Architecture of Human Cognition: A Recursive Domain-Axis Model Based on Structured Realism

1 Introduction

The idea that knowledge proceeds in stages is old. Kant puts it this way in the Critique of Pure Reason: “All our cognition starts from the senses, goes from there to the understanding, and ends with reason” (B 355; our translation) [95]. Twentieth-century cognitive psychology adopted this order in stage and filter models of information processing [27]; Hurley described the underlying pattern as the “classical sandwich” of perception, thought and action [86].

There is an equally old counter-tradition to this picture. As early as 1896, Dewey argued that stimulus and response form not an arc but a circle, in which action determines what counts as a stimulus in the first place [52]. Von Uexküll described the relationship between organism and environment as a “functional circle” of perceptual world and effector world [156]. Today, predictive processing gives this counter-tradition a formal shape: the brain does not merely receive signals, it predicts them and processes above all the deviations [41, 68].

What has been missing is a framework that connects the conceptually rich theory of faculties of classical philosophy with this recursive picture while taking an explicit ontological position. The present contribution proposes such a framework (Fig. 1, above). We keep apart what a claim is about, and we distinguish what it achieves: it can describe a phenomenon by giving it a place in the model, explain it by stating a mechanism, or predict something that is independently testable. A model that only describes can classify anything and is therefore of little value.

2 Ontological foundation: Structured Realism

The model rests on three theses. First, the external world exists independently of our observations. Second, the mind structures, filters and interprets reality without producing it. Third, the structures of cognition were themselves shaped by evolution in contact with the world; they therefore fit it without fully mirroring it. The third thesis builds on Lorenz, who interpreted Kant's a priori forms of intuition as adaptations acquired in the course of evolution [112], and on Vollmer's evolutionary epistemology [158].

Two criteria follow. The constraint criterion reads: real is what constrains us independently of our experience – a wall one walks into, an experiment that turns out differently than expected, a program test that fails. The invariance criterion reads: what stays the same across different cognitive systems belongs to the world; what differs between them belongs to the system. A falling tree produces a pressure wave that a dinosaur, an early mammal and a bird, with very differently built ears, register at the same time and place. The sound they hear differs in each of these systems; the pressure wave, its location and its timing do not. The heard sound is a quale – the experienced quality, which belongs to the hearing system; the pressure wave belongs to the world. The invariance criterion thus does not deny qualia; it assigns them to the system. The distinction has a precursor in Locke's primary and secondary qualities: extension, shape and motion belong to the thing, colour, sound and warmth arise in the perceiver [111]. The invariance criterion turns this list into a test: primary is what stays the same across differently built systems. Nozick defined objectivity in this sense as invariance under transformations [129]. Invariance counts only across differently built systems. When identically built brains agree under identical conditions – for instance when many people under DMT report similar entities (Section 5.3) – the agreement reveals their shared architecture, not the world.

Relation to Kant. We adopt Kant's theory of faculties and read it realistically. Kant himself called his position empirical realism (Critique of Pure Reason A 370) [95], and more recent readings understand the unknowability of things in themselves to mean that only their relational properties are accessible to us [105]. We go one step further: because the structures of cognition were shaped by the world, we regard things in themselves as knowable from a perspective. Kant's insight that cognition is always already structured remains the starting point of the model.

Convergence across traditions. The invariance criterion can also be applied to theories of mind. Traditions that developed independently of Kant arrived at similar basic distinctions. In Indonesian educational philosophy, Ki Hajar Dewantara's triad of cipta, rasa and karsa – thinking, feeling, willing – corresponds to Kant's three faculties of the mind (the faculty of cognition, the feeling of pleasure and displeasure, the faculty of desire) [97] and to our domains of understanding, emotions and will [51]. Yogācāra Buddhism describes a layered architecture with a “storehouse consciousness” (ālaya-vijñāna), in which impressions continue to act unconsciously as “seeds”, and a consciousness (manas) that constantly clings to a self [114]. Wang Yangming's “unity of knowing and acting” (16th century) anticipates the objection to the linear sandwich [161]. Such convergence is no proof. Since the traditions describe the same human architecture, their agreement speaks for distinctions that track this architecture – which is precisely what the model claims to describe.

Distinctions. Epistemologically, Structured Realism is a form of indirect realism: we know the world through inner models, not immediately. It differs from naive realism, which takes the world to be as it appears, in exactly this respect, and from constructivism in that these models are tested against the world and constrained by it. Against idealism, Structured Realism denies that reality exists only in the mind; against radical constructivism, that the world is merely a construction of the knower. Interface theory and the simulation hypothesis are treated in Section 9.2. Despite the similar name, it differs from structural realism in the philosophy of science [165]: for us, structure resides not only in world and theory but also in the knowing subject, and the world remains a world of things. There is a bridge to Russell's thesis that perception conveys above all the relations of the external world [141].

Place in the spectrum. If the positions of epistemology are ordered by whether they take mind or matter to be fundamental, the range extends from Berkeley's immaterialism [14] and Hegel's absolute idealism [79] to eliminative materialism, which would abolish concepts such as belief and desire [40]. Structured Realism lies beyond the middle, on the realist side: the world exists independently of the mind, but experience is not explained away; it is acknowledged as an open problem (Section 8). Its closest neighbours are the evolutionary epistemology of Lorenz and Vollmer [112, 158] and Popper's critical rationalism [134]; with Kant it shares the question, with Lorenz the answer.

3 Method

This contribution is theoretical. The model was developed over several iterations, beginning with a freely designed graphic. Each iteration was checked for consistency; the central criterion is that every element must have exactly one place in the model. Double placements were resolved, for example emotions appearing both as a domain and as a modulating dimension. The model's claims lie on three levels:

Two rules connect the levels. A phenomenon may not force new building blocks into the architecture: if a case can only be described with an additional building block, this counts against the architecture. And the architecture does not decide ontological questions: whether a space experienced in the DMT state exists independently is a question for the criteria of Section 2. Figure 1 is generated from a single data source containing all domains, axes and connections, so graphic and text cannot drift apart.

