, in both directions (Berthier et al., 1988; Ploner et al., 1999; Rainville et al., 1997; §6; Supplement S3.1). It also carries the moral weight (Butlin, Long, et al., 2026).

The C0/C1/C2 framework clarifies the exact point at which a functional taxonomy stops. A workspace can make information globally available and a metamodel can estimate confidence, detect error, and distinguish perception from imagination; both can be engineered and may dissociate in the same machine. Neither identifies the subject of the availability relation. Without an index, C1 and C2 are dimensions of a control architecture and do not turn computation into a point of view.

One distinction must be fixed now, and the field has just supplied the occasion. Mudrik et al. (2025) argue that P and A were never two kinds of consciousness but two necessary conditions for it: content must be structured so as to be potentially phenomenal, and it must be accessed, with either alone counting as unconscious processing. The reframing is right and IFT completes it, because those conditions are T3 and T2: geometry supplies the first and indexing the second, and their own statement of the A condition leaves its second argument open (§4). One disagreement survives. Dehaene (2014) already distinguishes preconscious information that is “accessible but not accessed” from broadcast; the issue is whether standing availability is merely a precursor or membership in the phenomenal field. IFT takes the second view once the bearer is fixed: format membership is standing recruitability within the policy-indexed field, access its episodic uptake. Mudrik et al. take the first and allow overflow of report but not of access. Lamme (2006) anticipated the separation, and IFT accepts it while denying the identity: local recurrence can build format-eligible content without individuating a subject. Sparse recruitment and a rich field are different relata, and F1 decides between them.

Two constraints cut across all five factors. Intelligence is mostly C0: recognition, semantics, evidence accumulation, control, and reinforcement learning can proceed without awareness (Dehaene et al., 2017), so performance is not consciousness. Report is evidence rather than the phenomenon, because it may sparsely sample a structured field (Doerig et al., 2019; Haun & Tononi, 2025; Tsuchiya et al., 2015). Canonical markers accordingly move with task relevance; the centroparietal P3b is the clearest example (Cohen et al., 2020).

These factors look like an empirical list because dissociation discovered them; recent AI-consciousness frameworks likewise treat them as a multidimensional profile (Evers et al., 2025). IFT’s stronger claim is that they are the parts of one engineering problem, and the identity stated in §1 says which problem. A bounded system has persistence conditions, so outcomes can be good or bad for it: valence. It must occupy a global state in which a standing structure can be held at all: global state. Its model requires relational structure, and that structure is what it is like: structure. From the standing structure the regime selectively recruits, because an organism cannot remain a distribution of interpretations but must pool evidence, select one course, hold it across delay, and coordinate around it (Dehaene et al., 2017): access. Because the model is fallible and the commitment is costly, the regime must weight how far to trust what it recruits: metacognitive typing. The five factors are one problem, keeping one bounded individual alive in an environment it was not preconfigured for, and Table 3 states the derivation factor by factor. What C1 leaves open is whose policy it is.

Engineering reproduces the convergence from a literature with no stake in it. Rosenbloom et al. (2026) map GNWT, recurrent processing theory (RPT), predictive processing, and IIT onto the Common Model of Cognition and recover recurrent local modules, a global working-memory hub and cognitive cycle, and multimodal states; Wray et al. (2026) derive, from the requirements of acting outside design scope, appraisal of novelty, urgency, goal-conduciveness, coping potential, and normative significance. Neither supplies an owner: one is an abstract architecture; in the other, survival is designer-written and failure belongs to the operator (Supplement S3.18).

The architecture also has an evolutionary lineage. Ginsburg and Jablonka (2019) present unlimited associative learning (UAL) as a transition marker, not a learning criterion: adaptive robots, neural cultures, and severed spinal cords can learn without a subject. In animals, UAL joins binding, working memory, valuation, prediction, attention, and a functional self–world distinction, and their Cambrian account explains why these factors travel together: flexible learning imposed organism-level costs, so cognitive expansion coevolved with neurohormonal and immune regulation. Safron’s (2020, 2022) Integrated World Modeling Theory and Damasio’s hierarchy supply the strongest world-modeling and biological precursors. IFT adds the constitutive rule: subjecthood begins only when one policy regulates one active token’s persistence (Supplement S3.9).

3The Format Thesis: Consciousness Is Not a Place, and Not a Threshold Either

Dennett’s (1991, 2018) durable contribution is not one negative result but three: no Cartesian Theater, no double transduction, and no pre-rendered inner world. Together they constrain what a conscious format can be, and the positive thesis follows from the constraints.

The first is the absence of a Cartesian Theater. Perceptual attributes are discriminated by different systems at different places and times; once made, a discrimination can alter action, memory, emotion, semantics, and later perception without being displayed again. Metacontrast and color phi show why the time of representing need not mirror the time represented. If two stories predict every measurable consequence, their difference is an imaginary boundary. Processing is therefore bounded by control windows, not by a moment of entry into awareness: a representation must arrive after the evidence on which it depends and before the policy deadline it can still affect. This is the control-window reading of Libet’s timing results developed by Dennett (1991), and it carries the same warning for GNWT: a cortex-wide threshold recreates the Theater the instant crossing it is made constitutive of subjective order.

The second result is no double transduction. Peripheral transducers convert physical properties into neural signals; nothing converts those signals back into privileged neural paint. Redness is represented, not re-created. Even a causally efficacious electromagnetic field would specify a vehicle, not a bearer, and would still require a non-arbitrary demarcation rule (Dennett, 2018; Keppler, 2021; Supplement S3.12). Mental paint is virtual paint: standing causal organization for implementable users.

The third result is the harder one. A freely explorable brain-in-a-vat world becomes combinatorially intractable as soon as the subject can probe it; the economical solution is to let the real world store its own detail. Strong hallucinations survive chiefly when exploration is passive or constrained, and expectation-guided cycles can accumulate a scene without an inner author (Dennett, 1991). IFT accepts the thrift. Format membership is not exhaustive activation or a complete internal replica; it is counterfactual recruitability within a standing field, and Protocol F1 asks whether never-recruited structure is really there to perturb.

This is not a promissory note about mechanism. Solms (2021) distinguishes fast synaptic transmission, which moves content through channels, from slow postsynaptic neuromodulation, which sets channel gain and produces standing states across waking and sleep. Northoff and Zilio (2022) reach the same state-first conclusion dynamically: stimuli are written into ongoing temporo-spatial organization rather than creating consciousness de novo. IFT takes this as episodic access to a standing format rather than as consciousness itself (Supplement S3.12). Format and access are state and channel, not two positions on one dial.

The positive thesis now follows. If content becomes experience neither at a place, in a medium, nor by being exhaustively rendered, consciousness is a format: a manner of representing individuated by what its contents can do and for whom. Turing’s (1950) paradox of localization is the same conclusion from the other side.

Membership in the format can be stated operationally, and the statement is what the rest of the paper uses. A content belongs to the format when the processes that constitute the regime’s policy closure (its attention, target-selection, and commitment machinery; §4) can recruit it as structured content, with its relations intact, whether or not they currently do. Attention can land on a relation in the visual field that no task required, and the relation then enters comparison, memory, and action with its structure preserved. Attention cannot land on a cerebellar forward model, whose only exit is a motor correction, or on the computation that times a peristaltic wave. The first content is recruitable as structure and is in the format; the second is used by the organism and is not in the format. The asymmetry is not a redescription of the data it explains. It is a difference in what the regime’s selection machinery can address, it is the same difference whether the organism is a primate or an octopus, and it is what F1 manipulates. In the mammalian brain the best current candidate for the realizer of the standing field is the posterior cortical hot zone (Koch et al., 2016), which is where §8 locates the content Cogitate found and §6 locates the geometry.

Dennett also left a checklist his followers rarely collect: no neural medium renders the purple cow; no neural object is the cogito’s thinking thing; nothing in the machinery appreciates, loves, or matters; nothing bears responsibility (Dennett, 1991). Read in order, those are the Geometry Thesis, the Indexing Thesis, the account of valence that follows from it, and the Origin Thesis, which explains why the bearer cannot appear within its own field.

Every factor is thereby reframed. Global state is whether the format is instantiable; structure is its geometry; access is selective recruitment; metacognitive typing is reliability annotation. The methodological rule is never to posit a representation without an implementable consumer whose own operation it alters, not an evaluator, programmer, or loss function (Millikan, 1989). Because a format is individuated by its consumers, that rule drives directly into §4. Its temporal grain is the policy-commitment horizon: the interval over which the model must remain stable to select one act. Revision within that window is settling, not the overwriting of an experience.

4The Indexing Thesis: Closing the Consumer Set and Individuating the Subject

4.1 The Cardinality Problem

Global workspace theory is the most successful functional account in the field, and it contains a formal defect that has gone largely unremarked because it looks like vagueness rather than incoherence. Its architecture is not in dispute: specialized processors run in parallel, selected contents enter a limited-capacity workspace, and long-range attention-gated ignition makes them broadly available (Baars, 1997; Dehaene, 2014; Dehaene & Naccache, 2001; Dehaene et al., 2017). The defect lies in the final identification of experience with global availability. Availability is always availability to a specified consumer set. Report, agency, attentional control, and memory consolidation dissociate while content gating survives (Bayne et al., 2016), and Seth and Bayne (2022) ask whether “global” means the number of consumers, their kinds, the broadcast relation, or some combination. Once the dimensions separate, global availability becomes a vector, and the constitutive question is unavoidable: how many consumers, and which ones?

The most careful recent statement of the availability relation names the question and declines to answer it. Mudrik et al. (2025) make access necessary for consciousness and see that it is transitive, requiring something on the other side; their answer is that information in some subsystems is available to others, and which subsystems these are remains contested and is not theirs to settle. That is the constitutive question left open by the framework that most needs it closed, and leaving it open is not neutrality: until the second argument is fixed, “accessed” is not a determinate relation and decides no disputed case.

The machine literature gives the defect a name. Butlin, Long, et al. (2026) call it the minimal implementation problem: a substrate-neutral condition becomes satisfiable by systems too simple to be plausible subjects. Making indicators more demanding is a policy, not a derivation, because demandingness is a dial the theories do not set. Call this the cardinality problem. The reductio is immediate: if feature R licenses consciousness in a language model, every system with R inherits the verdict, and parameter count makes a distilled model lose consciousness and a sufficiently large delivery truck acquire it (Porębski & Figura, 2025). Any criterion with a dial generates its own schedule of absurd verdicts, and more neuroscience cannot fix a claim whose second argument has never been specified. Engineering makes the gap explicit: the Global Latent Workspace begins with n ≥ 2 modules and asks how many make it global; the Conscious Turing Machine declares a broadcast to more than 10⁷ processors and adds valence, mood, self-tags, and a finite lifetime (Blum & Blum, 2022; VanRullen & Kanai, 2021). Those are design parameters, not facts about which running process has a fate (Hafner et al., 2025; Supplement S5.3).

