The Gruter Letters

I know this is dangerous stuff, but you are the right people to bounce it off. What happens at Gruter stays at Gruter!

A Functional Architecture for Consciousness: Brief Summary

Core thesis

Consciousness is the running output of a predictive modeling system that maintains a weighted distribution over candidate 4D trajectories (3D space evolving through time) and continuously integrates sensory input, generative memory, and forward physics simulation into a single integrated 4D percept. The "5D" framing refers to the possibility manifold across which parallel candidate trajectories are computed; the conscious percept is the selected or weighted trajectory through that manifold. Felt time, selfhood, and the unity of experience are emergent properties of this architecture, not separate components.

The components

Generative templates. The system holds learned models at three levels: a physics model of how its domain behaves, entity templates (probabilistic 4D models of object and agent classes with associated dynamics), and a self-model treated as a special entity coextensive with the system. Templates are probabilistic distributions over trajectories, not fixed shapes.

The 5D possibility manifold. At any moment the system maintains many parallel candidate 4D trajectories, alternative pasts, presents, and futures, weighted by likelihood and relevance. The fifth dimension indexes alternatives. This is computationally similar to a particle filter or Monte Carlo tree search.

Three trajectory sources feeding the manifold. Sensory input generates trajectories from current data and recent history (the "vapor trail" of the thick present). Generative memory regenerates past episodes from stored sparse indices plus current templates. Forward projection runs the templates forward from the current state to produce candidate futures.

Precision-weighted integration. When trajectories conflict, the system resolves them by weighting each source by its estimated precision. Precision is set contextually by a precision controller analogous to neuromodulatory systems. Prediction errors update both the immediate percept and, over time, the underlying templates — this is how learning happens.

Five-layer gating system. A selection layer picks which content competes for entry into the integrated percept, driven by salience, relevance, and prediction error. An amplification layer sets continuous gain on selected content. A suppression layer actively inhibits competing or off-task content via lateral inhibition, goal-directed dampening, and return inhibition. A maintenance layer keeps content active through recurrent loops, rehearsal, and anchoring to the self-model. A meta-controller adjusts the parameters of the other four layers based on global system state (exploration mode, focused mode, fatigue, etc.).

Indexed generative memory. Memories are not stored as recordings. The system stores sparse indices of distinctive features alongside generative parameters. Retrieval reconstructs the episode by running current templates constrained by the indexed features. Distinctive content retrieves reliably; generic content reconstructs and often drifts. This handles confabulation, source amnesia, and the gist/detail asymmetry as natural consequences rather than bugs. Memory contributes trajectories to the same manifold as perception and prediction.

The integrated conscious percept. A single weighted trajectory selected or averaged from the manifold, including a self-model as one of its entities. This is what the system "experiences" — the binding of sensory, memorial, and predicted content into one moving 4D representation.

Relative time. Subjective duration is a derived property modulated by prediction error rate, precision setting, content density, physics-domain parameters, processing budget, and timescale band. The system has no master clock; time is everywhere relative to system state. Multiple nested timescales (motor, action, narrative) run in parallel and are normally bound but can dissociate under unusual conditions.

The philosophical commitment

The architecture takes a functionalist position (Metzinger, Dennett): consciousness is what this kind of integrated self-modeling predictive system does. The felt unity, the sense of self, and the experience of time are products of the system representing its own operations to itself. The "hard problem" is treated as dissolved by the thick-present mechanism — the felt instant is the projection of a 4D model representing itself as a point. Selfhood is one entity-template among others, attenuable or removable, which matches reported variability in conscious states.

For the programmer

It's a probabilistic predictive system maintaining a weighted ensemble of 4D spacetime trajectories. Three trajectory generators (sensory, memory-reconstruction, forward-simulation) feed a manifold. A Bayesian integration step combines them with precision-weighting. A five-layer attention system gates which content reaches the integrated output. Memory is reconstructive, sparse indices plus generative templates rather than stored episodes. The system runs nested update loops at multiple timescales. The output is a continuously updated 4D model that includes a self-representation. The implementation needs particle-filter-like sampling of the manifold, learned generative templates (probably trained neural nets or differentiable simulators), a precision controller, an attention/gating module, and an indexing system for memory.