4 The model

4.1 Domains

The Cognitive Field consists of nine domains. Perception forms the interface to the world and to one's own body (exteroception, interoception, sensory integration). Memory stores and retrieves. Understanding works with language, logic and concepts. Judgment applies knowledge to the case: it selects what fits here and relates cases through their relations. Reason reflects on one's own processes (metacognition) and forms worldviews, ethical principles and meaning. Prospection simulates what is possible. Social cognition models others and their mental states. Emotions evaluate affectively and motivate. Will decides, sets goals and exercises self-control. What sets the field in motion at all are the drives; they form not a domain but a separate layer beneath it (4.9).

The core is Kantian. We understand prospection as a modern version of the productive imagination; remembering and imagining largely use the same systems [78, 142]. Social cognition receives a domain of its own because modeling other minds is an independent, early-developing and selectively impairable capacity [7]. Empathy is distributed over two domains: affective empathy belongs to the emotions, cognitive empathy to social cognition [19].

Dissociations in the Kantian core. Critics object that understanding, judgment and reason lack the clear lesion profiles of memory. There is evidence nonetheless. After lesions of the anterior prefrontal cortex, performance in perceptual decisions remains intact while the accuracy of judging one's own performance – metacognition, an activity of reason – is impaired [64]. After damage to the ventromedial prefrontal cortex, intelligence and knowledge of rules are preserved while practical and moral judgment changes [9, 101]. We read this as evidence for the coupling of emotions and judgment: without emotional signals, judgment loses its weighting, not its knowledge. Nor does prospection simply coincide with memory: in semantic dementia, imagining the future is impaired while remembering the past is relatively preserved [87]. Not every domain is equally homogeneous, however. Memory contains several dissociable subsystems – episodic, semantic, procedural, working memory. We treat a domain as the level at which systems dissociate, and the subsystems as the level below it.

Understanding and judgment are the most closely related domains. Judgment applies knowledge to the case, in two sub-functions: it selects from available knowledge what fits here and suppresses what does not, and it relates cases through their relations (analogy). Both can be separated from knowledge itself. After stroke aphasia, conceptual knowledge is preserved but cannot be deployed appropriately; in semantic dementia it has itself decayed [88]. Prefrontal lesions impair the integration of relations while knowledge is preserved, anterior temporal lesions impair knowledge while integration is preserved [160]. Retrieval and integration use separable prefrontal systems [31] and a dedicated semantic control network that is not merely linguistic [128]. Kant's distinction fits these sub-functions remarkably well: understanding is the faculty of rules, judgment the faculty of subsuming under rules (Critique of Pure Reason A 132/B 171) [95] – determining when the rule is given, reflecting when it has yet to be found [97]. The boundary with reason remains open: like metacognition, the integration of relations relies on the frontal pole. Since knowledge without application would remain empty, understanding and judgment are directly coupled. Perception and emotions are also directly coupled, dashed in the diagram, because sensory and bodily signals are evaluated quickly and largely preconsciously [108]. We do not introduce a direct coupling of perception and judgment, not even for Kant's reflective judgment, which seeks the universal for a given particular [97]: Kant himself routes it through imagination and understanding, and lesion findings show that classification fails when the route via memory and understanding is interrupted (Section 5.7). In total, the field contains nineteen couplings.

Language. Language runs through almost all domains: we remember, plan and judge in language. We treat it as a medium anchored in understanding – which is why Figure 1 names it there – and, as speech, in action. Its best-known correlates are Broca's area in the left inferior frontal gyrus, whose damage impairs speech production and grammar [28], and Wernicke's area in the left posterior temporal lobe, whose damage impairs comprehension [163]; today both are understood as parts of a wider left-hemisphere language network [59]. In fairness, one finding speaks against treating language as part of understanding: in aphasia, language fails while thinking and planning are partly preserved [59]. Whether language deserves to be a tenth domain therefore remains open (Section 10).

Embodiment. The Cognitive Field is always embodied in an organism, in humans above all in the brain. The body is therefore not a layer of the field of its own but its carrier. In Figure 1 it appears as a band beneath the field – what carries the field, not another layer of it – and additionally in three places: in perception as interoception, in the unconscious as bodily processes, and in the motor system. The states of the carrier – wakefulness, arousal, homeostatic need – act on all domains at once. Luria's division of the brain into three functional units, derived from lesion findings, has the same structure: one unit for tone and wakefulness, one for receiving and storing information, and one for programming and controlling activity [113]. The domains are defined functionally; the network correlates in Figure 1 [166] are correlates, not localizations.

4.2 Modulating axes

Domains describe what the mind does. The axes describe how it does it at a given moment. Every domain can work consciously or unconsciously, automatically or in a controlled way, directed at the past or the future. Emotions are therefore a domain, whereas valence is an axis: a tone that any content can have. We propose seven axes: access (unconscious – preconscious – conscious), control (automatic – controlled – reflective), time (past – present – future), perspective (self – others – group), valence (negative – neutral – positive), modality (possible – real – certain) and authorship (self – other – no one). Established measures exist for each axis, such as confidence judgments, reality-monitoring tasks and everyday thought sampling [63, 91, 145]. The seven axes are a proposal, not a complete list like Kant's table of categories.

Functions and categories. Domains and axes differ in kind. Domains are functions with neural correlates; lesions can separate them from one another (Section 5.7). Axes are categories under which every function operates, loosely modelled on Kant's forms of intuition and categories: time is modelled on the form of intuition of time, modality, with its poles possible, real and certain, on Kant's modal categories of possibility, actuality and necessity, and authorship on his category of causality within the categories of relation. The analogy is meant as orientation, not as identity: Kant's forms make experience possible in the first place, whereas our axes describe how already formed contents are processed. No brain area with content of its own corresponds to an axis; its current value is set by diffusely acting modulatory systems and a few hubs (Section 4.8). The difference shows in what a disturbance affects. A lesion of a domain preferentially affects one kind of content or computation – recognizing objects, say. A shift on an axis changes the same parameter in many domains at once, whether it is brought about by sleep, intoxication, stress or a lesion. To guard against an inflation of axes, we add a test: for every axis there must be at least one state that shifts it independently of all other axes. The axes themselves are common to all cognitive systems; their range has grown in evolution (Section 6). The criterion also shows where the list is uncertain: control is close to a processing mode, valence close to the emotions; further candidates are discussed in Section 10. The will is the domain that sets values on the control axis, just as the preconscious as gatekeeper regulates access: mechanism and value are kept apart. In Figure 1, six axes appear as sliders; access appears as the depth of the field (Section 4.3).