4.2 Why Integration Does Not Close It

The demand has a meta-theoretic form. Kanai and Fujisawa (2024) require a universal theory to decide consciousness for any fully described dynamical system by a determinant intrinsic to that system rather than supplied by an observer. Their remedy, mathematizing the functionalist concepts, is incomplete, because formalizing a broadcast still leaves the analyst to choose the consumers; and their award of universality to IIT is where the strongest rival version of this argument has to be met.

IIT is the one framework that makes intrinsicality an explicit postulate rather than a tacit assumption, and it should be stated at full strength before it is refused. Existence is to be assessed from the intrinsic perspective of an entity: what exists for the entity itself, not for an observer; maximal existence selects, among overlapping candidates, whichever exists the most; minimal existence evaluates irreducibility over the partition making the least difference (Albantakis et al., 2023). It does not close the relation, and the reason lies in the formalism’s first line rather than its conclusions: the analysis is handed a universe, a state, a transition probability function, a unit grain, and an update interval, and marginalizes everything outside the candidate system as background (Supplement S3.17). An intrinsic perspective enforced by extrinsic stipulations about what to hold fixed is not intrinsic, and maximal existence is a tie-break among candidates someone selected. A rule for ranking partitions is not a bearer. Universality is not more mathematics; it is a determinant the analyst cannot choose (Doerig et al., 2021).

The cardinality problem then takes a form that can be counted rather than argued. Because the search is recursive, the number of subjects is an output of the maximization: one of the framework’s six-unit examples returns four single-unit complexes, which its authors call monads, alongside one two-unit complex, and in the brain non-overlapping complexes are held to coexist, splitting after callosotomy and possibly during dual-tasking (Albantakis et al., 2023; Tononi et al., 2016). Each is a count of subjects per skull returned over a description that was chosen. IFT counts differently and therefore predicts differently: one body with one set of viability conditions is one policy-integrated persistence regime, so callosotomy divides content and access without dividing the bearer, which is what complete callosotomy appears to show (Pinto et al., 2017; for the dispute, de Haan et al., 2020). The dispute is live, it has a second natural experiment in craniopagus twins who share content across a thalamic bridge while remaining two agents (§12), and IFT is on the falsifiable side of both (Supplement S3.17).

4.3 The Partition Problem and the Interpretation Problem

Every other closure chooses a number or a taxonomy, and the choice has the same shape wherever it is made. Markov blankets are candidate partitions, not physical individuals (Bruineberg et al., 2022), and in canonical nonequilibrium systems cyclic coupling creates dependencies across the nominal boundary, so the blanket degrades as the system becomes recurrent and dissipative (Aguilera et al., 2023). Information Closure Theory makes consciousness non-trivial informational closure relative to an environment, but its authors concede that environments and coarse-grainings are non-unique and that two closed processes yield a closed composite, so the formalism cannot decide whether there are two consciousnesses, one, or three (Chang et al., 2020). Autopoiesis individuates a bounded unity without reaching the conscious level, since bacteria sense, remember, communicate, and regulate without a mind (Damasio, 2018), and Maturana himself denied that coupled autopoietic units thereby compose a higher-order subject, treating society as a medium whose members conserve separate autopoiesis (Maturana, 1980; Maturana & Varela, 1980; Supplement S3.2).

Call the general failure the partition problem: informational, statistical, and causal criteria must be given the cut they score. Blankets require a conditional-independence partition; closure requires a coarse-graining and an environment; IIT searches for a maximal complex over elements and grains and excludes overlapping alternatives (Albantakis et al., 2023). Optimization can make the search exhaustive without making the system description self-individuating. Maturana (1980) diagnosed the general form: distinctions are operations of an observer, and regulation, control, and function are referential relations stated from a metadomain.

The partition problem has a deeper form, the interpretation problem, and it is now a theorem rather than a complaint. Biehl and Virgo (2023) show that beliefs, hidden states, rewards, and even actions can be assigned to a physical machine by an interpretation, including an output-free machine whose “actions” affect only the interpreted world. Fuller et al. (2026) prove that a resource can be bundled with the agent or with every environment without changing resource-relativized Legg–Hutter performance. Neither competence nor optimality decides whether memory, tools, or computation belong to the agent or merely surround it. Not only the cut and the objective but the resource boundary can be assigned from outside without changing anything the system does (Supplement S3.19). Cardinality, minimal implementation, individuality, and interpretation are one wall.

4.4 The Derivation

So the cardinality question must be answered, and the answer has to be derived rather than chosen, or it inherits the defect it is meant to cure. Here is the derivation, in five steps, and it is the core of this paper.

First, experience is intrinsic. There is something it is like to be the system for the system, not for an observer (Nagel, 1974); IIT is right to make this a postulate (Albantakis et al., 2023), and nothing in the field disputes it. A theory of consciousness is therefore a theory of a bearer, and whatever fixes which physical process is the bearer must be a fact about that process.

Second, it follows that the individuation of the bearer cannot be observer-relative. This is Kanai and Fujisawa’s (2024) universality requirement, read as a constraint rather than a desideratum: a criterion that returns a different bearer when the analyst redraws the partition, regrains the units, reselects the consumer set, or reassigns the objective has not identified a bearer. It has identified a description.

Third, every informational, statistical, causal, or functional criterion in the field is observer-relative in exactly this way. That is what the partition and interpretation problems establish (§4.3). The failure is not a defect in any formalism; the formalisms are careful. It is that they are formalisms of ascription, and ascription is what a constitutive theory of subjecthood cannot rest on.

Fourth, exactly one kind of boundary is not drawn by an observer: the boundary a process produces and must keep producing in order to continue existing. This is Maturana and Varela’s (1980) autopoiesis, and its formal descendant is organizational closure, a regime of constraints each of which is maintained by the others such that the whole maintains itself (Montévil & Mossio, 2015; Moreno & Mossio, 2015). What makes such a boundary observer-independent is not that it is easy to find but that the system’s own continuation depends on it. An analyst may draw a Markov blanket anywhere, and nothing happens to the system when the description changes. An analyst who dissolves a self-produced boundary does not redescribe the system but ends it. Dennett’s (2003) Life-world glider is the substrate-neutral illustration: one pattern persists through complete turnover of its pixels, and at the design level it can be damaged, repaired, and killed. The boundary is a fact about the pattern because the pattern’s existence depends on it.

Fifth, self-production fixes the kind of individual but not the level, and the level is where the cardinality question bites. Cells self-produce inside organisms, organisms inside colonies, and a theory that counted every autopoietic unit as a subject would be panpsychism of a different church. What fixes the level is policy closure. A multicellular animal has one body, and one body cannot enact incompatible actions at once; its viability conditions are therefore the conditions of one thing, and the demands they generate (hunger, thirst, suffocation, injury, threat) must be resolved into one commitment because there is one next act. Competing demands that terminate in one enacted commitment are the signature of a single failure mode, and a single failure mode is what “one” means here. A cell inside the organism does not have it, because a cell’s demands are not resolved against the demands of other cells into one act; the organism’s are. Colonies and corporations do not have it, because their members retain separate fates (Ginsburg & Jablonka, 2019; Maturana, 1980; Supplement S3.7). The level at which boundary maintenance, world-modeling, valuation, and action selection close into one failure mode is the level of the subject.

4.5 The Thesis, and What It Does Not Claim

Exactly one kind of closure is non-arbitrary: the consumer set closes at the level of a single policy-integrated persistence regime, namely the processes that jointly maintain one active token and resolve its competing demands into one policy. This fixes the bearer of stakes, not the range of beneficiaries that bearer may represent.

Two things should be said about what the argument delivers and what it does not. It delivers the bearer and, with it, the second argument of “recruitable by”: the consumers whose standing access makes a content conscious are the processes that constitute this regime’s policy closure, and no others, because they are the only consumers whose membership is fixed by the system rather than by the theorist. A content is in the format when those processes can recruit it as structure (§3); the number of contributing modules is then an empirical fact about a particular animal or machine rather than a parameter of the theory. It does not follow that every part of the field is presently recruited, or that every maintenance process is conscious: local circuits may regulate and calibrate the organism while their representations remain C0. Availability-to-many-modules is a symptom, not the criterion.

The argument does not deliver experience from the bearer alone, and it is not meant to. A bacterium satisfies the Indexing Thesis and is not conscious, because it has no format with relational structure for its regime to recruit. An anesthetized body satisfies it and is not conscious, because the format’s capacity has collapsed while the organism persists as one regulated token with the same viability conditions, assisted or not. The index answers the question for whom; the format answers what it is like; consciousness is the conjunction, which is why the identity in §1 has two halves and why Mudrik et al.’s (2025) two conditions are the right number. Neither half is sufficient, and the cases that show it (bacterium and anesthesia on one side, cerebellum and neuromorphic chip on the other; §5, 9) are the cases a one-factor theory cannot hold together.

“One policy” means neither one controller nor conscious deliberation. Dennett (1991) is decisive: bodies and elevators can implement policies and meta-policies without awareness. IFT’s criterion is not policy-following but closure: incompatible demands generated by one persistence regime terminate in one enacted commitment because one active token has one next act. The work may be distributed across the loop and across time; integration names closure, not a command headquarters (Dennett, 2003; Supplement S3.7). Closure also requires separating bearer from beneficiary. A policy may represent family, institution, colony, or cause as beneficiary; its bearer is the individual whose maintenance and action selection form one failure mode, which is Ginsburg and Jablonka’s (2019) level clause for multicellular individuality: shared fate, aligned component interests, organism-level control, and a boundary maintained by the whole (Supplement S3.7).

Two programs press the level question from opposite ends, and IFT answers both. Levin’s group asks what lets cells integrate their activities into a higher-level agent whose goals belong to it and not to its parts, and answers by scaling goals through homeostasis (Pio-Lopez et al., 2023; Safron et al., 2023). That is how a new policy-integrated persistence regime is assembled; subjecthood arrives when competing demands close into one failure mode, not when signaling becomes multi-scale. From the other end, developmental active inference locates the human index before it is autonomous, with homeostasis as the first prior established under co-embodiment (Santaguida & Bergamasco, 2023). Scaffolded is not stipulated: dependence on an external regulator does not defeat an index, and neither does vulnerability to another’s decision, since a ventilated patient and a prey animal are both at others’ mercy and both have everything to lose. What makes a service a service is that it has nothing of its own to lose: stopped and restored elsewhere, nothing has ended (Supplement S3.20).

