Chapter 11 — The Offline Mind
§11.1 The Frozen Junction
Consider a mind that generates trajectories it cannot feel the cost of. A simulation engine turns, with whatever fidelity its generative machinery affords, producing candidate futures, counterfactual pasts, inferred perspectives of absent others, and it does this at speed and at volume, with the internal coherence of the outputs often exceeding the coherence of anything the organism's actual environment has recently supplied. The junction where these trajectories would be priced — where the interoceptive apparatus Chapter 10 installed would read the simulated trajectory as if it had been traversed, compress the reading into a precision-weighted valence signal, and broadcast that signal across the distributed cognitive regime that weights attention, memory, and action — is frozen. Not broken. Frozen. The simulation runs. The Witness is there. The pricing does not pass. The trajectories are produced without felt consequence, and the organism, or the architecture, proceeds to act on the unpriced products as though the pricing had been done.^1
The phenomenology from the inside is neither vivid nor conspicuous, which is the first thing to say about it. A patient with bilateral ventromedial prefrontal damage, on an Iowa Gambling Task, reports that deck A and deck B look roughly the same, reasons about them aloud at unimpaired propositional quality, draws the inference that the long-run expected values should converge under the experimental design, and then continues to draw from the disadvantageous deck across a hundred trials without the anticipatory skin-conductance response controls develop within the first forty. The patient is not confused about the task. The patient is not confused about the reasoning. What is missing is the thin continuous pricing that would ordinarily have weighted the simulated next draw before the draw was committed — the simulated outcome's valence signal, published into the broadcast limb, biasing the action selection by a margin the patient does not introspect but that governs the selection nonetheless. The simulation runs. The junction where the simulation's output would have been weighted against the organism's own felt stakes is not operating. The trajectories emerge, and the organism acts.^1
A healthier case sits closer to the reader's own phenomenology, and is closer to what the present chapter is obliged to describe. Consider rumination — three in the morning again, but a different three in the morning from the one Chapter 10 opened on. The earlier three in the morning was acute autonomic alarm at maximum precision gain on a mis-calibrated interoceptive channel. This one is cognitively furnished rather than somatically flooded. A simulated conversation runs — a scene with a colleague, or a parent, or a version of oneself the self is no longer on speaking terms with — and the scene iterates. The first iteration produces something that would, under ordinary pricing, register as resolution or as sufficient account-taking to close the loop; the second iteration undoes the first; the third rebuilds the first; the fourth undoes it again; and somewhere around the thirtieth the ruminator notices that the iteration has been going for forty minutes and that no version of the simulated scene is being priced at a weight that would permit the loop to terminate. The Witness is present — the ruminator is miserable, and knows it — but the misery is not specifically about any individual iteration of the loop. It is the chronic expenditure of running the loop whose pricing does not terminate it. The simulation engine produces trajectories; the Witness prices a common aversive flatness across all of them; the differential pricing that would have selected between trajectories and exited the loop by landing on one is absent. The junction is not frozen in the VMPFC-lesion sense. It is stuck in a different way — the Witness continues to deliver a global valence reading, but the local differentiation that would have adjudicated among simulated alternatives has been lost to the loop's own metabolic dynamics.^1
These are two shapes of the same architectural failure, and they are the chapter's opening case because they disclose, by what they lack, the closure the chapter has to install. A functioning offline mind is not a simulation engine. It is a simulation engine coupled, at a specific and diagnostically legible seam, to the affective stratum installed in Chapter 10, in such a way that the Witness prices simulated trajectories under the same broadcast regime it prices traversed ones, with the fidelity of the pricing falling off in specifiable ways as the distance from the traversed increases. The coupling is the closure. The closure is what Chapter 11 installs. The frozen junction — in its acute lesion form and its chronic ruminative form — is what the stratum looks like when the closure is installed and operating at the seam. A mind that did not have the closure at all would not ruminate; it would simply not simulate. The failure is a failure of a real capacity, and the capacity's reality is what the failure is evidence of.
What this chapter therefore owes is the specification of the capacity at the resolution its failures require. The capacity has two faces, and the chapter will install them as one. The first face is the decoupling of the cognitive stream from its real-time sensorimotor coupling, such that the organism can generate trajectories whose motor commitment has not been made and whose sensory returns are not being checked against the world's feedback — the substrate of counterfactual planning, of prospective simulation, of episodic re-enactment, of what the literature calls, with varying degrees of theoretical loading, mental time travel. The second face is the decoupling of the cognitive stream from its first-person frame, such that the organism can generate trajectories from within a simulated other mind — the substrate of perspective-taking, of theory of mind, of the recursive mentalisation that social primates run at some of the world's highest computational densities. The two faces look, from most contemporary literatures, like two separate capacities served by two separate networks. The frameworks's reading is that they are the same closure at two orientations of its decoupling operator, with the same Witness-Canon coupling problem at their shared seam, and the same failure taxonomy when the coupling fails. §11.3 and §11.4 install the faces separately at the level of empirical mechanism; §11.5 reads them together as the one closure they are. The frozen junction of the present section is what the closure is not, seen from the inside of the architecture where the closure almost is.
§11.2 Why Any Organism Would Pay for This
The installation cost of an offline stratum is not small, and the framework's anti-teleology clause, inherited from Chapter 8's prohibition against free Mediation, requires the cost to be named before the capacity's yield is described. The cost is thermodynamic, architectural, and representational at once. Thermodynamically, the mammalian brain's default-mode network — the cortical midline and lateral-parietal ensemble that activates during task-negative conditions and that the offline-cognition literature, since Marcus Raichle's 2001 identification of the network's baseline metabolic signature, has mapped as the substrate of spontaneous prospection, autobiographical recall, and mental simulation — runs at a metabolic rate comparable to the task-positive networks whose operation the organism's immediate survival depends on, and runs continuously, at a cost the organism pays whether or not any specific simulation delivers an adaptive return. Architecturally, the coupling conditions that permit the hippocampus to replay place-cell sequences decoupled from the animal's current location, and that permit the prefrontal cortex to sustain an internal model of a future or counterfactual state across tens of seconds against the pressure of incoming sensory drive, require a specific precision-weighting regime on the inputs from primary sensory cortex and from the limbic afferents that would otherwise dominate the ensemble's activity — a gain-control architecture whose tuning is itself metabolically expensive and whose failure, in either direction, produces specific clinical signatures the chapter will catalogue. Representationally — the word used here loosely, and with a reservation 11.6 will sharpen — the organism that decouples from the present sensorimotor stream forfeits, for the duration of the decoupling, the real-time error correction that the embodied stream would have been providing; the simulated trajectory's fidelity to the world is no longer being continuously checked against the world's feedback, and the architecture therefore requires an additional mechanism by which the simulation's drift from the world is bounded, or the simulation becomes, within a small number of iterations, decorrelated from any reality the organism might usefully act on. The three costs — metabolic, gain-control, drift-bounding — are joint, and a stratum that pays any one of them without paying the other two delivers no adaptive yield. The installation is all-or-nothing in the specific sense that a partial installation is worse than none.\2
The temptation, at this point, is to narrate a yield — to describe the new niches the offline stratum opened, the delayed-reward environments it permitted the organism to exploit, the social-contingency landscapes it made legible for the first time, and to proceed as though the narrative explained why the installation occurred. The framework's anti-teleology clause forbids this narrative order. The yield is real, and the chapter will describe it. But the chapter's obligation is to open the discussion of yield with a plateau, and the plateau is to describe an environment in which the stratum's yield is neutral or negative — in which the organism that could pay the installation cost would find its metabolic budget eroded without proportionate return, and in which selection pressure would therefore drive the stratum toward regression, not elaboration. The plateau case is the case the selection-theoretic literature has tended to elide, and its elision is what makes the elaboration-story look like explanation when it is description.