Modality. This axis corresponds to reality and source monitoring in memory research [91, 92] and, in Kant's terms, to modality. It captures the status the system attributes to a content: merely possible (imagined, planned, fictional), real, or certain beyond doubt. Where a content actually comes from – perception, memory, imagination – is a fact about world and system; it can only be established from outside, by the criteria of Section 2, and is not a value on the axis. A whole type of phenomena thus becomes describable as a gap between actual origin and attributed status. In hallucination, dreaming and the DMT state, internally generated content receives the status of reality; in delusion, even of certainty. In the confabulation of AI language models, generated content is output as fact. In cryptomnesia, someone else's idea is taken for one's own invention [29]. In Anton's syndrome, patients who are cortically blind deny their blindness and describe what they supposedly see [2]. The axis passes the independence test in the Charles Bonnet syndrome: people with severe visual loss see vivid, detailed images, yet usually know that they are not real [152]; the images arise in visual cortex, as in seeing [60]. Perceptual content is fully present while the attributed status remains “not real” – the reverse of hallucination in psychosis. Modality thus varies independently of perception. The status “merely possible” can also be set deliberately, as in children's pretend play or in reading fiction [109]. Earlier versions called this axis “information source” (external – internal – imaginary); three external reviews showed that this mixed origin and status. We now name it after Kant's modal categories. Who is the author of a content is a separate question, captured by the next axis.

Authorship. This axis captures whom the system takes to be the originator of a thought, an action or an event: itself, another agent, or no one – chance or mechanism. It corresponds to the sense of agency [76] and, in Kant's terms, to causality. It passes the independence test: in the alien hand syndrome, a movement is experienced as not one's own while the sense of self and perspective remain [73]; in thought insertion, one's own thoughts are attributed to another agent [69]; in possession trance, words and actions are attributed to a spirit. In the opposite direction, social cognition attributes events to agents where there are none – the “someone is there” of the DMT state (Section 5.3). Authorship is related to perspective but not identical with it: perspective asks from whose standpoint something is experienced, authorship who brought it about.

Several candidates from earlier versions are not axes: metacognition is an activity of reason, attention a selection mechanism of the preconscious; motivation arises from the drives and is evaluated by the emotions (4.9). Approach versus avoidance is not an axis either, but the motivational direction within the emotions. It can be separated from valence – anger is negative yet drives approach [35] – but demonstrably shifts only emotions and will; findings according to which it also organizes attention and frontal brain activity have not held up in large samples [131]. Wakefulness is not a value that each domain takes individually but a state of the whole organism (Section 4.3). The same holds for arousal: as graded wakefulness, it acts on all domains at once and in the same direction, and is therefore a state of the carrier rather than an axis on which each domain has its own value. Allostasis also belongs to the carrier: the adaptation of the organism to strain, whose price under chronic strain is allostatic load [118, 149]. In Figure 1, these states of the carrier appear in the lowest band.

4.3 Consciousness in four aspects

“Consciousness” refers to several things, and many misunderstandings arise because they are conflated [22]. The model gives each aspect a place of its own:

Languages confirm this division from two sides. Indonesian uses one word, sadar, for being awake, being aware of something and realizing it – the aspects blur easily. Chinese separates them lexically: 清醒 (qīngxǐng, awake), 意识到 (yìshí dào, to become aware), 自觉 (zìjué, self-aware) – they can be kept apart.

Two dissociations show that wakefulness and access differ: in unresponsive wakefulness syndrome, sleep and waking alternate without demonstrable access [106]; in dreams, contents are consciously accessible while wakefulness towards the world is low. In Figure 1, access is shown as depth: below the domains lies the preconscious as gatekeeper, and below that the unconscious, which pervades all domains. This adopts Freud's topographic model [67] but interprets it as a degree rather than a place, and it corresponds to the access function described by global workspace theories [5, 49].

4.4 The preconscious as gatekeeper

In neuroscientific terms, the preconscious corresponds most closely to the salience network of anterior insula and dorsal anterior cingulate cortex [144]. It detects which events are currently relevant, switches between the inward-directed default mode network and the outward-directed control network [121], and integrates bodily signals with emotion and attention. It is thus exactly what the model requires at this point: a layer without content of its own that selects and releases. We thereby separate mechanism and value: as a layer, the preconscious is the gatekeeper; conscious, preconscious and unconscious are the values on the conscious-access axis that a content receives through this gate. Each element thus has exactly one place. Dysregulated salience makes a class of disorders intelligible: in psychosis, irrelevant stimuli are charged with significance, and delusions explain this significance [99]; in the model, the path leads from the preconscious via judgment to a misattribution of reality status.

4.5 Metacognition: the field turns back on itself

What makes the field recursive in the strict sense is metacognition: the system makes its own processes the object of processing. It monitors how certain it is, notices errors, judges what it knows and does not know, and adjusts its control accordingly [63]. In the model, metacognition is the core activity of reason, but it reaches into every domain: there is metamemory (knowing whether one knows a name), metaperception (confidence in what one sees) and the monitoring of one's own actions. It can be measured as the agreement between confidence and accuracy, and it can fail selectively: after lesions of the anterior prefrontal cortex, perceptual decisions remain intact while the accuracy of judging them suffers [64]. Metacognition also makes it possible to hand over control without giving it up entirely: in hypnosis and trance, one's own actions are experienced as automatic while a higher level maintains the frame in which this happens. The cold control theory reads hypnosis in just this way – as the execution of an intention whose accompanying meta-representation is missing [53]; in the model, the control axis thus becomes two-tiered: handed over below, held above.

Metacognition connects several parts of the model. It is one of the two components of self-consciousness (4.3); its reach defines recursion depth in the comparison of species (Section 6); its failure appears in anosognosia and Anton's syndrome, where a deficit goes unrecognized [2]; its return marks the lucid dream (5.6); in AI language models it is demonstrable but unreliable [110]; and in the meta-problem it is the capacity that turns transparent models of the world into a puzzle (Section 8). In Figure 1, metacognition therefore heads the description of reason.