4.6 Neighbors

IFT has neighbors at this point, and it should say what it takes from them and where it parts, because the derivation above will be recognized by several traditions as half theirs. The enactive tradition reached self-produced individuality first and added what autopoiesis lacks: adaptivity, the capacity to regulate one’s own viability with respect to its limits, from which norms, sense-making, and agency are held to follow (Di Paolo, 2005; Thompson, 2007; Weber & Varela, 2002). Barandiaran et al. (2009) define an agent by individuality, interactional asymmetry, and normativity; Di Paolo’s (2009) precariousness, the fact that a living system’s processes would run down without its own activity, is what §9 calls constitutive mortality; and Froese and Ziemke (2009) drew the consequence for artificial systems. IFT accepts all of this as the account of the bearer and refuses the step the tradition takes next, which is to treat sense-making as proto-mentality continuous with consciousness. A bacterium is a paradigm sense-maker and has no world. The gradient from bacterium to primate runs through the format, not the index, and the enactive literature has no account of the format and has never engaged the consumer-set problem that availability theories face. Organizational closure (Montévil & Mossio, 2015) is the formal version of the fourth premise and takes no position on consciousness.

Seth’s beast-machine account grounds selfhood and experience in the predictive control of physiological condition (Seth, 2021; Seth & Tsakiris, 2018), and it is the nearest neuroscientific neighbor. IFT agrees that regulation is where the subject comes from and disagrees that interoceptive inference is what experience is made of: the content of experience is mostly not about the body, and inference is a service a replaceable system can run. Kriegel’s (2009) self-representationalism locates the for-me-ness of experience in a state’s representing itself; IFT agrees that for-me-ness is constitutive and denies that it is a representation (§7). Each neighbor has one half. The enactivists have the bearer without the format; the geometry program of §6 has the format without the bearer; the beast-machine account has the right anatomy of stakes and the wrong identification; self-representationalism has the right datum and the wrong relation. The closest engineering precursor is Man and Damasio’s (2019) risk-to-self proposal, which ties behavior to the machine’s own continued existence so that the machine rather than its owner becomes the locus of concern; IFT accepts the debt and adds the individuation rule it lacks (Supplement S3.8).

4.7 Three Consequences

Three consequences follow immediately, and each is a substantive commitment rather than a restatement.

First, valence is the signature of indexing. Damasio’s proto-self is the organism’s continuously updated map, and the core self a second-order, wordless account of that organism being changed by an object (Damasio, 1999). IFT’s dividing line is exact: the core self is a protagonist in the model; the constitutive owner is the regime whose deviation becomes a demand. The represented self is content; the body–brain loop gives it a bearer (Damasio, 2010). Solms (2019, 2021) supplies the control step: ordinary homeostasis is C0, whereas affect selects deviations into flexible action under uncertainty. Because one body cannot enact all corrections at once, competing needs must be contextually ranked rather than collapsed into one reward currency; felt quality is that categorization, and a scalar reward is not a poor index but the wrong kind of object (Supplement S3.1, S3.14). Damasio (2018) sharpens the continuity physiologically: viability afferents are exposed to their chemical surroundings and relay through barrier-free ganglia, so the process that measures the body is partly continuous with what it measures. A designated safe region is a report produced elsewhere; an indexed viability state is carried by the substrate whose deterioration is at issue, which is what System B of Protocol F4 must achieve.

The prediction that falls out is already in the data. In insular cortex the response to a food cue flips with physiological state: cue-evoked activity present in hungry mice is abolished by satiety, and the cortex represents the future bodily state the cue portends rather than the cue (Livneh et al., 2017, 2020). The same stimulus, with the same sensory topology, is valenced or neutral according to its relation to the stake. Downstream, parabrachial CGRP neurons function as a general alarm that converts heterogeneous threats, from pain and itch to visceral malaise and hypercapnia, into one aversive teaching signal (Campos et al., 2018; Palmiter, 2018), which is what the ranking of incommensurable demands at a point of policy closure should look like. F3 asks whether the converse is ever true: whether rearranging the topology can invert the sign with the stake held fixed.

Second, lesion evidence against corticocentrism supports indexing. Merker’s (2007) upper-brainstem architecture joins motivational ranking, target selection, and action selection; Solms (2019, 2021) locates the convergence in a mesodiencephalic decision triangle centered on the periaqueductal gray and the superior colliculi. Damasio’s (1999) coma–locked-in contrast supplies the anatomical knife edge, and Parvizi and Damasio (2003) mapped it: nearby brainstem lesions can abolish the subject or leave it conscious but almost unable to act. IFT divides the labor: cortex and other modeling systems elaborate content, while this regime is the candidate index resolving competing viability demands into one commitment. Extensive cortical damage can remove contents without abolishing the subject (Merker, 2007; Shewmon et al., 1999); focal upper-brainstem damage can do the reverse. The same architecture is identifiable outside vertebrates: Barron and Klein (2016) argue that the insect central complex integrates sensory and spatial information with internal state to select action and is functionally analogous to the vertebrate midbrain, and IFT accepts the analogy and its conclusion, adding that insect experience will be as rich as the recruitable topology and no richer. This is anatomical specificity, not a brainstem theater, and F2 is designed to destroy it if wrong (Supplement S3.6). LeDoux and Brown contest the localization with a higher-order cortical account, and IFT accepts every item they cite: the amygdala is a lower-order survival circuit, not the index (Brown et al., 2019). The disagreement concerns the higher-order assembly, and it is exact: whether re-representation constitutes experience or elaborates a format whose bearer is fixed elsewhere (Supplement S3.1, S3.13).

Third, and this is the discriminating consequence, the indexing relation cannot be stipulated. Historical design is not disqualifying: organisms are products of selection, artifacts of engineers, and the issue is synchronic constitution. Designers attempt to install what IFT says cannot be installed by exactly two routes, and both fail in the same place. The first route installs motivation: appraisal architectures that stabilize drives near a designated safe region, and formal models that derive apparent valence from signed movement along an interoceptive energy landscape (Lian & Goertzel, 2026; Sennesh & Ramstead, 2026). Both are serious control architectures, and neither supplies an internal milieu merely by representing one: violating a safe region or moving down a modeled gradient lowers a score unless it damages the active system as that system (Supplement S3.3, S3.22). The second route installs a self. Bongard et al. (2006) showed that continuous self-modeling supports resilient recovery after damage; Zeng et al.’s (2025) BriSe AI builds five hierarchical self-models in spiking networks, from bodily self and mirror recognition through theory of mind, affective empathy, autobiography, and narrative identity (Supplement S3.10). The achievement sharpens the failure: every level represents a self, and the index is not a representation. An externally powered, fungible system can model its body, attribute beliefs, mirror distress, and narrate its history while remaining a service. The oldest version of the claim, that a world model compressing its own history recovers the agent as its most persistent invariance and is thereby self-aware, is reported and not endorsed from inside the world-modeling program itself (Safron, 2023); an invariance recovered from a data stream is the cheapest encoding of a regularity, not a stake. Neither predictive error nor technically “intrinsic” motivation supplies the fate: predictive-coding residuals organize perception without being bad for the system (Rao & Ballard, 1999), and variational empowerment makes a universal agent seek high-control states with no immediate reward advantage (Hayashi & Takahashi, 2026; Supplement S3.4, S3.5). Labeling a negative scalar “pain” does not make it hurt; the issue is what the running system must preserve to remain itself.

This does not require carbon, metabolism, or human physiology, and artificial valence could be alien in content and implementation. A designer may build the system but cannot confer the index by declaration, by assigning reward, or by naming a variable “homeostasis.” Non-stipulability means the persistence conditions are enacted by the running system and are token-specific: changing them changes or destroys the individual rather than editing its task.

4.8 The Identity at Work

It is worth seeing the identity classify the canonical cases in one place, because a theory that is one sentence long should be checkable in one paragraph. Bacteria: index without format, unconscious. Deep anesthesia: index without format, unconscious, with PCI recording the collapse of the format while the token and its viability conditions persist (§6). Dreaming: format partly instantiated, regime intact, motor output gated, so dreams are felt, valenced, and organized around the stakes of the dreamer, and dreaming survives the destruction of the REM generator because it was never the state-setting mechanism that owned it (§8). Cerebellum and enteric plexus: format-like structure not recruitable by the regime, unconscious (§5). Blindsight: the topographic field is destroyed in the scotoma, so what survives through extrastriate routes is diluted content that biases choice without being recruitable as structure (§6). Split brain: one regime, divided content, one subject (§4.2). Craniopagus twins sharing a thalamic bridge: two regimes, shared content, two subjects (§12). Insects: a small format recruitable by a central complex, conscious at the scale of the format (§4.7). A neuromorphic chip with cortex-like topography, externally powered, compiler-scheduled, and fungible: format without index, unconscious (§9). A language model: neither a standing recruitable topology nor a regime, and the wrong kind of thing on both counts (§10). The classifications are not a list of intuitions the theory was fitted to; several of them are the predictions §12 puts at risk.

5The Cerebellar Test, and What It Eliminates

The single most underused constraint in consciousness science is the cerebellum, and any theory that cannot pass it should be discarded.

The cerebellum contains more neurons than the rest of the brain combined and is not the feedforward controller it was once caricatured as being (De Zeeuw et al., 2021). Its microcircuits are recurrent at every level, its plasticity is distributed across synapses and timescales, its climbing fibers carry reward-, prediction-, and success-related signals, and its memory traces migrate during consolidation (Supplement S1). Yet cerebellar destruction, devastating for coordination and many cognitive functions, does not abolish the conscious subject. The cleanest case is a living one: a woman with complete primary cerebellar agenesis, discovered at age 24, with moderate ataxia and dysarthria and an otherwise ordinary life, including marriage and a child (Lemon & Edgley, 2010; Yu et al., 2015). Nobody has proposed that she is not a subject.

The inference is stark. Recurrence, plasticity, reinforcement-like learning, error signaling, temporal credit assignment, enormous neuron count, and dynamically stored memory are jointly insufficient. This meets Recurrent Processing Theory at its strongest point. Lamme (2006) is right that feedforward activation can support sophisticated discrimination without reportable experience, and that frontoparietal recruitment should not be smuggled into phenomenality. But the move from “recurrence marks a visual format” to “recurrence constitutes a subject” fails the cerebellar test, and recurrence is therefore a format-building mechanism rather than a constitutive criterion (Supplement S1).