Consider the ecology of a small-bodied, short-lived animal whose entire trophic niche is characterised by reliable cue-reward contingencies at the timescale of the organism's foraging excursions. A shrew, to take a concrete case, or a small insectivorous bird in a stable understory. The action-space is well-sampled across a cohort's foraging lifetime. The reward schedule is densely decorated with proximal cues whose covariance with nutritive content is high enough that the interoceptive pricing 10.2 installed — graded, precision-weighted, continuous — absorbs essentially all of the adaptive learning the organism's niche rewards. There are no trajectories whose first traversal is lethal at rates the organism's cohort cannot amortise. There are no conspecific-contingency problems whose solution requires a model of another animal's concealed state, because the relevant conspecific behaviours are either transparent to direct observation or are not load-bearing for the organism's fitness in the niche. In this ecology, the metabolic budget an offline stratum would consume is budget spent on a capacity whose adaptive yield is zero or worse. Not "selected against because simulation is wasteful"; selected against because the specific architectural achievement that would let simulation pay for itself has no environment in which to pay. The shrew does not simulate. The shrew's phylogenetic lineage, across tens of millions of years in the niche the shrew inhabits, has been under continuous selection pressure not to. The absence is not an absence of evolutionary capacity; it is the specific presence of a niche in which Stratum 4 and Stratum 5 are a fitness cost with no compensating yield. The plateau is the environment in which the offline mind does not get installed, and it is the majority environment across most of the animal phylogeny the chapter is discussing.
The selective regime under which Strata 4 and 5 do get installed — the regime in which the installation cost is paid because the cost is less than the yield — has three specific architectural features, and each of the three has to be present for the installation to be favoured. The first is delayed reward. An organism in a niche where the highest-yield trajectories have their payoff at a temporal offset greater than the interoceptive pricing apparatus's calibration window — where the food cache, the seasonal migration, the mating aggregation, the toolmaking sequence all deliver their payoff at a delay measured in hours to weeks rather than seconds — has a niche in which offline pricing of trajectories is adaptive because the online apparatus cannot weight the trajectory at the timescale the trajectory requires. The corvid caching literature is the canonical case: scrub jays' capacity to cache food items at specific locations, to track which caches contain which items, to remember which caches have been observed by which conspecifics, and to re-cache items observed-being-cached by potential pilferers, is the behavioural fingerprint of a stratum whose offline pricing machinery is running on a timescale the animal's immediate interoceptive apparatus would not reach. The second feature is environmental novelty. An organism whose niche is continuously re-described by environmental change — by ecological transition, by migration, by the entry of novel predators or prey, by the construction of environments the organism's own conspecifics are reshaping at generational speed — has a niche in which the traversed-trajectory pricing Chapter 10 installed cannot keep pace with the environment's drift. The organism that can simulate never-taken trajectories, price them against a learned generative model of the environment's likely dynamics, and commit the action before the traversal is made, has a fitness advantage over the one that has to traverse in order to learn. The third feature is cached social contingency. An organism whose fitness is load-bearing on the behaviour of other organisms of its own species — on coalitional dynamics, on reciprocal exchange, on deception and its detection, on the nuanced reading of intent from cue — has a niche in which the affordances of the social environment are not legible from the organism's own first-person interoceptive apparatus, because what the affordances are is the other organism's concealed state, which the organism has no direct channel to. The organism that can simulate the other's mind — build a generative model of the other's likely action under the other's likely internal state — has a fitness advantage over the one that must observe the behaviour and infer post hoc. The social-brain hypothesis in Robin Dunbar's formulation, the coalitional-intelligence literature in Frans de Waal's, the Machiavellian-intelligence programme in Whiten and Byrne's, each specifies versions of this third feature with empirical stratification across primate lineages.\3
The three features are not independent, and their joint presence is what the selective regime requires. An ecology with delayed reward but no social contingency and no environmental novelty produces capacities for prospective cacheing and temporal discounting without full Stratum 4 elaboration — a stratum whose simulation capacities are narrow, whose bandwidth is small, and whose failure modes are not the ones the rest of the chapter catalogues. An ecology with cached social contingency but stable environments and short reward horizons produces capacities for direct-perception reading of conspecific state without the recursive simulation Stratum 5 installs — a lower-depth mentalising stratum whose signature is present in many social mammals and whose limits the primate lineage exceeds. An ecology with all three, at population-scale density across a lineage's evolutionary history, produces the selective pressure under which the metabolic, gain-control, and drift-bounding costs of the offline stratum are paid because the yield — legible to the organism in the form of survival through trajectories not traversed, exchanges not yet made, and coalitions not yet engaged — exceeds the cost at a margin the lineage's continuation depends on. The primates are the clearest instance. The corvids are a convergent instance at a very different lineage with similar selective features. The cetacean-lineage instance is likely and the empirical record is thinner. The hominin lineage is the instance at the book is downstream of, and the elaboration of Strata 4 and 5 across that lineage is the substrate on which Chapter 12 will install the Stratum 6 symbolic apparatus.
What the plateau case is for is to prevent the chapter from being read as a story in which the offline stratum arose because it was adaptive. The offline stratum arose in the lineages whose ecologies paid for it, and remained absent from the lineages whose ecologies did not. The causal order is the ecology's, not the capacity's. The stratum is not the answer to a question evolution was posing; it is the specific architectural achievement whose installation was selectively favoured in the specific niches where its three-feature pricing conditions obtained. The narrative tempation — to read the stratum as adaptive because it is elaborate — is the same temptation Chapter 8 named at the bioelectric stratum and Chapter 9 named at the navigational one. The framework refuses the temptation at each stratum on the same grounds: architectural closure is not teleological achievement; it is the specific structural outcome of specific priced conditions whose absence produces no closure and whose presence produces the one the chapter specifies. The plateau precedes the yield because the yield is not the explanation.\3\2
§11.3 Mechanism Sketch for Stratum 4: Decoupling from the Present Sensorimotor Stream
The neural substrate on which Stratum 4's decoupling is installed is not a single region and not a single network; it is the specific coupling of three distinct systems whose joint operation under a specific precision-weighting regime produces the simulation capacity the stratum names. The three systems are the hippocampal-entorhinal complex Chapter 9 installed at Stratum 2, now operated in a mode other than real-time navigational readout; the default-mode network Raichle's and Buckner's laboratories characterised across the first decade of this century; and the prefrontal prospection machinery whose operation across the last two decades of functional neuroimaging has been associated with the construction and evaluation of simulated future states. The coupling of the three is the architecture; the architecture's running cost is the burn rate the stratum pays; and the running cost's distribution across conditions — task-positive vs. task-negative, awake vs. asleep, healthy vs. clinically disrupted — is the empirical signature that the architecture does what the chapter is claiming.