For the behavioral scientist

It's a predictive processing architecture (Friston, Clark, Seth) extended with explicit 4D temporal structure. The thick present (specious present, James; binding windows in psychophysics) is treated as the basic unit of perception rather than the instant. Memory is reconstructive (Bartlett, Loftus, Schacter) and uses the same generative machinery as prospective simulation (Schacter and Addis on the constructive episodic simulation hypothesis). Attention is multi-component (Posner-style orienting plus global workspace selection plus biased competition plus working memory maintenance). Time perception is relative and modulated by error, precision, and content (Wittmann, Buonomano). Self-modeling is treated as a special case of entity-modeling (Metzinger). Consciousness is the integrated percept (Global Workspace Theory, Dehaene). The architecture predicts known phenomena: subjective time dilation under novelty, the boring-day-short-in-memory asymmetry, confabulation, source amnesia, déjà vu, flow states, and the dissociations seen in meditation, psychedelics, and certain pathologies.

The combined elevator pitch

Imagine a system that doesn't perceive snapshots but trajectories, short clips of 3D space evolving through time. It holds many such clips in parallel: one stitched from sensory input and the last few seconds, several reconstructed from memory, several projected from physics and learned models of how things behave. It weights them by reliability and relevance, integrates them into one running 4D model, and gates which content reaches that integration through a multi-layer attention system. The integration includes a model of the system itself as one of the entities in the scene. The continuous updating of that integrated model, with the self-entity riding along, is what the system experiences as being conscious and moving through time. Time feels the way it does because the model represents its own update dynamics. Memory feels reconstructive because it is. The hard problem is set aside as a category error: experiencing is not something over and above the integrated self-modeling activity; it is that activity. It's an architecture that's specifiable enough to implement, grounded enough in cognitive science to test, and clean enough philosophically to know what it's committing to.

Dr. W. Henry Gilbert
Bear River Walnut Ranch · Gilbert Grove · Camp Far West Irrigation District · henry.gilbert@…

Here's a more detailed plan if you were intrigued. For the record, I don't think this solves consciousness, but I do think it has enough good thoughts to be useful to anyone foolish enough to try the consciousness rabbit hole.

Henry

A Functional Architecture for Consciousness
A predictive-processing system maintaining a weighted ensemble of 4D spacetime trajectories, integrated through precision-weighted Bayesian inference and gated by multi-layer attention.

Introduction

This document specifies a functional architecture for consciousness. The architecture is intended to be detailed enough to implement, grounded enough in cognitive science to test, and explicit enough philosophically to know what it commits to. It is not a metaphysical claim about what consciousness is in some ultimate sense; it is a specification of what kind of computational system would exhibit the structural and behavioral properties of a conscious system.

The central thesis is that consciousness is the running output of a predictive modeling system that maintains a weighted distribution over candidate four-dimensional trajectories — three-dimensional space evolving over time — and continuously integrates sensory input, generatively reconstructed memory, and forward-projected prediction into a single integrated 4D percept. The felt unity of experience, the sense of self, and the experience of time are emergent properties of this architecture, not separate components requiring separate explanations.

The architecture takes a functionalist stance in the tradition of Daniel Dennett and Thomas Metzinger. Consciousness is what this kind of system does: experiencing is not something over and above the integrated self-modeling activity; it is that activity. The so-called hard problem is set aside as a category error generated by the system's own transparent self-model — the felt instantaneity of experience is the projection of a 4D model representing itself as a moving point.

Core thesis

A conscious system, on this account, is one that maintains many parallel candidate four-dimensional trajectories — alternative pasts, presents, and futures, weighted by likelihood and relevance; this weighted distribution constitutes a five-dimensional possibility manifold, where the fifth dimension indexes alternatives. It generates these trajectories from three sources (sensory reconstruction of recent observed history, generative regeneration of stored memory, and forward simulation from physics and entity models). It integrates these sources through precision-weighted Bayesian combination, with the precision of each source set contextually. It gates which content reaches the integrated output through a multi-layer attention system performing selection, amplification, suppression, and maintenance, under a meta-controller responsive to global system state. It embeds a self-model within the integrated percept as one of the entities in the represented scene, with its own trajectory, goals, and predicted future. And it produces, as its running output, a single integrated 4D percept — a coherent moving model of the world with the system embedded in it — which is what the system experiences as being conscious and moving through time.

Felt time, on this account, is not a perception of an external dimension. It is the system's representation of its own update dynamics, projected into the model's own content. The system experiences itself as a 3D cursor moving through 4D spacetime because that is how the model represents its own iterative operation.