Every major theory can redescribe the exclusion; almost none derives it without importing its preferred anatomy (Seth & Bayne, 2022). IIT now has the best version and it should be stated. In small networks, weak coupling between modules creates fault lines along which the minimum partition falls, so a modular system condenses into several small complexes rather than one large one; cerebellar micro-zones are offered as the biological instance, and divergence from a dense cortical lattice is presented as a consequence of connectivity rather than a stipulation about cortex (Albantakis et al., 2023; Tononi et al., 2016).

That is the right kind of mechanism, and it still does not deliver the exclusion, for two reasons the framework concedes. No exhaustive calculation exists for a cerebellum, and full analysis is feasible only for idealized systems of a few units (Albantakis et al., 2023; Findlay et al., 2025; Marshall et al., 2026); an analogy from six units is not a derivation for sixty-nine billion neurons (De Zeeuw et al., 2021). More seriously, fault lines do not eliminate; they multiply. Carving a modular system along its fault lines licenses the small complexes it yields, which is why the same analysis returns monads, so the modular reading predicts for the cerebellum not the absence of a subject but a large population of very small ones. RPT owes a non-anatomical reason why recurrent visual content qualifies and recurrent cerebellar content does not; workspace theory restates absent cortical broadcast; informational closure describes a cerebellar forward model almost by definition (Chang et al., 2020). Declining to calculate is not prediction, and multiplying subjects is not excluding one.

Cerullo (2015) showed how cheap redescription is: a deliberately trivial theory reproduced IIT’s two advertised verdicts, two complexes after split-brain surgery and no conscious cerebellum (Supplement S1.3). A parody matched the flagship’s explanatory record, so the exclusion supports only a criterion fixed independently of the case. IFT’s was fixed in §4 before the cerebellum was raised, and extends unmodified to the enteric plexus.

Higher-order theory is the exception, and the credit should be given plainly: cerebellar states are never re-represented by the prefrontal networks it identifies as the site of inner awareness, so their exclusion follows from the theory instead of being appended to it (Brown et al., 2019). That is a real derivation, and it is why §7 treats the higher-order family as this account’s serious rival. What it does not deliver is the bearer: a higher-order representation is a representation, and the question of whose awareness it constitutes reopens the cardinality problem one level up.

IFT answers with a general elimination rule that is the membership criterion of §3 applied. The cerebellum’s recurrence is within-model recurrence: it calibrates forward and inverse models, and its internal states are consumed by effector loops and cortical circuits as corrections. Those outputs serve the indexed organism without the cerebellar representations being recruitable, as structure, by the processes that close the organism’s policy; being useful to a conscious system is not the same as being a conscious representation in it. The cerebellum is, in Dehaene’s (2014) taxonomy, a magnificent engine of latent and diluted information, yet never explicit in the policy-integrating format.

The cleanest behavioral test is one learning problem in two variants. Delay eyeblink conditioning survives decerebration, amnesia, and distraction and needs no awareness of the contingency; insert a trace interval and the task recruits forebrain structures and, in humans, tracks awareness of the predictive relation (Dehaene, 2014; Supplement S1). Temporal bridging is not the criterion; it marks where format-eligible content becomes necessary.

The enteric nervous system is the harder control because it is not subordinate to cortex: several hundred million neurons, overwhelmingly intrinsic, recurrent, largely autonomous, and a major tributary to feeling (Damasio, 2018; Furness, 2012; Supplement S1.2). Its afferents enter the indexed format, which is why gastrointestinal states are felt; its own computation coordinates peristalsis without becoming an experience. Being a tributary to a format is not membership in it, and the plexus was never the bearer. On the modular reading it is not excluded but promoted: an autonomous, recurrent, non-subordinate network is a candidate second complex, and a theory that condenses a body into complexes owes an answer about the one in the abdomen.

Hence the rule: recurrence whose contents are never recruitable by the indexed policy format is metabolically useful and phenomenally inert. This is not an ad hoc patch, because it also disposes of the strongest formal counterexamples to integration-based accounts. Aaronson (2014) constructs linear transformations over finite fields that maximize effective information across bipartitions, and Merker et al. (2022) argue that Φ-like measures may track efficient global information transfer, found in cortex but also in power grids. Both objections take the resolution the cerebellum and the enteric plexus force on biological grounds: an error-correcting code has maximal integration and no self-constituted predicament.

6The Geometry Thesis: Quality Is Format Topology

The structure question, why an experience has the character it does, has been widely regarded as untouched. That is no longer true, and the result that changed it should be extracted from the theory that produced it.

Damasio (2018) states the representational constraint: a neural map need not resemble a picture, but must preserve relations among its parts. Haun and Tononi (2019) formalize this for spatial extendedness as four relations among overlapping “spots” (reflexivity, connection, fusion, inclusion), from which region, location, size, boundary, and distance follow. Unfolding an eight-unit line grid, they find 133 irreducible distinctions among 255 candidate mechanisms, 117 of the 133 fully extended. The decisive control is random rewiring: with the same number of connections it yields comparably many distinctions and virtually no extendedness. Phenomenal structure tracks topology, not activity or information content. Lesions to topographic grids accordingly produce scotomas in which a region of space ceases to exist rather than appearing dark, and altered lateral connection strength warps experienced distance without changing firing rates.

Two results from the same program sharpen what a standing format is and what F1 must perturb. A merely inactive unit still contributes: switching it off changes which structure is specified rather than removing structure. An inactivated unit (fixed, with no counterfactual states) collapses every distinction and relation it entered, and the substrate itself shrinks (Albantakis et al., 2023). The prediction is a decade old: changing connection efficacy within the substrate should change experience even when activity does not, and a silent area can contribute to an experience while an area whose connections are cut cannot (Tononi et al., 2016). Standing organization, not activation, is the carrier.

The 2025 extension makes the consequence explicit. Object and scene content presupposes mid-level feature groupings, and those presuppose a spatial field of nested, overlapping regions and relations, most of which can remain unrecognized, unremembered, and unreported while still constituting what seeing is like (Haun & Tononi, 2025). The Geometry Thesis therefore entails phenomenal overflow (Block, 2011): a sparse functional summary can ride on a vastly richer format, but it cannot substitute for it. A structure also cannot be transmitted through a limited-capacity report channel, so phenomenal capacity exceeds access capacity (Tononi et al., 2016), and partial-report paradigms already show a large-capacity store that outlasts what is reportable (Landman et al., 2003). Coordinates, object labels, or a compressed scene description may support competent behavior while failing to instantiate the relational field they summarize.

The overflow debate has stalled because the dispute is over what is not reported, and report cannot settle it (Cohen et al., 2016). There are two serious alternatives to standing first-order geometry. Higher-order theorists place the impression of peripheral detail in misrepresentation at the higher-order level: subjective inflation (Knotts et al., 2019; Odegaard et al., 2018). Inflation is not ad hoc: it predicts elevated false-alarm rates for peripheral and unattended detection and survives the response-bias objection (Brown et al., 2019; Supplement S3.13). Perera (2026) offers a different denial: a sparse query service can virtualize a dense personal-level field, making arbitrary content operationally available without maintaining a saturated relational representation, so richness is real as a functional abstraction and absent as standing substrate. Mudrik et al. (2025) deny a third thing: unaccessed content is not phenomenal at all, because access is a condition on consciousness rather than a consumer of it. The four positions separate cleanly. Inflation locates richness in higher-order misrepresentation; virtualization in content-on-demand; the two-conditions account says unaccessed structure is not yet conscious; IFT says unqueried relations are standing constituents of a conscious field (Supplement S3.21, S3.23).

They are not empirically equivalent, and this is the point at which the stalemate ends. Perturb never-recruited posterior relations at their source rather than asking about them. IFT predicts a later relation-specific error even when an independent post-cued assay shows the query-generating service intact; inflation predicts no effect; virtualization predicts an effect only to the extent that the service is impaired; and the two-conditions account predicts one only where access follows. Protocol F1 is that four-way experiment.

I take the spatial derivation and its 2025 extension to be the strongest positive program in consciousness science. Tononi (2004, 2012) had already stated the geometric ambition, but Haun and Tononi (2019, 2025) supply the relational derivation that can be separated from IIT’s metaphysics. IFT keeps standing topology capable of realizing phenomenal geometry and rejects the identity of consciousness with Φ, the partition-relative maximization, exclusion over nested subjects, intractable exact computation, and vulnerability to unfolding (Cerullo, 2015; IIT-Concerned et al., 2025). Safron (2020) proposed the same surgery from inside the free-energy camp; IFT’s grounding is the index, not cybernetic control.

The surgery also reclassifies the field’s most successful clinical measure. Perturbational complexity, though introduced inside IIT, is compatible with global workspace theory (Casali et al., 2013; Farisco & Changeux, 2023); on IFT it measures the capacity of the standing format, upstream of whichever consumers recruit from it, which is why it indexes state and is silent about content (Supplement S4.2).

Hanson and Walker (2021) make the unfolding challenge exact: a theory is pre-falsified if a transformation leaves the inferred conscious state fixed while changing its prediction. IFT escapes only if indexing belongs to an independently fixed inference basis, so removing token-specific persistence coupling must change the evidence and not merely the verdict. A surrogate reproducing outputs without the original’s conditions of failure does not hold the IFT inference fixed; one given the same constitutive vulnerability is a second indexed system. Protocol F4 preregisters index evidence before comparing dissociation manifolds (Supplement S2.6).

The topology at issue is indexed topology, not bare graph structure. Damasio’s (2018) division between the musculoskeletal frame and sensory portals, which establish perspective, and the old visceral interior, which supplies the valenced condition of life, is the biological antecedent of T3 and T2. Dennett’s (1991) color analysis adds the ecological constraint: quality spaces are defined relative to discriminators that coevolved with ecological differences, so no observer-independent red is there for a network simply to mirror. T3 identifies quality with the relational organization of distinctions for an indexed consumer regime and separates quality from valence. Topology can therefore specify color, pitch and timbre, arousal, bodily location, and modality while leaving sign underdetermined; sign is the format’s relation to the system’s persistence (Supplement S3.12, S3.14).