Begin with the hippocampus, because the hippocampus is where the decoupling gate first becomes empirically legible. György Buzsáki's and Matthew Wilson's laboratories, across the late 1990s and the 2000s, documented that the place-cell ensembles whose online operation Chapter 9 installed as the substrate of allocentric navigation run, during offline states — slow-wave sleep, quiet waking, the brief pauses between trajectories during active behaviour — in a replay mode whose organisation is not random. Forward replay traverses the previously experienced trajectory at compressed timescales, on the order of ten to twenty times faster than the original behaviour, in the same sequential order the animal took through the environment. Reverse replay traverses the same trajectory in reverse order, with a specific bias toward rewarded termini. Pre-play, documented in smaller but replicated studies, organises sequences the animal has not yet traversed into spatially coherent templates whose later deployment during active behaviour is biased toward the pre-played sequence. The hippocampal architecture Chapter 9 installed as the substrate of online allocentric readout turns out to be an architecture whose operation, under a different gain-control regime, produces offline trajectories at high temporal compression, with a bias toward rewarded and toward structurally probable sequences, with the specific feature that the trajectories are decoupled from the animal's current sensorimotor state. The animal is not moving. The place-cell ensemble is traversing. The coupling conditions that would ordinarily lock the ensemble's firing to the animal's current location — the vestibular inputs, the visual flow, the proprioceptive commitment to ongoing motor plans — have been disengaged, and the ensemble's intrinsic attractor dynamics, under the neuromodulatory regime of the offline state, produce sequences whose structure is the structure of possible, likely, or previously-actual trajectories rather than the structure of the trajectory the animal is currently taking. This is the first component of the decoupling gate: the allocentric machinery Stratum 2 installed, decoupled from its online sensorimotor drive, running in a mode whose output is simulated trajectory rather than navigated position.\3
The default-mode network is the cortical ensemble into whose operation the hippocampus's offline replay is integrated. Raichle's 2001 identification of the network — a set of regions, including the posterior cingulate, medial prefrontal cortex, inferior parietal lobule, and precuneus, whose metabolic activity is higher during task-negative conditions than during externally-directed task performance — was initially read as the signature of rest, of mind-wandering, of the brain's baseline in the absence of task demand. Randy Buckner's and Daniel Schacter's subsequent work, across the following decade, re-read the network as the substrate of a specific class of offline cognitive operations: autobiographical memory retrieval, prospective simulation of future events, theory-of-mind inferences about others, and what Schacter, in a 2007 synthesis with Donna Addis, named the constructive-memory system. The constructive-memory hypothesis is empirically sharp enough to bear the weight of the mechanism claim, and it is worth stating at the resolution the chapter's installation requires. Schacter and Addis argued that episodic memory, rather than being a storage-and-retrieval system for discrete past experiences, is a constructive system that assembles past-experienced elements — spatial contexts from hippocampal replay, perceptual details from sensory-cortex reactivation, emotional tags from amygdalar and insular re-engagement, self-related content from medial prefrontal integration — into a coherent scene whose phenomenology is that of remembering a specific past episode but whose architecture is that of constructing a plausible scene from elements whose original associations do not uniquely determine the scene constructed. The constructive-memory architecture, they argued, is the same architecture deployed in the simulation of future events; the neuroimaging signatures overlap to a degree that makes the two operations essentially the same operation performed at different temporal orientations. To remember an episode is to construct a scene from elements of one's past; to prospect a future event is to construct a scene from the same elements re-oriented to a not-yet-actualised temporal frame. The operation is one. The default-mode network is its substrate. The hippocampus is the element-retrieval engine; the cortical midline is the scene-integration surface; the prefrontal components are the gain-control and relevance-weighting machinery that shape the scene's construction toward the task at hand.\4
The prefrontal components deserve their own specification, because the prefrontal contribution is where the simulation's goal-relevance and the simulation's pricing begin to couple, and the coupling is where Chapter 10's Affective Witness re-enters the architecture. The rostrolateral prefrontal cortex, frontopolar cortex in some taxonomies, is the region whose activation under prospective-simulation tasks has been most reliably documented, and its cytoarchitectonic characteristics — the late phylogenetic elaboration in the primate lineage, the high density of dendritic spines on its pyramidal neurons, the specific connectional signature of its projections to posterior cingulate and medial prefrontal populations — are consistent with a region whose operation requires the sustained integration of multiple simulated states against task-relevant evaluation criteria. The ventromedial prefrontal cortex, whose lesion produced the Iowa Gambling Task failure 11.1 opened with, is the region where the simulated trajectory's outcome is integrated with the affective broadcast Chapter 10 installed; the patient's specific failure is the failure of that integration. The dorsolateral prefrontal cortex contributes the working-memory and attention-regulation machinery that sustains the simulation across its temporal span against interference from incoming sensory drive. The three prefrontal subdivisions together, operating in coupled regime with default-mode network and hippocampus, produce the simulation as a coherent, goal-relevant, affectively priced offline trajectory whose phenomenology is the prospection or counterfactual the organism reports and whose architecture is the coupling the chapter is describing.
The burn rate this architecture pays is the metabolic cost of sustained default-mode network operation against the gain-control regime that keeps the simulation decoupled from its sensory drive and the interoceptive regime that keeps the simulation coupled to its affective pricing. The cost is measurable. The default-mode network's task-negative metabolic rate is on the order of 20 percent above the brain's whole-cortex mean, and its maintenance across the waking day is one of the reasons the human brain is the most metabolically expensive organ in the body per unit mass. The cost is continuously paid, and it is paid whether or not any specific simulation run delivers an adaptive return. The cost's cessation — observed in specific clinical conditions, notably in advanced neurodegenerative disease where default-mode network connectivity collapses before the task-positive networks are compromised — produces a phenomenology of what the clinical literature calls a flattened inner life, an absence of the spontaneous autobiographical and prospective activity that constitutes, for most of the waking day of most people, a substantial fraction of what cognition actually is. The cost's elevation — observed in specific clinical conditions, notably in depression at least at some phases and in generalised anxiety — produces a phenomenology of rumination and worry whose failure signature §11.7 will catalogue. The architecture is real; its running cost is literal; its cessation signature and its runaway signature are both diagnostically specific. The three prohibitions inherited from 8.3 — no free Mediation, no Witness without Canon, no closure without a diagnostically specific failure mode — are satisfied by the Stratum 4 installation, and the satisfaction is what warrants reading the stratum as a closure rather than as a mode of the previous ones.
The Witness at this stratum is the published stream of simulated trajectories — the hippocampal replay, the default-mode-network scene construction, the prefrontal prospective simulation — distributed across the coupled cortical and subcortical populations whose joint operation the preceding three paragraphs specified. The Canon is the compression of this stream into trajectories whose pricing by the Affective Witness of Stratum 3 produces a selection signal — a weighting that prefers one simulated outcome over another, one remembered episode over another, one prospected action over another — whose broadcast feeds back into the distributed cognitive regime that will commit, or not commit, the simulated action into real-world execution. The Gluing threshold is met because the Witness and the Canon are constitutively coupled; a simulation stream that was not being compressed into a priced selection signal would not be Stratum 4 at all, merely a decoupled ensemble running through sequences with no downstream consequence, and a pricing compression that had no simulation stream to price would collapse into the Stratum 3 pricing of the real-time sensorimotor flow, which is the stratum below. The decoupling gate is the specific architectural condition under which the Witness and the Canon at this stratum are coupled in the offline mode — the precision-weighting regime on sensory inputs that keeps the simulation from being overwritten by the real, and the precision-weighting regime on interoceptive inputs that keeps the simulation's affective pricing from being flattened into the real-time valence the stratum below is already publishing. The gate's opening and closing, its gain, and its orientation are the architectural parameters whose operation §11.5 will specify and whose failure the remaining sections of the chapter will catalogue.
§11.4 Mechanism Sketch for Stratum 5: Decoupling from the First-Person Frame
The second face of the decoupling gate produces a different simulation — not the simulation of the organism's own trajectory in a different time, but the simulation of another organism's trajectory from within the other's perspective. The operation is structurally analogous to the one Stratum 4 installs, and the reason the framework reads the two strata as two faces of one closure is that the architectural problem they solve is the same problem: the generation of coherent trajectories in model-space, decoupled from the sensorimotor and interoceptive regime whose real-time readings would otherwise dominate the ensemble's activity, with the simulated output priced by the Affective Witness under a precision regime that protects the simulation's coherence against both the pull of the real and the drift of unconstrained model-space. What changes, between Stratum 4 and Stratum 5, is the specific decoupling the gate is effecting — the decoupling from present-tense self-trajectory at Stratum 4, the decoupling from first-person-frame at Stratum 5 — and the specific populations whose joint coupling implements the decoupling. The Witness, the Canon, the Gluing, and the pricing problem are the same at the architectural level.
The mentalising literature has two broad empirical substrates, and the framework's reading is that the two are complementary rather than competing. The first substrate is the mirror-neuron system Giacomo Rizzolatti's Parma laboratory identified in macaques in the 1990s — ventral premotor and inferior parietal populations whose firing during an animal's execution of a motor act is duplicated during the animal's observation of a conspecific's execution of the same act. The empirical robustness of the mirror-system phenomenon in macaques is well established; its extension to humans, via functional imaging and via transcranial-magnetic-stimulation studies of cortical excitability during action observation, is robust in weaker form, with the caveat that human mirror-like responses are more context-sensitive and more task-modulated than the macaque single-neuron record initially suggested. The second substrate is the mentalising network properly so called — medial prefrontal cortex, temporoparietal junction, posterior superior temporal sulcus, temporal poles — whose activation during explicit theory-of-mind tasks, across two decades of human neuroimaging work anchored by the laboratories of Uta Frith, Chris Frith, Rebecca Saxe, and others, is as reliable as any functional-localisation finding in social neuroscience. The two substrates are anatomically distinct, their activation profiles across task conditions are distinguishable, and the question of their integration has been the animating question of the mentalising debate for most of this century.