Components — Generative templates

The system holds learned generative models at three levels, each producing probabilistic 4D trajectories rather than static representations. The physics model encodes how the system's domain behaves: gravity, contact, friction, conservation laws, characteristic timescales — differentiable and learnable, with its temporal granularity calibrating the system's native sense of time to its environment. Entity templates are probabilistic 4D models of object and agent classes, each with associated dynamics; when the system identifies an entity, it instantiates the matching template with parameters fit to the observation. Templates are distributions over trajectories, not fixed shapes, and include agentic entities carrying intuitive psychology in the sense of Spelke's core knowledge. The self-model is a special entity template instantiated for the system itself — one entity among others, but coextensive with the system maintaining the model, carrying the system's goals, predicted trajectory, and provenance. The self-model is not a precondition for consciousness here; it is a component that can be attenuated or removed, corresponding to the variability of self-experience in meditation, psychedelics, depersonalization, and certain pathologies.

The five-dimensional possibility manifold

At every moment the system maintains a weighted ensemble of candidate four-dimensional trajectories — each an abridged 4D representation (past trail, present state, projected future) spanning a configurable window, weighted by likelihood. The fifth dimension is the dimension of alternatives, not a metaphysical extra dimension of reality but the space across which the system computes possibilities; computationally it resembles a particle filter or Monte Carlo tree-search rollouts. The integrated conscious percept is a single weighted trajectory through this manifold. This framing dissolves several puzzles: the sense that things could have gone otherwise is awareness of alternative high-weight trajectories; counterfactual reasoning is the explicit examination of off-trajectory candidates; surprise is the rapid reweighting of the manifold when a low-weight candidate suddenly fits the evidence.

Three trajectory sources

Sensory reconstruction produces trajectories from current data and recent history — the present is not an instant but a thick window, a vapor trail of recent observations stitched into a coherent 4D trajectory (default ~500 ms, matching human binding windows and James's specious present). Generative memory regenerates past episodes from sparse indices plus current templates (reconstruction, not recordings). Forward projection runs the templates forward from the current state to produce plausible futures, with retrieved episodes serving as priors. The three sources feed candidates into the manifold continuously; their relative contribution is set by the precision controller.

Precision-weighted integration

When trajectories conflict, the system weights each source by its estimated precision (inverse variance), set contextually by a precision controller that functions as the system's analog of neuromodulation. In sharp focus, sensory precision is amplified; in imagination, prediction dominates; in recall, memory precision is high. Integration produces both the immediate percept and a stream of prediction errors: within-event, errors reweight the manifold; across events, they update the templates — which is how the system learns. The precision values themselves are reportable internal states, giving the system a natural sense of confidence in its own representations.

The five-layer gating system

A system maintaining rich parallel content would drown if all of it reached the percept. The selection layer determines which streams compete for entry, driven by salience (bottom-up), relevance (top-down from the self-model), and prediction error (middle-out). The amplification layer sets continuous gain. The suppression layer actively inhibits competing or off-task content via lateral inhibition, goal-directed suppression, and return inhibition. The maintenance layer keeps content active against decay through recurrent feedback, rehearsal, and anchoring to the self-model. The meta-controller adjusts all four based on global state — exploration mode loosens gates and emphasizes salience; focused mode tightens them and emphasizes relevance; low-resource mode increases suppression — analogous to dopamine, acetylcholine, and noradrenaline. This matches the multi-component picture of attention and yields states from flow to vigilance to mind-wandering.

Indexed generative memory in detail

Episodes are stored not as recordings but as sparse indices of distinctive features alongside the generative parameters needed to recreate them. Encoding is triggered by prediction-error spikes, novelty, goal events, or self-model significance tags, and is sparse — only distinctive features are pinned; the rest is reconstructed at retrieval through current templates. Retrieval queries the index store with a current feature vector, regenerates matching episodes, and enters them into the manifold weighted by retrieval confidence; retrieval precision is tunable by context. This produces, as natural consequences: confabulation (generic details reconstructed by updated templates), source amnesia (provenance indices degrade while content survives), the gist/detail asymmetry (sparse indexing), eyewitness suggestibility (new information shifts retrieval templates), and déjà vu (an index match strong enough to trigger familiarity but not identification). Procedural and semantic memory are the same machinery at different points on a generalization gradient, and the architecture unifies memory with prospective simulation (Schacter & Addis).