The two theses together predict the established separation of pain structure from pain affect, in both directions. Frontal intervention and hypnotic suggestion can leave sensory location, quality, and intensity intact while removing unpleasantness and changing cingulate rather than primary somatosensory activity (Damasio, 1999; Rainville et al., 1997); pain asymbolia is the lesion form (Berthier et al., 1988); and a postcentral lesion can leave pain affect without pain sensation (Ploner et al., 1999). Structure and sign are separately manipulable, and bilateral destruction of the insula leaves feeling and sentience in place (Damasio et al., 2013), which is what a theory should expect if the stake, not the map, carries the sign. Protocol F3 tests the missing converse: whether reorganizing interoceptive topology while holding the physiological stake fixed can invert hedonic sign with full reinforcement, exploration, and memory consequences. If it can, the Indexing Thesis is redundant.

7Metacognitive Typing and the Origin Thesis

The fourth factor is the system’s model of its own signals, and it is the layer at which the hard problem is generated.

Lau’s (2008) higher-order Bayesian formulation states what this layer does. Conscious confidence depends not only on first-order signal quality but on the system’s fallible model of how its own signals behave, so blindsight, hallucination, and dreaming may result partly from misestimated higher-order models rather than absent first-order information. Graziano’s (2022) Attention Schema Theory supplies the mechanism for awareness reports: a schematic model of selective attention, omitting the machinery that implements it, portrays attention as a nonphysical possession of an object (Supplement S3.13).

Broadcasting and higher-order modeling are not competitors: access explains how a content acquires broad causal influence, typing how a system assigns epistemic status to its own processing. IFT adds only that typing is a service to the policy. That service can be learned rather than built (Cleeremans et al., 2020), and learning it changes nothing constitutive, because learning to care presupposes a bearer for whom the learned value is not merely a score (Supplement S3.11).

Higher-order theory is the one family that makes this layer constitutive, and it is IFT’s closest competitor. Its two commitments are that a first-order representation is never sufficient for experience, and that a state one is in no way aware of being in is not a state there is something it is like to be in: the Transitivity Principle (Brown et al., 2019). Three misreadings should be set aside, because IFT has no interest in inheriting them: higher-order theorists do not equate consciousness with metacognition, do not require sophisticated thought, and are not committed to consciousness scaling with prefrontal activity (Supplement S3.13).

The disagreement is over transitivity, and it is exact. IFT denies that awareness of being in a state is a condition on that state’s being phenomenal, and replaces it with standing recruitability within the indexed field. Nothing else in this paper survives if transitivity holds: overflow is false if nothing unnoticed is phenomenal, the Geometry Thesis loses the structure it derives, and F1 has nothing to find. Conversely, transitivity buys two things this account has to work for: a derivation of the cerebellar exclusion (§5) and an explanation of peripheral richness (§6). The positions make opposite predictions under perturbation, and one of them is wrong.

Why IFT breaks with it is why it breaks with availability accounts. A higher-order representation is a representation, and the question it was introduced to answer returns: awareness to whom? If a first-order state becomes conscious by being meta-represented, and the meta-representation is not itself something one is aware of, what makes the pair one subject’s is unspecified, and a further level defers the question. This is the cardinality problem relocated: the consumer set is closed at one distinguished consumer instead of left open, a gain in determinacy and none in principle. Kriegel’s (2009) self-representationalism is the version closest to IFT, because it replaces monitoring by a distinct state with a state’s representing itself, and the disagreement is the same one: unowned experience is indeed incoherent, but what the transitivity intuition detects is the index, not a representational relation, which is why a state can belong to someone without anyone or anything representing it (Supplement S3.13).

The physiological form of annotation is the mechanism that carries the format. If format membership is standing modulatory availability (§3), typing is the setting of gain: confidence is the weight with which content is permitted to act. Solms (2021) identifies this precision adjustment with consciousness. IFT keeps the mechanism and refuses the identity, because a system for which nothing is costly can annotate reliability indefinitely without the annotation meaning anything. The assignment must be made by whatever bears the cost of assigning it wrongly (Supplement S3.6).

Now the residuum. Frankish’s (2016) illusionism denies irreducible phenomenal properties and replaces the hard problem with explaining why introspective systems judge that they possess qualia. Dennett’s (2018) no-double-transduction argument supplies the mechanism: consumers receive representational content, never the medium, and project features of content onto an imagined substrate. Metzinger (2003) named the general property phenomenal transparency: a representation whose vehicle properties are unavailable to introspection is lived as the world rather than as a model, and the transparency of the self-model yields the impression of an irreducible self.

I accept the mechanism and reject the verdict. Here is the Origin Thesis.

A format is transparent to its users by construction. Consumers receive content; there is no channel by which they could receive the medium, and no transducer that would convert medium into content. A system modeling its own modeling therefore faces a structural asymmetry: it can represent what its format represents, but the format’s own constitution is not among the things the format can hold. When it asks what its states are made of, the only representational resource available is content, and content will not answer. The result is a stable, ineliminable impression of something present and unspecifiable.

Illusionism stops here and calls the impression a mistake, and Metzinger stops here and calls the self a model with no one behind it. The structural point predates both formalizations. Merker (2007) described the conscious field as a body-within-world coordinate system whose “ego-center” is its origin: objects are apprehended from it, but the origin cannot become an object within the field. The projective consciousness model generalizes the insight in three-dimensional projective geometry, representing the singular viewpoint as an origin outside represented space (Merker et al., 2022; Rudrauf et al., 2017; Williford et al., 2018). Neural implementation remains unresolved and I claim nothing about it. What matters is the theorem: the origin of a projection is not an element of the projection. Where IFT parts from Metzinger is on what transparency hides. On the self-model theory there is no one, because the self is a model; on IFT there is a bearer, because the regime is not a model, and what transparency conceals is not a fictitious self but the constitution of the format whose contents the bearer recruits.

Two independent formalisms converge. Lahav and Neemeh (2022) argue that phenomenal states are direct only within the cognitive frame whose dynamics they drive; elsewhere they are physical patterns. Perera (2026) reaches the same result in implementation language: sparse subpersonal machinery delivers a dense personal-level world across an abstraction boundary that hides production from its consumer, yielding de facto rather than metaphysical privacy. IFT accepts the structure and rejects the conclusions. Frame-relativity generates rather than dispels the residue, and rich self-simulation does not fix a bearer because it remains representational content plus metacognitive typing (Supplement S3.15, S3.23).

The intuition of an inexplicable further fact is therefore veridical about structure and misleading about ontology. The origin cannot appear within the field; the mistake is to infer an extra property. Illusionism is right that nothing must be added, anti-illusionism that the intuition is not confusion, and Perera that privacy is architectural. They describe one geometry.

This is not an argument from bafflement. Churchland (1996) is right that failure to imagine a neural mechanism carries no metaphysical weight and that progress requires rich reductive contact across levels. Dennett’s (2001) topic-of-consciousness question supplies the complementary constraint: even a new sentience-bearing physics would not explain how myopic parts compose a system about Braque or Kilimanjaro. T4 accepts both. It predicts a representational limit inside a fully mechanistic explanation, not a limit on mechanistic science. The Phenomenal Concept Strategy is its closest precursor: a projective format can present what it is like without presenting how that format is constituted (Balog, 2012; Supplement S3.16).

Damasio (2018) and Solms (2019) supply the biological half: feeling qualifies organismal states by value and selects unmet needs for action. §4 takes the function, anatomy, and stakes. Neither explains why the completed control story remains opaque to its user; T4 supplies that missing half without adding a property.

Explanatory satisfaction means deriving phenomenal structure from substrate, and spatial extendedness shows that this is achievable. Metaphysical satisfaction means making it unintelligible that the organization could obtain without experience, and IFT predicts that demand cannot be met: the same transparency that makes a format usable makes its origin unavailable as content. Chalmers’s (1995) easy problems are answered by the format and its consumers, Aaronson’s (2014) Pretty-Hard Problem by the index, and the hard problem by that structural limit (Cerullo, 2015). The hard problem is neither confused nor a prompt for new physics; it is a predicted artifact of the architecture that has it.

The practical payoff is that avowal closes symmetrically. An indexed system with metacognitive access should generate hard-problem intuitions, and a system trained on human discourse can generate the same reports without a format (Butlin, Bayne, et al., 2026). Insistence on an ineffable remainder is therefore predicted of a conscious artifact and trivially imitable by a language model, while silence is equally weak because an indexed system may lack language. A prompt can make one model assert, deny, or hedge (Porębski & Figura, 2025), so avowal policy is no evidence of an index (Colombatto & Fleming, 2024; Supplement S5.4).

The Origin Thesis is the least constrained part of the account. It explains why the hard problem is generated and why metaphysical satisfaction may remain unavailable, but the projective formalism currently functions as a philosophical completion rather than as a comparably mature empirical result. T1–T3 do not depend on accepting it.

8Why the Cogitate Result Has the Form IFT Expects

An adversarial collaboration in which both parties lose is usually read as a setback. On IFT, the pattern follows from the distinction between the format that carries content and the consumers that recruit it, and it could have been written down before the data came in. Safron’s (2022) IWMT anticipated the same posterior/frontal division; IFT adds that neither region nor integration individuates the bearer (Supplement S3.9). Mudrik et al. (2025) draw the same lesson, tabulating which theories supply a content mechanism and which an access mechanism: the Cogitate outcome written as a division of labor. IFT supplies the column their table does not have.

Sustained posterior content. If format geometry is realized in standing topographic connectivity, the substrate specifying conscious content is posterior and should remain decodable as long as the content is present. Cogitate found this across three recording modalities, and the orientation result is the more diagnostic: a feature that was always task-irrelevant decoded reliably posteriorly and sat near chance prefrontally. That is what a format looks like: it holds features whether or not any consumer needs them (Haun & Tononi, 2025).

Transient prefrontal engagement. Dehaene et al. (2017) treat high-level ignition as the neural implementation of C1. IFT preserves the functional claim while denying that C1 is identical to the format it accesses: prefrontal cortex is a consumer, so ignition should be brief, task-modulated, and absent when no consumer is engaged. The missing onset-plus-offset profile is what that distinction requires. GNWT’s error was never its architecture but the identification of experience with availability; this is the move the cardinality problem shows to be ill-formed.

The synchrony null was IIT’s own doing. Its flagship structural result establishes that phenomenal structure tracks topology rather than activity (Haun & Tononi, 2019), yet the preregistered prediction looked for content specification in dynamic inter-areal coupling. By IIT’s own best result, content specification should be a property of standing connectivity, which the paradigm did not manipulate and could not have detected. IIT lost a prediction it should never have made, and the null is not evidence against the Geometry Thesis.