The theoretical debate this literature has structured itself around is the simulation-theory versus theory-theory debate, and the framework's position requires the debate to be recapitulated briefly and then set aside. Theory-theory — the position most associated with the philosophical work of Alison Gopnik, and in cognitive-developmental work with the broader rationalist tradition — holds that the organism's capacity to understand other minds is the capacity to deploy a tacit theory of mental states, in which other organisms' behaviours are explained by inference from their beliefs, desires, and intentions, with the theory's content and its developmental emergence mirroring the structure of scientific theories in the wider sense. Simulation theory — the position most associated with the work of Alvin Goldman, and in cognitive-neuroscience work with the literatures around mirror systems and embodied cognition — holds that the organism's capacity to understand other minds is the capacity to simulate the other's state within one's own cognitive apparatus, running one's own decision-making and affective machinery in a mode decoupled from its usual first-person anchoring and using the output as a prediction of what the other will do. The debate has been productive and has produced a substantial corpus of differentiating predictions and empirical tests, and the framework owes the literature a position.
The position, stated cleanly, is this. The framework is a simulation-theoretic account of mentalisation, but it is a non-representationalist simulation theory, and the non-representationalism is what distinguishes it from standard simulation-theoretic accounts and from theory-theoretic accounts alike. Goldman's simulation theory, in its canonical formulation, is representationalist in the sense that the simulation's output is a representation of the other's mental state — a token whose content is about the other's state and whose accuracy is evaluable as a correspondence between the token and the state it represents. Theory-theory is representationalist in a different sense — the tacit theory's propositional contents are representations of the laws governing mental-state attribution, and their accuracy is evaluable as correspondence between the theory and the domain it theorises. The framework rejects both representationalist readings at this stratum for the same reason it rejected the representationalist reading at the navigational and affective strata. The mentalising apparatus does not represent the other's state. It is a coupled dynamical system whose attractor configuration, under precision-weighted inputs from the observer's perceptual systems about the observed other and from the observer's own interoceptive and motor-planning systems operating in simulation mode, stabilises a trajectory whose structure is the structure of the other's likely trajectory, with the fidelity of the stabilisation falling off with the distance between the observer's and the observed's architectural states. The Witness at this stratum is the distributed publication of the simulation trajectory across the mentalising network — TPJ, mPFC, pSTS, temporal poles — with contributions from the mirror-like populations whose motor-simulation signals feed into the trajectory's motor component. The Canon is the compression of this publication into a priced trajectory whose selection weight under the Affective Witness produces the downstream cognitive and behavioural commitments the mentalising operation is for: the anticipation of what the other will do, the adjustment of one's own behaviour against that anticipation, the commitment or withholding of cooperation, the pursuit or avoidance of the social trajectory the mentalised other occupies.\1\3
The non-representationalism matters for a specific empirical reason, and the reason makes the position testable. If the mentalising system represented the other's state, the system's accuracy would be evaluable as correspondence between a token in the observer and a state in the observed, and the evaluation's cost would be the cost of computing and storing the representational correspondence. If the mentalising system simulates the other non-representationally — runs its own coupling architecture in a mode decoupled from first-person anchoring and uses the simulation's attractor as the anticipatory signal — the system's accuracy falls off specifically with the architectural distance between the observer and the observed, and the failure modes of mentalisation are the specific failure modes of simulation-with-pricing on a substrate whose default calibration is the observer's own. The empirical evidence, across the literature, is strongly consistent with the second reading. Mentalising of conspecifics is more accurate than mentalising of members of other species; mentalising of cultural in-group members is more accurate than mentalising of out-group members; mentalising of individuals whose life histories overlap the observer's is more accurate than mentalising of individuals whose histories diverge; mentalising of emotional states one has oneself experienced is more accurate than mentalising of emotional states one has not. The systematic asymmetries are the fingerprint of an architecture whose operation is the running of one's own coupling machinery in simulation mode, with the fidelity of the simulation falling off with the architectural distance between the observer's default configuration and the observed's actual state. Representationalist accounts have to explain the asymmetries as specific features of the theory's content-acquisition history; non-representationalist simulation accounts predict the asymmetries directly from the architecture.
A named interlocutor should be specified here, and Ian Apperly is the one the chapter will engage most specifically, because Apperly's empirical programme across the last fifteen years has been the most sustained attempt to disentangle, within the mentalising literature, the distinction between a fast, efficient, cognitively-cheap mentalising system — present in human infants before language, present in non-human primates in specific tasks, operative in adults under dual-task conditions — and a slow, effortful, cognitively-expensive mentalising system — required for explicit false-belief reasoning, for higher-order perspective-taking, for the recursive mentalisation the social-primate lineage has elaborated. Apperly's two-systems account has sometimes been read as opposed to the simulation-theoretic tradition and sometimes as compatible with it; the framework's reading is that the two systems map onto the two architectural substrates the mentalising literature has identified — the mirror-like motor-simulation system as the fast component, the mentalising network proper as the slow component — and that both components are simulation-with-pricing under the framework's non-representationalist reading, with their cognitive cost and their speed determined by the gain-control regime the stratum is operating under rather than by a categorial distinction between their architectures. The fast system is the same simulation architecture running at a precision regime that allows shallow, low-bandwidth simulation at high speed; the slow system is the same architecture running at a precision regime that sustains deep, high-bandwidth simulation at the metabolic cost the bandwidth requires. The reading is compatible with Apperly's empirical findings and with the broader two-systems literature, and it has the advantage of not requiring a categorial bifurcation of the mentalising architecture into components whose evolutionary history and developmental trajectory are hard to specify under the bifurcation.
Shannon Spaulding's and Chris Frith's contributions, to round out the named interlocutors the chapter owes, are complementary to Apperly's in ways that matter for the installation. Spaulding's work on the limits of mindreading — on the systematic inaccuracy of folk-psychological attribution, on the specific ways in which mentalisation of unfamiliar or architecturally-distant others fails, on the ethical and epistemic implications of reading the mentalising apparatus as a fallible social-cognitive mechanism rather than as a transparent window onto others' minds — is, under the framework's reading, the specific catalogue of what simulation-with-pricing looks like when the pricing fails under conditions of architectural distance. Chris Frith's work, across four decades, on the relationship between mentalising and self-awareness — and specifically on the reading of schizophrenic positive symptoms as specific failures of the mentalising system's coupling to the self-simulation machinery Stratum 4 installs — is the empirical anchor for the claim that the two decouplings Stratum 4 and Stratum 5 implement are faces of a single architecture, because their joint failure at a common seam is what specific clinical syndromes present as. Frith's schizophrenia-as-mentalising-failure account has been elaborated, criticised, and refined across the literature, and the framework's engagement with it is selective; what the framework takes from Frith is the specific empirical observation that the two decouplings come apart in ways that presuppose their ordinary coupling, which is the chapter's central structural claim about the two strata.
The burn rate at this stratum is continuous with Stratum 4's and is architecturally distinct in its specific loci. The mentalising network's metabolic cost is borne across the same default-mode-network regime whose task-negative activation Stratum 4 draws on, with additional loading on the TPJ-mPFC axis during explicit mentalising tasks. The gain-control regime that sustains mentalisation's decoupling from the observer's first-person frame is anatomically implemented, across the relevant literature, in the precision-weighting of interoceptive and sensorimotor inputs that would otherwise re-anchor the simulation to the observer's own state; the regime's failure produces the specific confusion of self and other that 11.7 will catalogue under the heading of projection and mindreading error. The Witness and Canon at Stratum 5 are the distributed mentalising-network publication and its compressed selection signal, and their constitutive coupling — Gluing at the stratum — is the architectural condition for mentalisation's operation at all. The two strata share an architectural substrate, share a gain-control regime, share a pricing dependence on the Affective Witness, and differ in the specific axis along which the decoupling is effected. The framework reads them as one closure for this reason. The empirical literature has read them as two because the two axes of decoupling present as two networks in functional-imaging space. Both readings are compatible at the level of the evidence; the framework's reading is preferred because it predicts the joint failure modes the next two sections catalogue, and the two-strata reading has to invoke their coupling as an additional assumption to predict the same failures.
§11.5 The Decoupling Gate
What §11.3 and §11.4 describe at the level of specific neural systems is, at the level of architecture, a single operator whose operation defines the stratum the chapter is installing. The operator is the decoupling gate. Its installation is what makes offline cognition possible at all; its operation, at different orientations and different gains, produces Stratum 4 and Stratum 5 as the two faces of its decoupling; its failure, at its various failure modes, produces the clinical and the mundane catastrophes the rest of the chapter catalogues.