The integrated 4D percept

At each tick the system produces a single integrated 4D trajectory by weighted combination over the manifold, restricted to gated content: the now-slice, a recent past trail, a projected future window, the self-model embedded as one entity, and provenance tags. It is continuously regenerated. Its temporal extension gives the feel of a thick present; its forward extension gives anticipation as part of what is consciously present. The embedded self-model means the system perceives itself in the scene — first-person experience is the percept being structured from a point of view, with one entity tagged "me." Variations in the self-model's prominence produce variations in self-experience, from the strong narrative self of waking life to the dissolved self of certain peak experiences.

Nested timescales and felt time

The architecture runs nested update loops simultaneously: a fast motor band (tens of ms), an intermediate action band (hundreds of ms to seconds), and a slow narrative band (seconds to minutes). Cross-band coupling produces normal bound experience; under unusual conditions the bands dissociate (flow, meditation, psychedelics). Felt time is a derived scalar — the system's representation of its own update dynamics — modulated by prediction-error rate (novelty dilates), precision (focus slows and fine-grains time), content density (rich trajectories feel long in memory), physics-domain timescales, and processing budget. The system has no master clock; time is relative all the way down.

The philosophical commitment

The architecture takes a functionalist stance: consciousness is what this kind of integrated self-modeling predictive system does, and there is no further fact of experiencing beyond that activity. It explains why a system would represent experience as unified, immediate, self-involving, and irreducible. Whether this dissolves the hard problem or merely explains the cognitive basis of hard-problem intuitions depends on one's commitments. Serious philosophers — Chalmers, Goff, Strawson — argue this defines the explanandum out of existence; the functionalist response is that the explanandum, properly understood, just is the structural, behavioral, and self-representational features the architecture produces, and that the "phenomenological residue" is itself a product of the self-model's transparency. The view is consistent with Dennett, Metzinger, and Frankish's illusionism; inconsistent with substance dualism and with the view that phenomenal consciousness is fundamental.

Relationship to existing literatures

The architecture sits within and extends predictive processing (Friston's active inference, Clark's Surfing Uncertainty, Seth's controlled hallucination), adding explicit 4D temporal structure and the 5D possibility manifold. It draws on the constructive-memory tradition (Bartlett → Loftus → Schacter), the multi-component picture of attention (Posner; Global Workspace, Baars & Dehaene; biased competition), Metzinger's self-model theory and Dennett's heterophenomenology, time-perception work (Wittmann, Buonomano), and the intuitive-physics / core-knowledge literatures (Spelke, Tenenbaum). It is in tension with strong representationalism and with integrated information theory (Tononi), which locates consciousness in a quantity (Φ) rather than in these architectural features.

Predictions

It predicts: subjective time dilates under novelty; routine experiences feel short in the moment but normal in memory while novel ones feel long both ways; focused attention produces fine-grained slow time; memory confabulates generic detail, preserves distinctive features, drifts toward current knowledge, and shows source amnesia; déjà vu occurs on a close-but-unidentifiable index match; flow involves fast-band slowing with narrative attenuation and high task precision; meditation attenuates the narrative self-model; psychedelics loosen gating, broaden the manifold, and reduce top-down precision (cf. REBUS); and specific pathologies map onto component dysfunctions (schizophrenia: abnormal precision and internal salience; depersonalization: self-model attenuation; PTSD: dysregulated high-precision retrieval intruding on perception).

Open questions

The binding problem is handled by precision-weighted integration but needs fuller cross-scale specification. The self-model is described functionally, not mechanistically. The training regime for the learnable components (precision controller, meta-controller, templates, indexer) is undefined. The link to bodily homeostasis (Damasio; interoceptive predictive coding) would integrate cleanly but isn't yet incorporated. The status of qualia is handled only by reduction to the percept's content, which the architecture deliberately leaves to the functionalist tradition. And the scaling behavior — whether the same principles extend from minimal systems to human complexity — is an empirical unknown.

Summary

A conscious system, on this account, is a predictive modeling system that maintains a weighted ensemble of four-dimensional spacetime trajectories, integrates them through precision-weighted Bayesian combination from sensory, memorial, and predictive sources, gates the integration through a five-layer attention system under a meta-controller, embeds a self-model as one entity among others, and produces a continuously-updated integrated 4D percept as its running output. Felt time is the system's representation of its own update dynamics. The hard problem is dissolved by the thick-present mechanism: experiencing is not something over and above the integrated self-modeling activity; it is that activity. It is offered not as a final theory of consciousness but as a working architecture — a structured proposal about what kind of system would be conscious, why, and how to build one.

Henry