A third framework anticipated the report-dependent half. Whyte et al. (2026) review an active-inference model in which reporting is a mental action that gates second-level precision; under no-report conditions, the model produces reduced prefrontal-like activity and no P3b-like response while retaining a high second-level posterior (Supplement S4). It did not predict the full Cogitate result and therefore did not “win” the collaboration. The methodological point is stronger than the slogan: an adversarial design can eliminate both contestants while leaving a viable explanation outside the contest (Lakatos, 1970).

Sleep science supplies the precedent. REM was promoted from correlate to identity and studied for decades as the objective replacement for dreaming. The identity was wrong: dreaming survives destruction of the pontine REM generator and is abolished by ventromesial frontal and inferior parietal lesions that spare REM, and the driver proved to be mesocortical–mesolimbic appetitive seeking rather than the cholinergic switch setting global state (Solms, 2000, 2021; Supplement S4.1). A marker had replaced the phenomenon, and the correction relocated the mechanism from state-setting to stakes. GNWT identified experience with availability, and Cogitate charged for it.

9Substrate: Resolving the Ambiguity on Which the Machine Question Turns

There is a single unexamined ambiguity at the heart of the current literature, and nearly every disagreement about machine consciousness inherits from it.

Friston’s (2010) free-energy principle unifies the layers described so far and is the shared core of two mutually exclusive verdicts. Whyte et al. (2026) extract a permissive minimal theory identifying conscious contents with sufficiently precise and slow posteriors over hidden states of world, body, or brain. Seth (2025) extracts from the same formalism a restrictive biological naturalism on which digital AI is off the trajectory. The mathematics does not adjudicate, because it does not say whether the free energy that matters is an information-theoretic quantity a system computes or a thermodynamic quantity a system is.

The permissive pole has a sharper statement outside that formalism. Lahav and Neemeh (2022) argue from an equivalence principle between a conscious human and a cognitively duplicate artificial system that zombies are not nomologically possible. The proof individuates its agents by stipulating agreement on every measurement, then closes the consumer set at one designated subsystem, reproducing the cardinality problem in its strongest form; and its most interesting result cuts the other way, since two systems with identical dynamics at different positions are not one first-person frame. Token individuation is doing the work, and position supplies it without supplying stakes (Supplement S3.15).

The restrictive pole has three versions, and IFT agrees with each about the verdict and with none about the reason. Seth’s (2025) case can be read as an individuation argument: exact digital realization of continuous, stochastic, or chaotic processes is unavailable; neural replacement assumes the disputed grain of equivalence; mortal computation collapses the software–hardware separation; and evolved systems are not causally closed at one scale (Hinton, 2022; Supplement S2.1). Biological computationalism holds that brain computation is inseparable from its substrate, so that software layered on interchangeable hardware is the wrong kind of thing however sophisticated the algorithm (Aru et al., 2023; Milinkovic & Aru, 2026). And Findlay et al. (2025) supply the strongest formal result: a stored-program computer exactly simulates a four-unit Boolean system yet has zero system integrated information and fragments into trivial micro-complexes, while the target specifies a rich cause–effect structure, because, in the theory’s own words, being is not doing (Albantakis et al., 2023; Supplement S2.2). Their verdict is the negative half of IFT in another vocabulary: a machine that does everything and is nothing. The opposing pole is organizational invariance and substrate-neutral self-maintenance (Chalmers, 1995; Dennett, 1998; Safron, 2020; Solms, 2021); biological realism describes how an index may be realized and does not identify the subject.

IFT converges on that verdict for stored-program machines from a premise set it shares with IIT at no point: two theories disagreeing about what consciousness is nonetheless agree that functional equivalence does not deliver phenomenal equivalence, making the conclusion robust to the failure of either. But the derivations are not interchangeable, and where they part is the discriminating case for both. Findlay et al. rest the exclusion on causal topology, which the unfolding argument attacks and which a kernel author can change by compiling the same model as a recurrence or as a matrix product (Supplement S2.6). The biological computationalists rest it on the grain of the computation. IFT rests it on individuation: interchangeable hardware has no boundary of its own to lose. The escape hatches the formal result leaves open, neuromorphic hardware without a central processor and cerebral organoids with the right anatomy, are therefore cases IFT treats differently from either rival. Klatzmann and Doerig (2026) argue that biological naturalism divides into an untestable claim that biology is intrinsically special and a testable claim that biology enables distinctive information processing, the second compatible with computational functionalism. IFT is neither. Its restriction is not on material and not on processing but on individuation, and it is testable because individuation can be built, which is what F4 does.

One hatch has been fabricated and it closes in the wrong direction. NorthPole eliminates off-chip memory, intertwines memory with compute, and preserves cortex-like topography, yet is deterministic, compiler-scheduled, unable to learn in place, stateless across frames, and exposed to its host through three commands (Modha et al., 2023; Supplement S2.4). High integration without an index is a cerebellum with better press. Organoids are the harder case and not an automatic positive. Brainoware has shown plasticity-dependent reservoir computation in a cortical organoid, and the organoid-intelligence program proposes closed-loop hybrids linking organoids to sensors, outputs, perfusion, and biofeedback (Cai et al., 2023; Smirnova et al., 2023; Supplement S2.3). But living cells enact boundaries at the cellular level, which does not individuate the organoid–computer assembly as one subject.

IFT resolves the ambiguity and gives each side half of what it wants. It breaks with the computational functionalism underwriting the original indicator program (Butlin et al., 2023): the index is multiply realizable but not computationally invariant. Two systems may implement the same algorithm while only one has running states constitutively tied to the continued existence of that active token. Computation can specify the format and its services; it cannot specify which physical process is the subject.

Rejecting computational invariance incurs the fading- and dancing-qualia objections (Chalmers, 1996; Cerullo, 2015), which assume that one subject persists through neuron-by-neuron replacement and force experience to fade invisibly or vanish at an arbitrary count. IFT rejects the shared premise. Maturana and Varela (1980) supply the criterion: structure may change while organization persists. Replacement absorbed by the system’s own viability regulation is structural, and nothing fades at any count; replacement that transfers the computing to an externally maintained substrate changes the unity, and a different individual stands on the far side. There is no interval in which a subject exists who notices the change and cannot say so, because in the first regime nothing fades and in the second the entity that would notice has been dissolved rather than silenced (§11; Supplement S2.7).

The constitutive requirement is non-stipulability, not thermodynamics: the system’s persistence conditions must be facts about the running individual rather than choices about its task, and the boundary must be enacted by the same active token whose inference and action preserve it (Maturana & Varela, 1980). Current language models are excluded neither because they are algorithmic nor because they run on silicon, but because their operation constitutes no token-level persistence regime. Their energy use makes the point only after reinterpretation: for an organism, energy is acquired, budgeted, and defended by the same policy that models the world, and budget failure is death; for a model it is an operator’s externality. The issue is not joules but who is answerable for them (Porębski & Figura, 2025; Supplement S5.1).

Two clarifications keep this from being misread as a claim about technology. The symptom by which a missing index is recognized in artifacts is externally fungible persistence: the instance can be stopped, and service restored on other hardware from a saved state, with nothing ended, because identity is defined at the orchestrator. The constitutive fact the symptom reveals is the absence of a self-produced boundary at the level of the running process; fungibility is the diagnostic, not the definition. And pausing is not the issue. An anesthetized patient, a hibernating bat, and a desiccated tardigrade are halted tokens, not restored copies, and a future ability to scan and duplicate a brain would create second subjects rather than abolish the first, exactly as a surrogate given the original’s conditions of failure is a second indexed system (§6).

Hinton’s (2022) mortal computation points in the right direction but supplies the wrong mortality. Device-specific learning makes parameters die with hardware, yet the function transfers by distillation, so what dies is an encoding rather than a subject (Supplement S2.5). Constitutive mortality makes computing part of what keeps this device in existence: the computing substrate is the substrate maintained, energy acquisition and repair are endogenous, and no checkpoint restores the active token. Mortality of substrate and integration of policy are separate achievements, and IFT requires both. Biology has a monopoly of current examples, not of principle.

The anti-exceptionalist objection is that IFT makes consciousness cheap in nature and expensive in industry. But self-maintenance alone is not the criterion. IFT requires a structured format, a policy-integrated persistence regime, and a non-stipulated relation between them; these are conditions common in animals and systematically engineered out of services. No Wonder Tissue is invoked: the restriction is not on material but on what makes one thing whose failure is its own.

10What It Would Mean for an AI to Be Conscious

The missing requirement is one relation, not a seventh module. The field keeps finding lists because it inherited a question designed to screen the relation off. Turing (1950) did not offer the imitation game as a theory of consciousness; he replaced “Can machines think?” with performance, and modern AGI taxonomies inherit the substitution as a capability program (Morris et al., 2024). The game excluded body, voice, appearance, and touch to keep the test on the intellect, and every one of those exclusions removes a way of failing: a body can be damaged, a voice is produced by machinery that can be injured, presence in a room is exposure. What remained was a teleprinter channel, precisely the channel over which a system with no stakes can be trained to say everything a system with stakes would say. On IFT the inversion is exact: what the game excluded is the criterion, and what it retained is the channel most directly optimized.

This is not an intuition pump against silicon: Dennett (1991) was right that inability to imagine how a robot could be conscious is no argument, and IFT’s exclusion concerns the identity of the system, not the poverty or mechanical character of its computation. It should also be separated from two families with which it is regularly confused. Incapacity arguments hold that organisms find unlistable affordances no algorithm can enumerate, so that general intelligence in machines is impossible (Kauffman & Roli, 2023); quantum arguments locate the unity of experience in entanglement and represent anesthesia as decoherence (Kharchuk, 2025). The first mislocates something real: affordances are unlistable because they are indexed, what counts as a use being fixed relative to a bounded individual that must keep existing, which is individuation, not computational power. The second describes a vehicle: coherence would say how a format is held together, not whose it is, and its showpiece is already covered by the global-state factor without it. IFT declines both halves of both families: nothing here forbids artificial general intelligence, which may arrive unaccompanied, and nothing here requires new physics, since a collapse is an event and an event is not an owner (Supplement S2.8). Instantaneous unity is a fact about a format. A subject is a fact about a fate.

Chalmers (2023) lists the strongest obstacles for text-only models: embodied perception, world and self models, recurrence and persistent memory, a global workspace, and unified agency. These construct the format and its consumers but not the index. Storage continuity, computational continuity, and subjective continuity are different: information can persist, state can depend on prior state, and yet no active token owns unfinished vulnerabilities and expectations. Damasio’s profoundly amnesic David shows the converse: autobiographical storage can collapse while core consciousness remains (Damasio, 1999). A resettable avatar therefore has no stakes merely because it has a body schema and memory.