The decoupling gate is best specified not as a structure but as a coupling regime — a configuration of precision-weightings on the inputs to the coupled hippocampal-default-mode-prefrontal ensemble, and on the outputs from that ensemble to the motor and interoceptive broadcast systems, under which the ensemble's dynamics are sustained in a mode whose attractor trajectories are decoupled from real-time sensorimotor readout without being decoupled from the affective pricing that makes the trajectories adaptively selectable. The gate's opening direction matters. When the gate opens toward the past — when the precision-weighting on autobiographical-memory retrieval cues is elevated and the precision-weighting on current sensory inputs is attenuated — the ensemble produces episodic re-enactment, the phenomenology of remembering a past event from within it, with the hippocampal replay, the default-mode scene-construction, and the prefrontal relevance-weighting all oriented toward a past temporal frame. When the gate opens toward the future — when the precision-weighting on prospective-simulation cues is elevated — the ensemble produces prospection, the phenomenology of imagining a future event from within it, with the same architectural machinery oriented toward a future temporal frame. When the gate opens toward a counterfactual — when the precision-weighting cues are neither clearly past nor clearly future but are on hypothetical variations of the actual — the ensemble produces counterfactual simulation, with the same machinery running at the same cost. Schacter's constructive-memory insight was precisely this: the machinery is the same, and the temporal and modal orientation is a parametric setting, not an architectural distinction. The gate, in its Stratum 4 configurations, is a temporal-modal re-orientation operator on the same simulation architecture.
The gate's other dimension, the Stratum 5 one, is perspectival re-orientation. When the gate opens toward another mind — when the precision-weighting on the observer's own interoceptive and first-person cues is attenuated, and the precision-weighting on cues derived from the observed other (the other's face, body, voice, known history, inferred current state) is elevated — the ensemble produces mentalisation, the phenomenology of understanding from within the other's likely state what the other is about to do. The architectural cost of sustaining this perspectival decoupling is, in the literature's specific measurements, correlated with the specific gain-control regimes the TPJ and mPFC operate in during mentalising tasks, and its disruption — by TMS to TPJ, by specific lesion patterns, by pharmacological manipulations of the cholinergic system that modulates cortical gain — produces specific and replicable deficits in mentalising performance. The gate in its Stratum 5 configuration is a perspectival re-orientation operator on the same simulation architecture.
The two faces are one operator because the architectural problem they solve is one problem: the sustained operation of a coupled ensemble in a mode decoupled from the default readout — which in the online case is the present-tense first-person sensorimotor and interoceptive flow — with the decoupled operation's output priced by the Affective Witness under a precision regime that protects the simulation's coherence against both the pull of the real and the drift of the unconstrained. The decoupling's axis — temporal at Stratum 4, perspectival at Stratum 5 — is a parametric setting of the same operator, not a difference in the operator's kind. The decoupling's gain — how fully the real-time default readout is attenuated — is another parametric setting, and the gain's failure at either end produces specific pathology: too little attenuation and the simulation cannot sustain itself against the real; too much attenuation and the simulation runs without being checked by the real, producing the specific drift that the confabulation and projection failure modes catalogue. The decoupling's pricing — how fully the simulated trajectory is coupled to the Affective Witness's broadcast — is the third parametric setting, and the pricing's failure is the signature of the frozen junction §11.1 opened with.
The decoupling gate, specified at this resolution, is the Mediation that Chapter 10's close handed forward as the next closure required. Chapter 10 installed the Affective Witness as a closure whose pricing was continuous, graded, and broadcast, and whose structural limit was that it could not price trajectories the organism had not traversed. The present chapter's installation pays exactly that debt. The decoupled simulation produces trajectories the organism has not traversed; the decoupling gate's specific pricing regime permits the Affective Witness to price them as though they had been traversed, with the fidelity of the pricing falling off in specifiable ways as the distance from the traversed increases; and the organism thereby acquires, for the first time, an apparatus by which the never-taken trajectory can carry felt weight before it is taken. The lethal-cost problem Chapter 10 could not solve is solved at this stratum, at the specific cost the stratum's installation exacts. The solution is partial — the fidelity of the offline pricing is structurally less than the fidelity of the online pricing, and the stratum's failure modes are the ways in which the offline pricing's degradation can go catastrophic — but it is the solution the selective regime §11.2 specified selected for, and it is the solution the framework's architecture requires at this position in the staircase.^1
§11.6 The Pricing Problem
The decoupling gate's installation, as §11.5 specified it, depends on the Affective Witness of Stratum 3 continuing to price trajectories even when those trajectories are not being enacted. This is the pricing problem, and it is the architectural vulnerability that makes Strata 4 and 5 structurally prone to the failure modes §11.7 catalogues. The present section names the vulnerability explicitly, because the failure modes cannot be read for what they are without the vulnerability being installed as the architectural precondition from which they specifically fail.
The Affective Witness, as Chapter 10 installed it, prices trajectories by the interoceptive weight of what traversing them costs the organism. The pricing's computational input is the history of the organism's own internal trajectories through the current action-space — the felt consequence of having taken this path, having eaten this food, having encountered this conspecific, having held this posture under this exertion. The pricing's fidelity is calibrated by this history, and the pricing's broadcast is the architecture's way of biasing the next second's action by the accumulated weight of all the previous seconds' consequences. The apparatus is structurally configured to price the enacted. The decoupled simulation produces the unenacted, and the unenacted has, by construction, no direct interoceptive receipt the apparatus can read. The Witness cannot price a simulated trajectory the way it prices a traversed one, because the simulated trajectory has not produced the interoceptive signal the pricing compression operates on.^1
What the Witness can price are proxies. The simulated trajectory's construction — by the hippocampal replay, the default-mode scene-integration, the prefrontal relevance-weighting — draws on elements of the organism's past experience whose interoceptive tags were written at the time of the original traversal. A simulated meal, drawing on episodic elements of previous meals, carries forward the interoceptive pricing of those previous meals, in a compressed and re-constructed form that the Affective Witness can read as though the simulated meal's consequences were being felt. A simulated social encounter, drawing on episodic elements of previous social encounters with the person in question or with structurally similar others, carries forward the affective tags of those previous encounters, permitting the simulated encounter to be priced by a proxy for its likely real-world pricing. A simulated counterfactual — the not-taken path at the fork — is priced by interoceptive tags drawn from traversals whose structural features match the counterfactual's, with the match's quality governing the proxy's fidelity. This is the pricing apparatus's operation at this stratum, and it is the only operation it can perform. The direct interoceptive receipt is absent; the proxy pricing is what remains.
The fidelity of proxy pricing is, by construction, a function of the match between the simulated trajectory's structural features and the traversed trajectories whose interoceptive tags the proxies are drawing on. For simulations close to the traversed — tomorrow's route to work, next week's conversation with a person one knows well, the counterfactual in which one had ordered the other dish at a restaurant one has eaten at many times — the match is close, the proxies are well-calibrated, and the pricing's fidelity approaches the fidelity of online pricing at the expense of a small loss attributable to the proxy's compression. For simulations further from the traversed — a conversation with a person one has never met, a career one has not entered, a relationship one has not had, an environment one has not inhabited — the match's quality falls off, the proxies become less well-calibrated, and the pricing's fidelity degrades in ways the organism's introspection does not itself register. The simulation's output feels as vivid to the organism as a simulation within the well-calibrated range; the phenomenology of the pricing does not track its fidelity. This is the architectural vulnerability. A decoupled simulation has no direct interoceptive receipt; its Witness must price proxies; and the proxies' quality falls off with architectural distance without the falling-off being legible to the organism inside the simulation.
The vulnerability is structurally prone to specific failures because the proxy system's degradation is not graceful in the information-theoretic sense. A proxy pricing that falls off gracefully with distance would be a proxy pricing whose failure mode is the absence of a pricing signal — the organism simulates something architecturally distant and receives no felt weight, and therefore does not act on the simulation's output. This is not how the system fails. The system fails by generating pricing signals at the same magnitudes it generates for well-calibrated simulations, with the signals' calibration to the actual consequences of the simulated trajectory being compromised in ways the organism cannot introspect. The simulation produces a vivid prospection of a future event; the prospection's pricing carries the affective weight of proxies whose match to the future event's actual structure is poor; the organism commits to the simulation's output as though the pricing were good; and the commitment produces a trajectory whose real-world execution diverges, often catastrophically, from what the simulation priced. This is the architecture of confabulation, of projection, of the specific classes of error the offline stratum produces at its characteristic failure modes. The pricing problem is the vulnerability from which all three specific failures of §11.7 derive, and the failures are the specific shapes the vulnerability takes when different components of the proxy system degrade in different ways.