4.6 The circle between world and action

The world supplies stimuli to perception and the preconscious. The will leads through action and the motor system back into the world, which is thereby changed and generates new input. Within this circle, the system predicts and compares; the oldest experimental example is the reafference principle [83]. Action is behavior as described from the outside; the motor system is its bodily execution via motor cortex, cerebellum and muscles. This assignment, too, names correlates, not localizations: cerebellar lesions also change thinking and affect [143]. The separation is not merely a matter of drawing: in Parkinson's disease, execution is preserved while initiation is impaired (Section 5.7).

4.7 Higher phenomena as couplings

Phenomena such as conscience, the self or intuition do not receive domains of their own. They appear as stable couplings of several domains with characteristic axis positions. This keeps the model parsimonious and testable: if a phenomenon is a coupling, disturbances of the domains involved must change it predictably. Intuition, for example, is a fast, largely unconscious interplay of memory, pattern recognition, prospection and judgment near the “automatic” pole of the control axis.

4.8 Neurotransmitters and connectome

If the axes are real dimensions, they should be shiftable neurochemically, and in several domains at once. Candidate hypotheses: noradrenaline and acetylcholine for access [4], dopamine for control and valence [15, 137], serotonin (5-HT2A) for perspective and modality [136]. Every neurotransmitter acts on several axes; a one-to-one assignment would be a simplification. This fits the distinction drawn in Section 4.2: the neuromodulatory systems arise from small nuclei in the brainstem and basal forebrain and act diffusely on the whole cortex, not on one area. Some axes also depend on hubs, such as perspective on the temporoparietal junction; stimulating it shifts perspective rather than abolishing it [21].

The model also makes a claim that can be tested against the connectome: directly coupled domains should be more closely connected via their network correlates than uncoupled ones. A first comparison fits: the cortex is organized along a gradient from the sensory areas to the default mode network, which lies furthest from them [115] – like reason in the model, which has no direct coupling to perception. And lesions in different places produce the same symptom if they hit the same network [65]; this supports the functional definition of the domains.

4.9 Drives: sources of impetus

A field that only perceives, evaluates and decides does not explain why it gets going at all. Drives fill this gap. We group them in four: bodily drives (food intake, thirst, thermoregulation, self-preservation), reproduction (sexual drive, care for offspring), social drives (attachment, rank, territory) and epistemic drives (exploration, play and, as a hypothesis, a metaphysical need for meaning). The groups largely match the basic systems that Panksepp delineated in mammals through brain stimulation and lesions: seeking, rage, fear, lust, care, panic (separation distress) and play [130].

Drives are not a tenth domain. They share no content and no computation; they are sources that act into different domains: exploration into perception and prospection, attachment and rank into social cognition, the need for meaning into reason, hunger and sexual drive into emotions and will. In Figure 1 they therefore lie as a separate band beneath the field and act diffusely upwards – the counterpart of the axes, which act on all domains from above. The model thus takes up Freud's concept of drive as a frontier concept between the mental and the somatic [66]: drives arise from the carrier but only become wishes, goals and plans in the field.

That drives are absorbed neither by the emotions nor by the will is shown by the dissociation of wanting and liking. If dopamine in the reward system of rats is almost completely removed, liking is preserved – their reactions to sweet tastes are unchanged – while wanting, the impetus to seek out a reward, collapses [16]. Wanting corresponds to the drive, liking to the evaluation by the emotions, and the choice between goals is the business of the will. Akinetic mutism shows the same boundary from the other side: perception and evaluation are preserved, impetus is missing (Appendix).

Two everyday phenomena show how drives act into the field. Boredom arises when the preconscious no longer finds anything salient; it is not idling but a signal that drives the search for new goals [12] – the exploratory drive making itself felt. Disgust begins as bodily protection against spoiled food and is taken over socially and morally: an unfair offer triggers the same facial muscle movement as a bitter taste [39]. Here a bodily drive travels via the emotions all the way into judgment.

Two neighbouring concepts do not belong to the drives. Homeostasis keeps set points constant; allostasis adapts the organism to strain and, under chronic strain, has its price, allostatic load [118, 149]. Both are mechanisms of the carrier. A drive, by contrast, is directed and sets the field in motion. Motivation, finally, is the drive as it is evaluated in the field and related to a goal.

The groups are developed to different degrees across species. Bodily drives and reproduction are found already in the fly, exploration and play in many mammals and birds, rank and territory in socially living species. A metaphysical need – for meaning, for the whole, for what comes after death – is documented only in humans. Kant called it a “natural disposition” to metaphysics (Critique of Pure Reason B 21) [95], and religious ideas are found in all known cultures [25]. Whether it is a drive of its own or the exploratory drive reaching, with reason, beyond all experience is open; we know of no dedicated circuit. We treat it as a hypothesis. AI language models have no drives in this sense: they lack a body that could lack anything, and their goals come from outside. When they speak about meaning, they reproduce the trace of human drives in their training data, not a drive of their own.

5 Case analyses

The following analyses can be traced in Figure 1 using the “lens”.

5.1 Conscience

Conscience is a coupling: reason supplies principles, judgment subsumes one's own action under them, emotions supply guilt and shame, social cognition contributes the internalized view of others, prospection and memory produce the warning and the bad conscience, and the will carries out the verdict. Kant defines conscience as “moral judgment passing judgment upon itself” (our translation) [98]; in Freud, the coupling corresponds to the superego [67]. The division of empathy implies: if cognitive empathy remains intact while affective empathy is impaired, harm is understood but not felt – the profile described for psychopathy [19]. When people with psychopathy are explicitly asked to empathize, their brain activity approaches that of controls [119]. Affective empathy is thus not missing; it simply does not run automatically. The difference lies on the control axis.