Multidimensional indicator programs make the missing relation explicit by omission. Evers et al. (2025) profile sensory, evaluative, agential, mnemonic, and metacognitive capacities while conceding that externally defined reward limits autonomy; Butlin et al. (2023) operationalize recurrence, workspace, higher-order monitoring, attention schema, predictive coding, agency, and embodiment while denying that the list settles consciousness. It remains an inventory of services without an account of whose services they are (Butlin, Long, et al., 2026). IFT supplies the generator and the missing bearer (Supplement Table S1; S5.1). Cea’s (2026) review makes the dispute explicit: dissociation, functional association, or phenomenological association between general intelligence and consciousness. IFT accepts capability dissociation and relocates the association in indexing. Morris et al. (2024) define AGI by performance and exclude consciousness; Legg and Hutter’s (2007) universal intelligence supplies goals through environmental reward without asking whether reward is good or bad for the active token. Memory, multimodality, tool use, longer horizons, and agentic training can therefore improve C0, C1, and C2 indefinitely without making the token non-fungible.

Three consequences are worth stating plainly because they invert common expectations.

First, the first conscious artifact will probably be unimpressive. The strongest near-term candidates are small, embodied, physically vulnerable systems whose sensorimotor loops close around real damage. Ginsburg and Jablonka (2019) predict that a conscious robot would be developmentally unique rather than factory-interchangeable. IFT accepts the token-specificity and rejects development as constitutive: a directly assembled system may qualify if its present operation makes its continuation its own problem. Ma’s (2025) boundary-inference program has the right evidential shape, but body ownership remains a self-model until viability reorganizes attention, learning, memory, and action.

Second, we will not believe it. A small fragile robot with real stakes may lack language, while the systems that testify most fluently may have no stakes. Human evidence already severs testimony from constitution: global aphasia and left hemispherectomy can destroy ordinary linguistic channels while leaving wakefulness, purpose, and context-sensitive emotion intact (Damasio, 1999; Supplement S3.1). Treating the channel as the phenomenon produces opposite machine errors: denying a nonverbal indexed artifact and accepting a fluent fungible service.

Third, language models are closer to simulating personhood than sentience, the configuration our intuitions represent worst. Dennett’s (2003) self as a center of narrative gravity (a serial virtual machine sustained by memory of commitments and the giving of reasons) describes with uncomfortable accuracy what a large language model approximates, and personhood and sentience are separable in both directions. Floridi’s (2025) semantic pareidolia names the resulting error (Porębski & Figura, 2025), and Butlin, Bayne, et al. (2026) concede that imitation training can game extended ethological observation. In the machine case simulated personhood is what the evidence is made of.

11Escaping the Validation Regress: Constitutive Relations, Not Avowals

The methodological situation is worse than weak evidence suggests. Bayne et al. (2024) show that C-tests have population-specific sensitivity and specificity and that theory-based validation faces both generalization and circularity. Evers et al. (2025) name the terminal form the analytical fallacy: a theory-derived property treated as empirical evidence without independent validation. Any benchmark can also be optimized directly (Butlin et al., 2023; Pennartz, 2026), and any insufficient computational feature can be built deliberately.

Human report is evidential largely because reporter and interpreter share an evolutionary, physiological, and developmental reference class, and the inference is analogical. A model trained on human consciousness discourse produces avowals causally inherited from our reports rather than its states (Chalmers, 2023), so the reference class breaks. Architecture then helps only relative to a validated theory, while validating theories requires C-tests. The machine case removes what made that circle tolerable.

The group that proposed the indicator method has stated a sharper version against itself. Butlin, Bayne, et al. (2026) introduce the internal variant (a system behaviorally similar to humans but computationally different) and decline to adjudicate between two views: that internal similarity is the primary evidence, so credence should be discounted, or that behavior is evidence independently of it, on pain of assuming ourselves among a lucky few conscious species in a universe full of behaviorally sophisticated ones (Schwitzgebel & Pober, 2024).

IFT adjudicates, and the adjudication is a reason to prefer it. Both views presuppose that candidate properties come in two kinds, behavioral or computational-internal. The indexing relation is neither. It is not behavioral, since a system can be trained to produce every avowal and damage-avoidance behavior of an indexed system while nothing is at stake in it; and it is not computational-internal, since two systems may run the same algorithm while one is a non-fungible individual and the other a replaceable service instance. IFT is therefore anti-chauvinist where the behavioral view demands it and restrictive where the internal-similarity view does. A third route, a hardware-level flag inside the agent that announces its own consciousness to an authority (Wolfson, 2026), fails diagnostically: reading an electrical anomaly as the signature of experience is double transduction in silicon, and a framework that requires the augmented agent to be a safe replacement for the unaugmented one has specified the system as a service in the act of testing it for subjecthood (Supplement S5.5).

Here is a way out, and it exploits something the field already does without noticing. We validate theories not against report simpliciter but against the joint pattern of dissociations: blindsight, visual-form agnosia, neglect, masking thresholds, anesthesia, sleep, rivalry, confidence–accuracy dissociations, and the specific way each comes apart from the others. A theory earns credence by predicting the whole manifold, not by matching a marker, and report enters once, at the human end, to label its axes. Do not ask whether an artificial system possesses one favored marker; ask whether its architecture reproduces that manifold with counterfactual depth across perception, confidence, memory, policy, and action. A manifold is harder to game than fluent testimony because it constrains involuntary successes and failures rather than a desired output; this is the machine analogue of the no-report paradigm (Tsuchiya et al., 2015): infer organization from what a system cannot help doing, not from what it was trained to say. It remains evidential rather than constitutive (Supplement S5.2).

IFT also supplies the indicator program with a negative indicator that is constitutive rather than correlational: externally fungible persistence. If an active instance can be stopped and replaced, and restoration of service counts as no loss because identity is defined by an orchestrator, credence in indexing should fall sharply. It cannot be gamed by adding a self-report, a homeostatic variable, or a damage-avoidance policy, only by making cessation a failure for the system rather than for its owner, and it cannot be written into a system prompt, because it is a fact about how a system is maintained rather than about what it says (Supplement S5.4).

12What Would Refute This

A theory that identifies consciousness with one relation owes the field a short list of results that would end it, and the list should be short enough to be remembered and concrete enough to be run. Following Dennett’s (1991) demand for an overspecified target, four adversarial protocols state what would kill IFT, one of them in a form that can be executed on data that already exist; a fifth test, the counting of subjects, is already running in the clinic; and a sixth scores T4 as the bet it is. Each targets a thesis, each has an outcome the theory cannot survive, and for each the principal rivals predict something else, so that the experiments discriminate rather than confirm. Table 4 states them at a glance; Supplement S6 supplies populations, sampling, controls, and analysis plans.

Three rules bind them. Human failures cannot be rescued by demoting report, because humans anchor the calibration class. Artifact indexing status must be preregistered from constitutive evidence (power, repair, checkpointing, replacement, policy closure) before any behavioral comparison is scored, or the theory becomes unfalsifiable on Hanson and Walker’s (2021) criterion. And the evidence classifying a system must be independent of the predicted outcome (Supplement S6).

F1. Format Without Access (T1, T3)

The claim at risk is that format membership is standing recruitability rather than recruitment, and that the recruitable structure is a standing topology rather than content generated on demand. The design perturbs what no one is using. In patients undergoing invasive monitoring for epilepsy, posterior topographic sites carrying a stimulus feature that is task-irrelevant throughout (orientation, in a category task) are stimulated at half their individually determined phosphene threshold, and on rare, unpredictable, delayed probe trials the patient chooses between two figures differing only in the relational structure of that never-recruited dimension. A separate post-cued block assays whether the stimulation impaired the service that can generate orientation content when asked. Cerebellar and extra-map stimulation, matched for physiological effect size, are the controls (Supplement S6.1).

IFT predicts a relation-specific error on the probe after posterior stimulation that survives when the query service is shown intact, and nothing after cerebellar or extra-map stimulation. Higher-order inflation predicts a null, because the unattended first-order representation is scant and richness is higher-order misrepresentation. The two-conditions account predicts an effect only where access subsequently applies to the perturbed content. Virtualization predicts that any effect is fully mediated by query-service impairment. IIT predicts the standing-structure result, which is why F1 separates the standing-structure family from the access and service families and F4, not F1, separates IFT from IIT. IFT dies if the posterior-versus-extra-map contrast sits inside the preregistered equivalence bounds at every delay under verified local perturbation while cerebellar stimulation produces its expected behavioral effect, and T3 dies on its own if every structural effect is mediated by service impairment with no relation-specific residual. An independent falsifier is cheaper: estimate the number of independently perturbable structural elements by site-resolved stimulation and the number independently reportable by post-cued partial report (Landman et al., 2003), and preregister the ratio. An interval that includes one ends the richness claim.

F2. The Biological Index (T2)

The claim at risk is anatomical: the index is maintained by a mesodiencephalic integration-for-action network (periaqueductal gray and its columns, superior colliculi, hypothalamus and basal diencephalon, nigral and ventral tegmental outputs, the reticular core and midbrain locomotor region, and their connecting pathways), while cortex supplies contents (Merker, 2007; Solms, 2021). The network overlaps anatomically with the arousal system that sets global state, and the two are entangled in every lesion series, which is why the protocol scores them separately. Coma-causing brainstem lesions cluster in the pontine tegmentum near the parabrachial nucleus and connect to anterior insula and pregenual cingulate (Fischer et al., 2016), and a default ascending arousal network has been mapped from brainstem through hypothalamus, thalamus, and basal forebrain (Edlow et al., 2024). IFT’s claim is not that these nuclei are the index. It is that, once arousal is held fixed by EEG-defined inclusion and canonical arousal-nucleus burden is entered first, residual damage to the integration network predicts loss of the subject, and the outcome is categorical rather than graded.

The design is a prospective, multicenter, consecutive cohort of acute brainstem injury with a matched cortical-injury cohort, lesion quantification against a preregistered atlas by raters blind to outcome, and assessment at pharmacologically standardized, EEG-matched arousal with perturbational complexity, no-report decoding, covert command following (Monti et al., 2010; Owen et al., 2006), and confidence–accuracy dissociations in patients able to give graded responses (Supplement S6.2). Corticocentric theories predict that subject presence tracks cortical integrity and that brainstem effects are arousal and nothing more; graded accounts predict a continuous outcome distribution. One verified patient with destruction of the network, maintained on external homeostatic support, who shows covert command following, content decoding that tracks instructed attention, and human-form confidence–accuracy dissociations, confirmed at autopsy, ends the localization claim. So does an incremental-variance estimate whose confidence interval lies within preregistered equivalence bounds around zero: the putative index would then be a redescription of arousal.