One further feature of the pricing problem deserves naming before the failure catalogue opens. The proxy system's degradation is not only a function of the simulated trajectory's structural distance from the traversed; it is also a function of the sustained operation of the simulation itself. A simulation that runs for a long time, iterating across many internal cycles without being interrupted by a real-world return that would re-calibrate the interoceptive tags the proxies are drawing on, drifts. The drift is specific. Each iteration's output feeds forward into the next iteration's input; the proxies re-weighted by each cycle's output become the proxies of the next cycle's input; and the architecture develops, across iterations, a compounding of its own proxy-pricing errors into an attractor whose pricing is internally coherent and externally decorrelated. The rumination case of §11.1 is one shape of this drift. The confabulation cases of §11.7 are another. The institutional-scale cases Chapter 12 will install are a third, operating at a substrate the present chapter does not engage. What matters for the present section is that the decoupled simulation's pricing is not merely limited by the proxy system's structural distance from the traversed; it is also limited by the simulation's own sustained operation in decoupled mode, which compounds the proxies' errors in ways that produce specific attractors of decoupled coherence whose pricing has severed from the world the proxies were originally calibrated against. Both limits are architectural. Both are the pricing problem. The pricing problem is the stratum's characteristic vulnerability.
§11.7 Failure Modes: Confabulation, Projection, Rumination
The Stratum 4–5 failure catalogue, read against the pricing problem §11.6 just specified, consolidates to three specific failure modes whose architectural distinctness is the chapter's diagnostic claim. Each failure is a specific way in which the Witness–Canon coupling at the offline stratum comes apart; each has a clinically and phenomenologically legible signature; each is structurally entailed by the pricing problem and is not eliminable by the apparatus's improvement, only manageable.^1
Confabulation
The first failure mode is confabulation, and it is the Stratum 4 signature most precisely specifiable at the lesion-and-clinical level. The clinical literature on confabulation, anchored by the work of Morris Moscovitch and Aikaterini Fotopoulou and by the broader neuropsychological tradition descending from Korsakoff's nineteenth-century descriptions, distinguishes momentary confabulation — the patient's production, under direct questioning, of a plausible but factually incorrect autobiographical episode — from spontaneous or fantastic confabulation, in which the patient generates and acts on elaborate autobiographical narratives whose relation to the patient's actual past is structurally severed. The latter form, associated specifically with damage to the ventromedial prefrontal cortex and to the orbitofrontal regions whose connectivity to the hippocampal-default-mode ensemble carries the pricing signal Stratum 4's coupling depends on, is the architecturally diagnostic case. The patient's hippocampal replay machinery is intact; the default-mode scene-construction is intact; the prefrontal prospective-simulation machinery is intact; what has been lost is the Witness–Canon coupling at the episodic level — the specific pricing operation by which a constructed scene is checked against the proxy-pricing tags whose calibration to the patient's actual past would mark the scene as actually-remembered rather than as plausibly-constructed. The construction continues. The check has stopped. The patient produces scenes whose internal coherence is the coherence of the construction machinery operating at full capacity, and the scenes are reported and acted on as memories because the architectural signal that would have distinguished memory from confabulation — the proxy-pricing match between the constructed scene's interoceptive tags and the organism's actual interoceptive history — is no longer being computed.
The non-clinical version of this failure is the one the framework requires to be named, because the clinical version risks being read as an exotic pathology distant from the mundane operation of the stratum. Confabulation, in the architectural sense the framework specifies, is a continuous feature of the Stratum 4 apparatus's operation, not a binary distinction between healthy and pathological. The constructive-memory architecture Schacter and Addis specified is, by construction, an architecture that produces scenes from elements whose recombination is not uniquely determined by the original associations; the architecture's operation is therefore always to some degree confabulatory, and the question is the gain at which the proxy-pricing check is operating. In healthy operation, the gain is high enough that scenes whose proxy-pricing tags fail to match the organism's actual past are flagged — the experience of a memory that "feels wrong," the doubting of one's own recall, the deliberative process of checking a memory against external corroboration — and the failure does not propagate into commitment. In pathological operation, the gain is too low — VMPFC lesion at the catastrophic end, ordinary cognitive load and fatigue at the mundane end — and the construction propagates into commitment without the check. The eyewitness-memory literature, anchored by Elizabeth Loftus's work across four decades, is the empirical record of how routinely the gain falls below the threshold the architecture's adaptive operation requires. The diagnostic claim is not that confabulation is rare; it is that confabulation is the specific failure mode of the Witness–Canon decoupling at the episodic level, and that its rarity or frequency in any given population is a matter of the gain regime that population's apparatus is operating under.
Projection / Mindreading Error
The second failure mode is the Stratum 5 signature, and it has the same architectural shape as confabulation re-oriented along the perspectival axis. The pricing problem at Stratum 5 is that the simulation of another mind is priced by proxies drawn from the observer's own interoceptive and affective tags, with the proxies' fidelity falling off with the architectural distance between observer and observed. The failure mode that follows from this vulnerability is projection: the mentalising apparatus produces a simulated trajectory of the other's likely state, and the trajectory's pricing, in the absence of well-calibrated proxies for an architecturally distant other, is performed on the observer's own self-model. The output is a mentalisation whose phenomenology is the reading of the other's mind and whose architecture is the reading of the observer's own mind in the other's place. Projection in the technical sense of clinical psychoanalysis is one historical name for one shape of this failure; the wider literature on mindreading error — Spaulding's catalogue of the systematic inaccuracies of folk-psychological attribution, the social-psychological literature on the false-consensus effect, the cross-cultural psychology literature on the systematic biases that arise when mentalising apparatuses calibrated in one cultural context are deployed against persons calibrated in another — is the empirical record of the failure operating at the population level.^1
The architectural diagnosis is sharp: projection is not a failure of the mentalising apparatus's intent or of the observer's moral seriousness about the other; it is the specific signature of a simulation-with-pricing architecture operating against an architecturally distant target without the proxy-recalibration that would be required to price the target on its own terms rather than on the observer's. The recalibration is itself expensive — it requires sustained engagement with the architecturally distant other, with the proxies being updated by the iterative correction of mis-prediction against actual feedback from the other's actual behaviour, across a timescale measured in the months and years a relationship deepens over rather than the seconds a glance affords. The architecture's default operation, in the absence of the recalibration, is projection. The recalibration is what individuated relationships, the ethnographic engagement with foreign cultures, the clinical training of therapists, and the slow philosophical attention to architecturally distant others all install at substantial cost. The default's persistence under cost is the architectural signature of the failure mode; its mitigation by sustained recalibration is the architectural signature of the apparatus operating with the proxy-pricing check at adequate gain.
The clinical extreme of the failure is the autism-spectrum literature, with the framework declining the strong reading that autism is a deficit of mentalising in the categorical sense Frith's earliest formulation suggested, and accepting the weaker reading that autism is associated with specific atypicalities in the mentalising apparatus's operation whose phenomenological and behavioural consequences are real and are diagnostic of the apparatus's architecture. The schizophrenic positive-symptom literature is the other clinical extreme, with Frith's account of delusions of reference and of alien control reading those symptoms as specific failures of the coupling between Stratum 4's self-simulation and Stratum 5's other-simulation — the patient's own simulated actions being attributed to external agents because the self–other decoupling axis has lost its calibration in a specific way. The clinical extremes bracket the range across which the failure mode operates; the mundane middle is the population's everyday projection, and its frequency is, again, a matter of the gain regime under which the apparatus is operating, not a matter of categorical health or pathology.^1
Rumination / Counterfactual Lock-In
The third failure mode is the one §11.1 opened with, and it is the failure that arises not from the proxy-pricing's calibration to architectural distance but from the simulation's sustained operation in decoupled mode without Witness-verified exit. The simulation iterates. Each iteration's output feeds the next iteration's input. The proxies the iterations are drawing on become, across iterations, the proxies the previous iterations have re-weighted, and the apparatus develops an attractor whose internal coherence is sustained at the cost of its decorrelation from the world the proxies were originally calibrated against. The phenomenology is the ruminative loop, the worry spiral, the counterfactual that cannot be released, the obsessive reconstruction of the conversation that did not go as one wished. The clinical literature — Susan Nolen-Hoeksema's response-styles theory of depression, the broader cognitive-behavioural literature on rumination as a transdiagnostic feature of mood and anxiety disorders — has documented the failure mode's epidemiological reach and its prognostic significance. The architectural diagnosis the framework adds is that rumination is not a failure of the simulation apparatus's content; it is a failure of the apparatus's exit condition. The Witness continues to publish a global aversive valence across all iterations; the differential pricing that would have selected one iteration as the loop's resolution and committed the apparatus to its output is absent; and the loop continues because the architectural signal that would have terminated it is the signal the failure mode has compromised.