5.2 Self and soul

The self is not a domain but the unity of the whole configuration. Kant frames this unity formally: “The ‘I think’ must be able to accompany all my representations” (B 131; our translation) [95]. Whether there is a soul as an immortal substance is something the model does not decide: such a soul can be neither represented nor tested in it, and Kant, too, treats its immortality as a postulate of practical reason [96]. What the model can explain is why belief in a soul and in the continued presence of the dead – in ancestor veneration, for example – is so widespread: social cognition attributes an inner life to persons, and this attribution does not automatically end with death; even children continue to attribute mental states to the deceased [13]. The explanation concerns the belief, not its truth.

5.3 The DMT state

The state induced by DMT is neither a place nor a new domain but an extreme shift on the axes. Internally generated content receives the status of reality, often of certainty; control is automatic, the boundary of the self dissolves, and the time axis drops out. The REBUS model describes this as a relaxation of high-level predictions [34], and imaging findings show a dissolution of the usual network boundaries [154]. Many describe their encounters as completely real; in a large survey, most respondents said the experience had altered their fundamental perception of reality [47].

The entities reported recur in a small number of types [107], which the model reads in three layers. The most frequent report is an unspecified presence – “someone is there” – without further content: social cognition detects an agent even without input, as with the illusory shadow person produced by stimulating the temporoparietal junction [3]. Above this lie species-typical templates: a female or caring presence (attachment), deities (superior power and meaning), animals and creatures (old detectors of predators and danger), geometric and luminous beings (patterns of early visual cortex that psychedelics reliably evoke [26]) and tricksters who break expectations. The uppermost layer is cultural filling: Hindu or Egyptian deities, aliens, robots and “machine elves”, a term popularized by Terence McKenna. That many people meet the same types does not make the entities part of the world: agreement among identically built brains under the same substance reveals their shared architecture (Section 2). Whether the experienced space exists independently is a question about its actual origin, to which the criteria of Section 2 apply.

5.4 Autism

The model describes autism as a different configuration rather than the failure of a domain. Social cognition runs controlled rather than automatic more often, which explains why social adaptation takes effort [85]; salience and prediction weight sensory details differently [132]. Since social cognition is calibrated to one's own type, the misunderstanding between autistic and non-autistic people is mutual [124].

5.5 Freud and Jung

Freud's topographic model is the depth axis; his structural model is distributed over couplings [67]: the id corresponds to the unconscious and the emotions, the ego to will and understanding, which mediate the reality principle via the modality axis, and the superego to the coupling of conscience. Defense is a blocking of access in the preconscious. Jung's four functions [93] can be assigned to domains: thinking to understanding, feeling to the emotions, sensation to perception, intuition to prospection together with unconscious pattern recognition. The model accepts the collective unconscious only in a naturalistic reading, as species-typical structures shaped by evolution [112].

5.6 Borderline cases of modality

In dreams, internally generated content is attributed to the outer world while reason and will barely check it; in lucid dreams, reflection and control partly return [159]. In the out-of-body experience, perspective shifts; that it could be induced by stimulating the temporoparietal junction [21] supports reading it as a construction of the self-model.

5.7 Lesions and stereotypies

Lesions are the hardest test for a functional architecture, because they show what can be separated. Split brain: after the corpus callosum is cut, the speaking left hemisphere readily invents reasons for actions initiated by the right [71]. An action whose actual origin is inaccessible receives an invented reason that is taken as real; explanations of one's own actions are constructed, not read off. Neglect: after right-sided lesions, the left half of space is ignored although vision is intact [43] – even when imagining a familiar square. This fits a disorder of the preconscious, a layer beneath all domains. Agnosia: objects are seen but not recognized [58]; a patient with form agnosia could not state the orientation of a slot yet posted a card into it accurately [74]. Perception can thus lead into action via will and the motor system, without access and without a concept. Parkinson's disease: automatic control of movement is lost first [137]; yet patients step confidently over lines on the floor. What no longer runs automatically is triggered in a controlled way; the motor system itself is largely intact.

The opposite pole is formed by stereotypies in animals kept in barren or confined conditions, such as the “weaving” of elephants after years in chains [117]. Here a movement pattern keeps running automatically and the will can no longer stop it, often long after conditions have improved. Parkinson's disease and stereotypy thus lie at opposite ends of the same control axis: in one, the automatic is missing; in the other, it can no longer be braked.

Paired cases. Lesions are most informative in pairs that mirror each other. In blindsight, visual information guides behaviour without conscious access [162]; in Anton's syndrome, patients claim to see although cortical vision has failed [2] – perception without access against “access” without perception. In locked-in syndrome, consciousness and will are preserved while the motor output is blocked; in akinetic mutism, the motor system is intact but the drive to act is missing [46]. In the Capgras delusion, a familiar person seems to be an impostor [56]; in the Fregoli delusion, strangers seem to be a familiar person in disguise [44]. Each pair separates what the model keeps apart: perception and access, will and execution, recognition and the feeling of familiarity.

5.8 Character and its attribution

We attribute character traits to other people and to our pets: someone is anxious, conscientious or suspicious; the dog looks guilty. The model offers a simple reading. A state is a momentary shift on the axes; a trait is their typical resting position, together with the strength of certain couplings. Anxious then means a resting position of valence towards the negative and a strongly reacting carrier; conscientious, control towards the controlled pole and time towards the future; impulsive, control towards the automatic pole with weak inhibition by the will; resentful, time fixed on the past and coupled to negative valence; gullible, a modality that quickly sets “real”; suspicious, an authorship that readily attributes hostile intent to others. Personality research describes traits in a similar way, as typical distributions of momentary states [62]. One trait does not fit: sociability, part of extraversion [90], cannot be described by valence alone; it additionally needs the motivational direction of approach versus avoidance within the emotions (4.2) and a strong social attachment drive (4.9).

With animals, the model predicts which attributions are plausible: only those that require domains the animal has. A dog's “guilty look” would require the coupling of conscience, and thus reason; in experiments it follows the owner's scolding, not the dog's own misdeed [84]. Jealousy-like behaviour, by contrast, needs only emotions and elementary social cognition, and it has been observed in dogs [77]. That we nevertheless see guilt or cunning in animals is the same tendency of social cognition that detects agents behind events (Section 5.3).