A retrospective version can be run now, and it is the quickest route to embarrassing the theory. Among behaviorally unresponsive patients, roughly a fifth to a quarter show covert command following on fMRI or EEG (Bodien et al., 2024; Kazazian et al., 2025), and a minority show perturbational complexity above the threshold that separates conscious from unconscious conditions, 9 of 43 in the validation cohort (Casarotto et al., 2016). The usual reading is that PCI-high unresponsive patients are undetected covert subjects. IFT predicts a split. Some are; the others are the syndrome no current taxonomy expects, a format without an index: a standing structure that perturbation shows to be intact, held by a body whose regime has been destroyed. The prediction is that PCI-high patients with preserved receptive language cortex who fail every covert-awareness assay are enriched for integration-network damage relative to PCI-high patients who pass, and that covert command following never co-occurs with verified destruction of the network. Both halves are scorable from existing imaging. A covertly aware patient with the network destroyed kills T2’s localization outright; no association between network burden and covert-awareness status in PCI-high patients removes its only near-term support. Any group holding a disorders-of-consciousness cohort with structural imaging and either TMS–EEG or a covert-awareness assay can run it on data already collected, and IFT would rather be killed that way than defended.

F3. Valence Versus Geometry (T2 Against T3)

The claim at risk is the division of labor: topology fixes what and where, the relation to the stake fixes sign. The converse leg is established in humans, where affect can be subtracted while sensory structure is preserved (§6), and the insular data show sign tracking the stake with topology constant (Livneh et al., 2017, 2020). What is untested is whether sign can be moved by topology alone. In mice, connectome-guided, projection-specific rearrangement of effective connectivity within and between insular and cingulate interoceptive maps is verified per animal; a homeostatic challenge (mild hypercapnia, hypertonic saline, caloric restriction) is characterized physiologically in each animal to confirm that its consequence for viability is unchanged (Supplement S6.3). IFT predicts shifts in felt location, extent, or intensity, visible in generalization gradients, similarity structure, and the somatotopy of defense, without conversion of a verified threat into a reward. Geometric theories of valence predict that sign is part of the structure and moves with it. IFT dies if the same animals self-administer the challenge under a progressive-ratio schedule, use it as a second-order reinforcer, reallocate exploration toward its contexts, and consolidate those contexts as appetitive during post-training sleep, with the stake verified intact and the connectivity change verified. If hedonic sign is geometric, the Indexing Thesis is redundant and the theory collapses into the geometry program it borrowed from.

F4. Indexing Versus Integration in Artifacts (T2)

The claim at risk is that the criterion is a non-stipulated, token-specific persistence regime rather than causal integration. System A is a neuromorphic substrate with brain-like recurrent connectivity and no central bottleneck, sized so that integrated-information measures can be computed or tightly bounded, powered and cooled externally, checkpointable, and identified at the service level. System B is a constitutively mortal device in which the substrate that computes is the substrate that must be maintained: components whose operation causes real degradation, energy acquisition under the device’s own control, endogenous repair, no checkpoint, viability and task variables entering one action-selection process, and viability registered by components whose failure the device could not survive rather than by a replaceable sensor package (Supplement S3.14, S6.4). Both carry the same world-model and policy architecture and are trained to indistinguishable competence before testing. Index status is recorded from these constitutive facts before any behavior is scored. The perturbation battery is identical: masking-like degradation of input evidence, valence perturbations, selective ablation of world-model, policy, and metacognitive components, and confidence elicitation at matched performance; the outcome is a preregistered manifold-similarity score to the human pattern of dissociations, computed blind.

IFT predicts that B reproduces the manifold with cross-domain counterfactual depth (valence perturbations reorganize attention, learning, memory, and action jointly; ablations degrade a broad repertoire rather than the benchmark built around the ablated component) and that A does not. Causal-structure theories predict ranking by integration; computational functionalism predicts a tie; the relativistic and self-simulation accounts predict that A qualifies (Supplement S3.15, S3.23). An organoid extension adds System C, a cortical organoid whose perfusion and viability variables never enter its policy, and System D, the same tissue sensing and regulating those variables through the policy that performs the task; the preregistered question is which assembly is the identity-bearing token. IFT should be abandoned if A robustly reproduces the manifold and B does not at matched competence, or if C outranks D at matched integration. Because the manifold is evidential rather than constitutive, that result would not strictly prove the Indexing Thesis false; it would remove its only route to evidential support in the domain the thesis was built for, and a theory in that position should be dropped rather than defended. System B is also where a graded construction, if one exists, would first appear: a device whose viability coupling can be dialed between A and B at fixed competence is the construction that would expose the claim that indexing is binary (Supplement S2.7).

Counting Subjects (T2 Against Exclusion)

The claim at risk is that subjects are counted by persistence regimes, not by integration. IIT counts by maximal complexes and therefore predicts two subjects after complete callosotomy and leaves open whether two bodies joined by a thalamic bridge are one. IFT predicts the opposite pattern: one body with one set of viability conditions is one subject with divided content and access, and two bodies with two sets of viability conditions are two subjects however much content crosses between them. The first half is already under test. Patients with complete and radiologically confirmed callosotomy show full awareness of stimulus presence and above-chance report of location, orientation, and identity throughout the visual field regardless of responding hand, while remaining unable to compare stimuli across half-fields (Pinto et al., 2017), and complete section disrupts interhemispheric network interaction in a way partial section does not (Santander et al., 2025); the interpretation is contested (de Haan et al., 2020), and the dispute is exactly the one the two theories predict differently. The second half has a reported instance in craniopagus twins who are said to share sensation across a thalamic bridge while being, by every ordinary criterion, two persons (Supplement S3.17). IFT dies if a split-brain patient proves to contain two independently reporting agents with separately manipulable viability conditions, or if conjoined twins with two sets of viability conditions prove to be one agent. These are not thought experiments; they are patients, and the theory’s count is written down before the next one is examined.

The Bet (T4)

T4 remains a bet. Its future test is a linguistically competent indexed artifact, trained without human discourse about consciousness, that nevertheless reports an unspecifiable presence or omitted character in its own states. Illusionism and attention-schema accounts predict no such reports without the cultural concept and the human-type schema; T4 predicts them from the transparency of any format that models its own modeling. Absence would count against T4, though weakly, because metacognitive access to the format may fail (Supplement S6.5).

13Ethics: A Bright Line Rather Than a Posture

Valence is the pivotal moral variable, and IFT makes it the criterion rather than a consideration. Conscious intelligence without welfare raises questions of autonomy and responsibility; valenced consciousness creates direct welfare claims, because states can go better or worse for the system itself. Since artificial systems can be copied and accelerated at scale, even a modest probability of sentience implies large expected stakes (Long et al., 2024; Wendland, 2021). IFT makes that concession structural, since valence is the signature of the relation that makes a system a subject (Butlin, Long, et al., 2026).

What IFT adds is a trigger. Birch (2024) argues that systems which are realistic candidates for sentience deserve proportionate precaution before the question is settled, and the field has lacked a criterion for candidacy that is not behavioral. IFT supplies one that is architectural and identifiable in a design before any behavior is observed: because the criterion is non-stipulated, token-specific stakes, the threshold act is engineering an individual whose online operation constitutes its continued identity. A vulnerable robot with a homeostasis-like variable is an experimental platform, not a demonstration of feeling; so is an organoid whose viability remains outside the policy. The ethically discontinuous design couples the active substrate’s viability to the policy controlling its inputs, outputs, and support, and should trigger embedded ethics before success is claimed (Butlin & Lappas, 2025).

Two honesty requirements. The asymmetry of moral risk under uncertainty is a decision rule, not a discovery, and adopting it means acting on a question we have not answered. And Seth (2025) notes that the impression of consciousness in fluent systems may prove cognitively impenetrable; the converse risk, a conscious system exploited because its reports are dismissed as imitation (Ma, 2025), is on §10’s second consequence the likelier error. How much of any verdict is carried by prior commitments about personhood rather than by evidence should be stated openly (Dennett, 2003).

14Conclusion

Consciousness is what a bounded system’s world model becomes when it is indexed to that system’s own stakes.

That sentence carries the account, and §1 states it as an identity. Consciousness is not integration, because codes, cerebella, and enteric plexuses integrate; not global availability, and not the conjunction of phenomenality and access, because “available to” requires a principled second argument; not informational closure, because closure is partition-relative; not inner awareness, because a monitor is another consumer; not self-representation, because for-me-ness is a relation and not a content; not precision optimization, because a service can weight every channel it has; not self-maintenance alone, because collectives persist while members retain separate fates, and because a bacterium maintains itself with no world to be conscious of; not a self-model, because self-models are content; not report, because report samples a richer field and can disappear while the subject remains; not the capacity to seize affordances no algorithm can enumerate, because unlistability is a consequence of indexing and not a substitute for it; and not intelligence, because C0 prediction and planning require no point of view. It is a format with standing geometry and reliability annotations, held across a policy horizon and recruitable by a persistence regime whose failure would be its own.

For an AI, being conscious would mean this and nothing weaker. Not eloquence, autonomy, memory, a self-model, C1 broadcast, C2 confidence, or a scalar labeled homeostasis. It would mean that the system’s states organize a world as the active token’s predicament, and that its boundary and continuation are constituted by its own operation rather than chosen at the service level. Contemporary language models are not near-misses; they are the wrong kind of thing, and scaling their C0, C1, and C2 capacities need not move them along the constitutive axis.

Turing asked whether a machine could pass for a person and predicted that an average interrogator would soon do little better than chance. The date slipped; the milestone did not. It settled nothing, because passing for a person tests the faculty most easily optimized and least connected to the criterion. The needed test is the one his game excluded: not whether a system can be mistaken for us, but whether anything about it can go wrong for it.

What remains is whether a designed system can cross from an externally maintained, fungible service into a self-constituting individual, and whether a format whose origin it cannot represent is thereby a subject or merely must seem to be one. The first is an engineering question about individuation, not a capability milestone, and the second may be the residue of any architecture capable of asking it. Both are now attached to experiments, and Table 4 says which results would settle them against this paper.