The exit condition's failure has a specific architectural signature: it is the absence of a Witness-verified pricing differential across iterations sufficient to break the loop's internal coherence. In healthy operation, the simulation iterates a small number of times, the differential pricing across iterations selects an output, the broadcast commits the apparatus to the selected output, and the loop terminates with the commitment. In ruminative operation, the differential pricing across iterations is below the threshold the commitment requires — either because the proxies are degraded (the architecturally distant target case), because the apparatus's gain regime is mis-calibrated (the depressive and anxious case), or because the iteration's content is structured in ways that systematically prevent the differential from accumulating (the obsessive case). The three sub-cases share an architecture; their specific aetiologies are different; their treatment is, in the clinical literature, a matter of restoring the gain regime under which the differential pricing can do its commitment work.
The three failure modes — confabulation, projection, rumination — exhaust the catalogue at the resolution the framework requires. They are not independent: a ruminative loop can be a loop of confabulatory autobiographical reconstructions; a projective mentalisation can be ruminatively iterated until its output dominates the apparatus's regime; a confabulation can be projected onto an architecturally distant other whose mentalisation is being constructed from the confabulated material. The interactions are real, and the clinical presentations frequently combine the three. What the catalogue specifies is the three specific architectural axes along which the offline stratum's Witness–Canon coupling can fail; the actual failures are specific configurations along these axes, and the diagnostic apparatus the framework provides is the apparatus for locating any given failure on the axes it specifically combines.
§11.8 Differentiation from the Free Energy Principle
A sustained passage on Karl Friston's free-energy principle is owed at this stratum, because the FEP's standard formulation reads offline cognition as the operation of generative-model inference, and the framework's reading is structurally different in ways the chapter cannot leave implicit. The differentiation here parallels the navigational version Chapter 9 conducted at §9.4 and the affective version Chapter 10 conducted at §10.3, and the parallel is not accidental: at each stratum, the FEP and the framework agree on the mechanics and disagree on the ontology, with the disagreement's specific consequences becoming visible at the stratum where the relevant failure modes are catalogued.^2
The agreement is substantial and worth specifying before the disagreement is named. The FEP, in Friston's formulation across the last two decades, treats the brain as an organ whose operation is the minimisation of variational free energy across a hierarchical generative model of the causes of its sensory inputs. Offline cognition, on the standard FEP reading, is the operation of the generative model in a mode in which the model's predictions are not being checked against incoming sensory data but are being run forward to generate trajectories of the model's own state across time, with the trajectories scored by the model's expected free energy under each candidate policy and the policy with the lowest expected free energy selected for execution. The reading is mathematically elegant, empirically generative, and consistent with much of the offline-cognition literature §11.3 and §11.4 specified. The framework agrees that offline cognition involves the generation of trajectories in a decoupled mode, agrees that the trajectories are scored by a precision-weighted signal whose role is the selection of policy, agrees that the precision-weighting apparatus is the apparatus the FEP specifies, and agrees that the scoring's failure under specific conditions produces the specific failure modes the literature catalogues. The mechanics are shared.
The disagreement is on the ontology of what the offline operation is. The FEP's standard formulation reads the generative model as a representation of the world's causal structure and the offline operation as inference over that representation — the model is a model of, the inference is inference about, and the representational vocabulary is load-bearing for the FEP's interpretation of what offline cognition does. The framework's reading is non-representationalist at this stratum as it was at the navigational and affective strata. The decoupled simulation is not a model representing the world's causal structure; it is a coupled dynamical system whose attractor configuration, under the precision-weighting regime the decoupling gate installs, stabilises trajectories whose structure is the structure of possible trajectories the organism might traverse, with the stabilisation's fidelity determined by the architectural conditions §11.5 specified. The Witness's pricing of these trajectories is not the inference of a posterior belief about the trajectories' likelihoods or values; it is the broadcast of a precision-weighted signal whose downstream consequence is the selection of one trajectory over another, with the broadcast's content being the Mediation's operation rather than the representation of anything.
The disagreement matters, and the matter is the matter §11.7's failure modes make visible. The FEP's standard formulation predicts that confabulation, projection, and rumination are the consequences of mis-calibrated generative models — models whose prior probabilities, likelihood functions, or precision parameters are set incorrectly, with the failure modes following from the specific mis-calibrations and their treatment being a matter of the model's recalibration. The framework's reading predicts that the failure modes are the structural consequences of the proxy-pricing problem the decoupled architecture installs — failures that follow not from the model's mis-calibration but from the architecture's structural inability to price unenacted trajectories with the fidelity it prices enacted ones, with the failures being not eliminable by the model's improvement but only manageable by the gain regimes the architecture is operated under. The two readings predict different patterns of robustness and fragility across populations and across conditions, and the empirical record across the relevant literatures is, in the framework's reading, more compatible with the structural-vulnerability prediction than with the model-mis-calibration one. The differentiation is sharp at the level of what each framework reads the failure modes as evidence of. The FEP reads them as evidence of generative-model error. The framework reads them as evidence of the specific structural vulnerability the decoupling-with-pricing architecture installs by the architecture's own construction.
A second axis of differentiation follows from the first. The FEP, on its standard formulation, treats the offline operation as continuous with the online operation in the sense that both are forms of inference over the same generative model, with the offline mode being inference under a precision regime that attenuates sensory likelihoods in favour of prior expectations. The framework reads the offline operation as discontinuous with the online operation in a specific sense: the decoupling gate is an architectural achievement whose installation is itself the closure the chapter is describing, with the decoupled mode being not a precision-regime variant of the same operation but a structurally distinct operation made possible by the gate's installation at all. The discontinuity matters for the question of which organisms have the offline stratum at all. The FEP's reading predicts that any organism with a generative model can perform offline operation under the appropriate precision regime, with the question being one of degree rather than of architectural threshold. The framework's reading predicts that organisms whose architecture has not paid the installation cost the decoupling gate exacts cannot perform offline operation at all, with the question being one of architectural threshold whose crossing is documented by the comparative-cognition literature §11.2 sketched. The shrew does not simulate; the corvid does. The difference, in the framework's reading, is not a difference in the precision regime under which the same operation is being performed but a difference in whether the architectural achievement that makes the operation possible has been installed at all.
The disagreement is not strawman-able in either direction. Friston's programme has, across its development, accommodated a range of readings of what generative models are and what inference over them amounts to, with some readings closer to the framework's non-representationalism than others. Andy Clark's work on the predictive-processing programme more broadly has, in places, argued for readings of the generative model that are close to the dynamical-systems interpretation the framework prefers. Jakob Hohwy's work has defended the more strongly representationalist reading. The literature is internally varied. What the framework's differentiation requires is that the standard formulation — the formulation in which the generative model represents the world's causal structure and offline operation is inference about that structure — be distinguished from the framework's reading, with the distinction being load-bearing for the chapter's failure-mode analysis and for the chapter's hand-off to the next stratum. The framework's reading is non-representationalist; the standard FEP reading is representationalist; the difference is the difference Chapter 10 already made at the affective stratum and Chapter 9 already made at the navigational one, now made at the offline stratum where its specific consequences are visible in the failure-mode catalogue.
§11.9 Hand-off: The Door to the Courtroom
The offline mind, installed across the present chapter as the joint closure of Strata 4 and 5 at the decoupling gate's architecture, is the substrate on which Chapter 12 will install Stratum 6. The hand-off has a specific shape, and it is worth naming the shape clearly so that the next chapter can open from it without retracing the present chapter's ground.