6 Comparative analysis: From flow to field

If the linear flow of information is taken as the basic form, the Cognitive Field is its more deeply recursive form. A comparison across species makes this visible. The assignments are our reading of the research, not measurements; we attribute a domain only when learning performance, neural findings and the exclusion of simpler explanations fit together. Instead of a single gradient we distinguish, as Birch and colleagues proposed for consciousness in animals [18], several dimensions: domain inventory, coupling density (how strongly the field is integrated), recursion depth (how often the system turns back on its own processes) and axis range (for example how far the time axis extends into the future). The gradient in Figure 1 is a simplified projection of this space.

Coral and fly. Corals have no brain, only a nerve net without a centre; they mark the zero point of the scale, the pure reflex arc. The brain of the fly Drosophila, with about 140,000 neurons, has been fully mapped [55]; perception, associative learning, valuation and action selection are demonstrated, prediction only in the short term in the sense of reafference. Whether insects experience anything remains open [8].

Fish and reptiles. Fish are the oldest vertebrates. They show pain responses that go beyond a reflex [146], and a cleaner wrasse passed a version of the mirror test whose interpretation is disputed [102]. In reptiles, spatial memory and first evidence of social learning are added [100].

Octopus. The octopus is the most distant comparison. About two thirds of its roughly 500 million neurons lie in the arms [72]; much control is decentralized. Nevertheless it uses tools, such as coconut shells carried along as later shelter [61], and findings point to affective pain experience [45]. The octopus shows why one dimension is not enough: a considerable domain inventory but little central integration.

Corvids and great apes. Ravens plan flexibly for the future [94]; scrub jays re-cache food if they were watched, especially if they have stolen themselves [57]. These feats rest on a pallium without a layered neocortex but with cortex-like circuits [148]. The Cognitive Field has thus arisen independently at least twice. Great apes save tools for later [125], recognize themselves in a mirror [70] and take others' false beliefs into account [103]. The boundary with humans lies not in the domain inventory but in recursion depth and axis range.

AI systems (as of October 2026). Large language models show a reversed profile: evolution built from the bottom up, language models start at the top, with language, understanding and judgment. Performance on theory-of-mind tasks has been documented [150]; its interpretation is disputed. In the research literature, by contrast, superintelligence denotes systems that surpass humans in almost all cognitive domains [24]¹; in the model, that would mean all nine domains at or above human level. In this sense, the term does not apply to today's language models.

At the bottom, almost everything is missing: no body, no interoception, no valuation that regulates survival. The weak point lies on the modality axis: generated content is sometimes output with the status of fact [89]. We describe this functionally, at the level of behaviour; it does not presuppose an inner attribution mechanism like the human one. Introspection is demonstrable but unreliable [110]. For embodiment we distinguish three stages: the language model in dialogue; the agent that acts through tools and meets real constraint, for example a failing program test; and the robot with sensors and motor system. The model expects the gap between actual origin and attributed status to shrink from stage to stage, because the world constrains the system's assumptions from outside more often. Whether such systems have experience cannot be decided with current methods [32]. Since this section was written with the participation of such a system, its self-reports should be treated as data of uncertain reliability.

¹ Since September 29, 2026, the US federal government has used the term “Super Intelligence” (SI) instead of “Artificial Intelligence” by executive order; for now, the existing statutory definition of AI applies to it [164]. This usage is unrelated to the concept in the research literature. ↩

7 Testable predictions

For each prediction we state a competing explanation of the same finding and the condition under which it would be refuted.

1. Gradient and convergence. Domains appear step by step in evolution and can arise anew on a different neural foundation; corvids plan and take observers into account without a neocortex [57, 94, 148]. Alternative: specialized adaptations to food caching rather than general capacities. Refuted if far-reaching prospection can be reliably demonstrated in species with low recursion depth, or if the feats of corvids can be explained as simple association.
2. Coupled domains fail together. People with hippocampal amnesia struggle to imagine new situations [78]; other patients, however, produced normal future scenarios [147]. Alternative: joint failure due to shared anatomy rather than functional coupling; the prediction only becomes discriminating for couplings between spatially separate networks (Section 4.8). Refuted if lesions of the connecting pathways regularly leave the coupled domains independent of each other.
3. States are axis shifts. Altered states of consciousness create no new domains but shift values on the axes; experience reports can be ordered in low-dimensional state spaces [126]. Alternative: other dimensional models such as the 5D-ASC [54] order the same reports without domains. Refuted if state reports cannot be ordered by a few axes or contain abilities that cannot be assigned to any domain.
4. Prediction acts across domains. If the prediction of one's own actions is disturbed, perception, will and self-experience are affected together; with passivity experiences, self-tickling is often not attenuated [20]. Alternative: a later, faulty attribution of authorship rather than disturbed prediction [151]. Refuted if such disturbances can be confined to a single domain.
5. Empathy is twofold. Affective and cognitive empathy can be impaired separately: in autism, cognitive empathy tends to be affected, in psychopathy affective empathy [19]. Alternative: a single empathy that in psychopathy is merely not deployed spontaneously [119]. Refuted if both forms always vary together.
6. The sense of inexplicability follows transparency. How strongly an experience is described as immediately given and inexplicable depends on whether its construction is transparent [37, 123]. Alternative: a culturally learned or general introspective illusion. Refuted if such reports remain equally strong where people experience their perception as constructed, for example in lucid dreams or in meditation.

8 The model and the mind–body problem

The hard problem is today's form of the old mind–body problem. The model answers half of it. The body is the carrier of the field, the mind is the field itself – what the embodied organism does; the soul, insofar as the model speaks of it, is the unity of this configuration (Section 5.2), not a second substance. A dualism like Descartes' [50] is therefore unnecessary: body and mind are not two things that must interact, but carrier and carried. In the terms of the philosophy of mind, this is a non-reductive monism: there is one world, but the mind is not simply physics. What remains open is the other half: why what is carried is experienced – why there are qualia.

Chalmers distinguishes the “easy” problems of consciousness – access, attention, reportability – from the hard problem of why these functions are accompanied by experience at all [36]. Wakefulness and access are easy problems; experience is the hard one. We place the model on three levels.