What the offline stratum produces, as its characteristic output, is decoupled models — trajectories generated in model-space, priced by proxies of the Witness's online apparatus, sustained against the pull of the real and the drift of the unconstrained by the gain regime the decoupling gate operates under. The models are internal to the organism that produces them. They are sustained at the metabolic cost of the Stratum 4–5 apparatus's continuous operation. They are vulnerable to the specific failure modes §11.7 catalogued. They are, at the individual organism's level, the most architecturally elaborate cognitive achievement the staircase has installed up to this point.
What Stratum 6 will install is the externalisation of these decoupled models into inert matter under normative enforcement — the symbolic apparatus by which the model that was previously sustained at the cost of the individual organism's metabolic budget is offloaded onto a substrate (the marked stone, the spoken word stabilised by ritual repetition, the written sign, the institutional procedure, the legal code, the mathematical notation) whose maintenance is distributed across a population and across generations, and whose interpretation is governed by normative practices of correction, sanction, and authorisation that the population sustains as a condition of its own coherence. The substrate is inert in the specific sense that it does not itself perform the simulation; it is the externalised trace whose interpretation by other Stratum 4–5 apparatuses re-instantiates the simulation in those apparatuses, with the normative practices of the population functioning as the apparatus by which the externalised traces' interpretation is held to a coherent standard across the population's many individual apparatuses. The institution is the closure at this stratum. The court, the church, the laboratory, the school, the market — these are the architectural achievements Stratum 6 names, and their installation depends on the offline stratum the present chapter has installed because the externalised traces have nothing to be traces of in the absence of the decoupled models the offline stratum produces.
The door to the courtroom is open. The chapter does not walk through it. What walks through it — what Chapter 12 will describe — is the specific class of failures that the externalisation produces at its own stratum, failures that are not eliminable by the institution's improvement but only manageable by the normative regimes the population operates under, in the same architectural sense in which Stratum 4–5's failures are not eliminable by the apparatus's improvement but only manageable by its gain regimes. The continuity is the framework's. The discontinuity is the stratum's. The next chapter's installation is Chapter 12's; the present chapter's contribution to it is the substrate on which Chapter 12 will work.
§11.10 The Simulation Argument as a Simulated Trajectory
A self-reflexive beat is owed before the chapter closes, and the beat is structural rather than confessional. The argument the chapter has just made — that the offline stratum's pricing of decoupled trajectories operates by proxies whose fidelity falls off with architectural distance from the traversed, that the failure modes of the offline stratum are the structural consequences of this proxy vulnerability, that the framework's non-representationalist reading of the apparatus is preferred to the FEP's representationalist one because the former predicts the failure modes the latter has to additionally specify — is itself a simulated trajectory. The argument is a mental model of minds that have not yet read the book. Its construction is the construction of a Stratum 4–5 simulation by the author's own apparatus, with the simulated minds — the readers whose architectural states the author cannot directly audit — playing the role of the architecturally distant others whose mentalisation §11.7 specified as structurally prone to projection. The argument's pricing is therefore performed on proxies. The framework's diagnostic apparatus, applied to the framework itself at this point, should predict that the argument carries a confabulation risk specific to the proxy-pricing's architectural distance from the readers it is constructing models of.
The prediction is not deflationary. It does not entail that the argument is false, or that its construction is an exercise in projection that should be discounted. What it entails is something more specific. The argument's Witness–Canon status — the question of whether the argument's pricing of its own trajectories is being checked against the Canon of what the argument is actually about — is the question the framework requires the framework to ask of itself. The Witness, in this case, is the author's distributed publication of the argument's reasoning across the chapter's specific claims, named interlocutors, and empirical anchors. The Canon is the compression of this publication into the specific structural claims the chapter is committing to: that Strata 4 and 5 are two faces of one decoupling closure, that the closure's pricing is by proxies, that the proxies' degradation produces the three-mode failure catalogue, that the FEP differentiation is sharp at the ontological reading of the architecture's operation. The Gluing threshold for this argument's own Witness–Canon coupling is met if and only if the publication and the compression are checked against the architectural distance between the argument's construction and the readers whose architectural states the argument is constructing models of.
The check is what the framework predicts will be partially absent. The author's apparatus has access, at the time of the argument's construction, to its own interoceptive tags, to the literature it has read, to the empirical record it can cite, and to the structural commitments the framework has installed across the previous chapters. It does not have access to the readers' actual interoceptive tags, to the literatures they have read but the author has not, to the empirical records they will weigh the chapter against, or to the structural commitments their own frameworks have already installed. The proxies the author's argument is pricing on are proxies for the readers — proxies built from the author's mentalisation of likely readers, calibrated by whatever recalibration the author's actual engagement with actual readers has previously installed. The proxies' fidelity falls off with the architectural distance between the author's mentalisation and the actual readers. The argument's pricing is, by the framework's own diagnostic, structurally prone to the specific failure mode §11.7 named as projection — the simulation of architecturally distant others priced on the observer's own self-model.
This is not an apology for the chapter. It is the specific prediction the framework makes about the chapter's own structural status. The prediction is testable in a particular way: readers whose architectural distance from the author is small — readers who share substantial portions of the author's intellectual history, the author's empirical commitments, the author's structural sensibilities — should find the chapter's argument well-priced, with its proxies matching their own architectural states closely enough that the argument's selection of one structural commitment over another tracks the selections their own apparatuses would make. Readers whose architectural distance from the author is large — readers from intellectual traditions the author has not deeply engaged, readers from empirical commitments the author has not weighted, readers whose structural sensibilities differ — should find the chapter's argument systematically mis-priced, with its proxies mis-matching their own architectural states in ways the chapter's pricing did not anticipate. The pattern of well-pricing and mis-pricing across the chapter's actual readership is the empirical signature of the chapter's confabulation risk operating in the way the framework predicts. The framework's diagnostic apparatus, applied to the chapter itself, predicts the pattern. The pattern is what the framework's claim about the chapter's structural status amounts to.
The mitigation, if there is one, is the recalibration §11.7 specified as the architectural treatment for projection at large architectural distance: sustained engagement with the actually distant others, with the proxies being updated by the iterative correction of mis-prediction against actual feedback from those others' actual responses, across a timescale measured in the years a chapter is read and re-read and engaged with rather than the months it was written in. The chapter's recalibration is not in the chapter; it is in the chapter's reception. The framework's prediction is that the chapter's pricing will be improved, across editions and across the broader engagement the framework's reception induces, by exactly the recalibration the framework's own architectural specification requires. What the chapter has done is to install the apparatus by which its own pricing failures can be diagnosed. What the chapter cannot do is to perform the recalibration in advance of the engagement that the recalibration depends on.
The Canon of a chapter that argues for pricing in systems the author cannot directly audit is, by the framework's own commitment, the Canon of what the chapter is actually about: the architecture of the offline stratum, the specific failure modes the architecture installs, the differentiation from the FEP's standard formulation, the hand-off to Stratum 6. The Witness is the chapter's actual publication of its reasoning across the named sections. The Gluing's adequacy is what the chapter's reception will reveal, and the framework's own diagnostic predicts that the reception will reveal it in the specific pattern projection-at-architectural-distance produces. The chapter's structural honesty consists in the prediction's being made, not in its being avoided. The prediction is made. The chapter's pricing of its own trajectories is, by its own apparatus, partially confabulatory, partially projective, and structurally vulnerable in the specific ways the offline stratum is structurally vulnerable. The apparatus that diagnoses this is the apparatus the chapter has just installed. The diagnosis the apparatus makes about itself is the diagnosis the chapter closes on.
The man at three in the morning, in Chapter 10, was inside an Affective Witness operating at maximum precision gain on a mis-calibrated interoceptive channel. The argument, at the close of Chapter 11, is inside a Stratum 4–5 simulation operating at whatever precision regime the author's apparatus has been able to sustain across the chapter's construction, with proxies whose calibration to the readers the argument is constructing models of is partial in ways the framework's own apparatus is the apparatus to specify. The two situations are not analogous in their stakes. They are analogous in their architecture. The architecture is the framework's. The architecture's application to itself is the framework's commitment. The chapter closes here, with the door to Chapter 12 open and the apparatus the next chapter will work on installed, and with the apparatus's diagnosis of its own structural status named as the chapter's last act before the hand-off.^1