Level 1: When and where consciousness occurs. Here the model is an integrative architecture: the preconscious corresponds to the access function of the global workspace [5, 49], the circle of prediction and comparison to predictive processing [41], reason as metacognition to higher-order theories [30]. That none of these theories prevails on its own is shown by a large preregistered adversarial test that failed to confirm parts of both global workspace theory and integrated information theory [42].

Level 2: The meta-problem. Chalmers proposed first explaining why beings like us consider the hard problem to be a problem at all [37]. Here the model makes a contribution of its own. On the perspective axis, the system experiences its states as those of a self. On the modality axis, it gives its own models not the status of constructions but that of the world as simply given – Metzinger calls this the transparency of the self-model [123]. A system with both properties must describe its experience as immediate, private and physically inexplicable, regardless of the metaphysical answer.

Level 3: The hard problem itself. Here the model remains open. Qualia are exactly what the hard problem would have to explain; the axes describe only how they are given – toned (valence) and marked as real or merely possible (modality). Structured Realism suggests a direction: if we know the world above all in its structure [141], its intrinsic nature remains open. Russellian monism suggests that experience itself could be this intrinsic nature [1] but inherits the combination problem. We name it as an option without committing to it.

9 Objections and replies

9.1 The Kantian circle

Objection. The evolutionary derivation of the forms of cognition presupposes space, time and causality – exactly the forms it means to explain. For a Kantian, evolution, too, is mere appearance.

Reply. We concede the circle and deny that it is vicious. There is no standpoint outside all cognition; Neurath compared science to sailors who must rebuild their ship on the open sea [127], and Quine turned this into the program of a naturalized epistemology. Such a circle is admissible if it could have failed: if our notions of space, time and causality were ill-fitting, predictions based on them would systematically fail. Moreover, the invariance criterion supports the realist reading independently of any single system. A tree falls at the time of the dinosaurs, some hundred million years before the first human. Dinosaur, early mammal and bird hear different sounds but agree in registering place, time and sequence: first the fall, then the pressure wave, then flight. This also answers the question associated with Berkeley's principle that to be is to be perceived [14]: a tree that falls unheard produces a pressure wave, for that belongs to the world, but no sound, for that belongs to a hearing system. Statements about events before any consciousness, such as the formation of the Earth, are in any case hard to accommodate for positions that tie being to consciousness [120]. All this is an inference to the best explanation, not a proof. The circle is not vicious, but it is not removed either.

9.2 Interface theory and the simulation hypothesis

Objection. In evolutionary game-theoretic models, Hoffman and colleagues show that perception tuned to fitness drives out perception that represents the world truthfully; space and time are only a user interface [82, 135]. The simulation hypothesis goes further: we probably live in a computer simulation [23]. Both deny that the world in space and time is the real one.

Reply. Hoffman's argument undermines itself, because it presupposes what it denies. His theorem is mathematically correct within its assumptions, but those assumptions are populations of organisms that reproduce over generations and compete for resources – at least a real world with time, succession and causality. Whoever uses such a model to show that space and time are not real saws off the branch they are sitting on [6]. Hoffman himself exempts logic and mathematics from his argument; but then evolved capacities can track the truth, and the only question left is in which domains. The invariance criterion offers an answer.

The simulation hypothesis cannot be refuted. Four considerations show, however, that it does not threaten Structured Realism and explains little. First, realism remains. Had no asteroid struck the Earth 66 million years ago and darkened the atmosphere for years, the dinosaurs would have continued to flourish and the small mammals would have stayed small. There would then be no primates, no humans, no mathematics, no physics, no computers – and therefore neither Hoffman's simulations nor a simulation hypothesis. Both ideas are late products of a contingent history of living beings on a planet that, over millions of years, brought forth different stages of the Cognitive Field, up to humans who began writing software for chips only a few decades ago. Every simulation presupposes a real base world with such a history. Bostrom concedes this himself: his hypothesis does not deny that there is a real world, only that we live in it. The question of reality is shifted, not answered.

Second, the hypothesis immunizes itself. Any finding that would count against it can be explained as part of the simulation. Russell's five-minute hypothesis, according to which the world might have come into being five minutes ago together with all our memories [140], and Gosse's Omphalos argument, according to which fossils were created together with the world [75], have the same structure. Such hypotheses cannot be refuted, but nothing can support them either; no wall can constrain them. By the constraint criterion of Section 2, they are therefore not empirical claims about the world. Even quantum physics, which shows that atoms are mostly empty space, does not refute the wall one walks into: it explains why the wall is hard. And it was developed by beings whose cognitive structures were shaped by such walls over millions of years.

Third, the hypothesis is selectively sceptical. Bostrom's probability estimate relies on our physics to gauge how much computing power future civilizations might have, while declaring the world this physics comes from probably simulated. There is no good reason for this choice of which evidence to trust and which not [17].

Fourth, even a simulated world would be real in the sense of Structured Realism. Chalmers has argued that objects in a simulation are not illusions but digital objects with genuine causal powers [38]. A simulated wall constrains too, and what stays the same across cognitive systems would stay the same then as well. The criteria of Section 2 therefore do not distinguish between a simulated and a non-simulated world, and they need not: they determine what is real for us and which structure we recognize. If Bostrom were right, Structured Realism would remain untouched; only the location of the base world would shift.

9.3 Embodiment and enactivism

Objection. According to Merleau-Ponty, the lived body is the medium through which we have a world at all [122]; enactivism understands cognition as the sensorimotor activity of an embodied organism, not as the processing of inner representations [153, 157]. A diagram of boxes is a “head model”.

Reply. We partly accept the objection. The field is embodied, interoception belongs to perception, the motor system closes the circle with the world, and that circle comes from Dewey and von Uexküll. We considered a separate body layer in the diagram and rejected it: the body is not a layer alongside the preconscious but the carrier of the whole field. Figure 1 therefore shows it as a band beneath the field rather than as another box – which matches Merleau-Ponty's point better. One difference remains: the model holds on to inner models, because prospection, memory and dreams can hardly be described without them. In dreams, the circle with the world is interrupted, and yet rich experience arises.

9.4 Parsimony

Objection. According to the free-energy principle, every self-organizing system minimizes prediction error; perception, action and learning are only different ways of doing so [68]. Why, then, nine domains and seven axes?

Reply. The approaches lie on different levels. Marr distinguished the task a system solves from the procedures and subsystems with which it solves it [116]. The free-energy principle applies to a bacterium as much as to a human being and therefore does not say which separable capacities a human being has; lesion findings show that this question must be asked. The model also incorporates predictive processing, in the circle with the world and in prediction 4. Our reply would be refuted if prediction models without a domain-like subdivision produced the same dissociations as real lesions.

10 Discussion

The model's strength lies in three points. It connects a conceptually rich tradition with the recursive picture of today's cognitive science. It makes explicit what is a domain and what an axis, and thus becomes open to criticism. And it classifies cases from ethics through lesions to AI without introducing new building blocks for each case.

The limits are just as clear. The contribution is conceptual; the predictions are derived but not tested by us, and two of them (2 and 4) cannot yet be distinguished from competing explanations. The boundaries between domains are debatable, especially between judgment and reason and in the case of language. Whether the seven axes are complete and independent of each other is open. The realist reading of Kant is a strong thesis against the usual interpretation. The assignments in Section 6 rest on our interpretation of the literature. And the hard problem of consciousness remains open.

Candidates under examination. External reviews of version 4.5 by ChatGPT, Gemini and four further models proposed additions. We examine them without adopting them yet, because every additional building block makes the model harder to refute. Domain candidates are language, defined narrowly as symbolic-compositional processing [59], and core number, with acalculia in humans and numerical abilities in animals. Drives, formerly also a candidate, have been adopted in version 4.8 as a separate layer rather than a domain (4.9); allostasis, as adaptation to strain, belongs to the carrier. Of the axis candidates, authorship has been adopted (4.2). Approach versus avoidance was examined and not adopted as an axis. It does not coincide with valence: anger is negative yet drives approach [35], and people deliberately seek out hot food, horror films and sad music, stimuli that are at the same time experienced as unpleasant [139]. But it demonstrably shifts only emotions and will, not all domains, and therefore counts as the motivational direction within the emotions (4.2). Two later reviews independently proposed processing direction – whether expectation (top-down) or input (bottom-up) dominates. It may pass the independence test, since the Charles Bonnet syndrome is driven top-down while modality remains “not real”; but in predictive processing it is the basic mechanism of all processing rather than one parameter among others. Both reviews also proposed a self-model (body ownership, sense of agency) as a domain; we examine whether it can instead be described as a bundle of the axes perspective and authorship. Construal-level theory distinguishes temporal, social, hypothetical and spatial distance [155]; three of these correspond to our axes of time, perspective and modality. Its core prediction, that distance makes representations more abstract, was however barely confirmed in a large multi-lab replication [33]; the parallel is therefore a suggestion, not support. One review suggested a more fundamental reorganization: separating what is represented (body, space, number, language, other agents), what is done with it (perceiving, storing, simulating, evaluating, selecting, executing – Kant's faculties would belong here as operations) and the state parameters under which this happens (the axes). We are testing this internally.

A leave-one-out test. To check whether our domains suggest themselves without prompting, we presented three language models of different sizes (Claude Opus, Sonnet and Haiku), in 45 runs, with the model minus one domain and asked them, without mentioning the deletion, to name the missing building blocks; partly with and partly without domain names, plus five control runs with the complete model. Perception, memory, social cognition, emotions and will were almost always restored in first place. Understanding was found only as language, reason only as metacognition and at middling ranks, prospection weakly – and judgment in no run at all. Conversely, almost all runs proposed spatial cognition, a self-model, working memory and number. The test measures what language models consider obvious from the literature, not what is correct. But it shows where our division depends most on the Kantian tradition and needs the most justification; in version 4.8 we have therefore re-grounded judgment without appeal to Kant, as the application of knowledge to the case (4.1); its boundary with reason remains open.

11 Conclusion

In this model, the mind is not a sequence of stages but a system: nine functionally defined domains in a recursive field, seven axes describing how they work at any given moment, and a circle closed through body and world. Consciousness is not one thing in it but four different things in four different places. The world exists independently of us, and our structures were shaped by it. The model explains experience without deciding about what lies beyond experience, and it names the observations that would refute it.

12 Outlook: testing and recognition

Origin. Sebastian Bohnen is an independent thinker without affiliation to a university or research institution. The model was developed outside academia, in an ongoing dialogue with AI systems that searched the literature, formulated counterarguments and translated. This is a strength and a weakness at once. The model could cross disciplinary boundaries – between Kant, Freud, neuropsychology, comparative cognition and AI – that an academic career usually keeps apart. But it has not yet passed the scrutiny that academic work normally undergoes: so far it has been reviewed by AI systems, not by human experts.

The path to recognition. What holds for the world in Section 2 holds for a theory: it proves itself against what pushes back. We see five steps. First, publication as a preprint on an established server with a persistent identifier (DOI), such as PhilArchive or PsyArXiv. Second, review by human experts from philosophy, neuropsychology and cognitive science, explicitly asking for refutation. Third, testing at least one prediction from Section 7 against existing data, most readily the connectome prediction (4.8) on publicly available datasets. Fourth, fixing predictions before each new case analysis, so that the model does not explain everything after the fact. Fifth, submission to a peer-reviewed journal, observing its rules for disclosing AI involvement.

AI and science. The capabilities of AI systems are growing fast: the length of tasks they complete autonomously has doubled roughly every seven months since 2019 [104]. That the US government has officially called such systems “Super Intelligence” since September 2026 (footnote 1) shows how seriously the development is taken politically – even though the label does not mean that superintelligence in the sense of the literature has been reached. Our model is an example of one consequence: theories can now arise outside institutions, with a breadth of literature and cross-checking that would once have required a research group. For universities, teachers and students this shifts the task – less providing knowledge, more testing, weighing and taking responsibility. They will have to redefine their role, and faster than their usual cycles allow.

As a testing authority they remain irreplaceable, and our own model says why. AI language models have a reversed profile (Section 6): strong in understanding and judgment, weak on the modality axis, prone to presenting generated content as fact. What they lack is the resistance of the world – the experiment, the measurement, the judgment of people who bear responsibility. A model developed with AI therefore needs, all the more, a test that does not come from AI. That is the test we ask for.