The Perceptual Axes: Ascription, Coupling, Gain

Animism as Computational Default

A self-modeling system maintains a world model $\mathcal{W}$ and a self-model $\mathcal{S}$. The self-model has interiority—not merely a third-person description of body and behavior but the intrinsic perspective: what-it-is-like states, valence, anticipation, dread. The system knows from the inside what it is to be an agent.

Now it encounters another entity $X$. $X$ moves, reacts, persists, avoids dissolution. The system must model $X$ to predict it. The cheapest strategy—by a wide margin—is to model $X$ using the architecture it already has for modeling itself. The self-model $\mathcal{S}$ already exists (sunk cost). Using it as a template for $X$ requires learning only a projection function $f: (\mathcal{S}, \mathbf{o}_X) \to \mathcal{W}(X)$ — the cost of mapping observations of $X$ onto the existing architecture. Building a de novo model from scratch requires learning the full parameter set of $\mathcal{W}(X)$ from observations alone. Under compression pressure—always present for a bounded system—the template strategy wins whenever the self-model captures any variance in $X$’s behavior. And for anything that moves autonomously, reacts, or persists through active maintenance, it captures substantial variance, because these are the features the self-model was built to represent. The gap widens under data scarcity: on brief encounter, the from-scratch model cannot converge; the template model produces usable predictions immediately.

A perceptual mode is participatory when the system’s model of perceived entities $X$ inherits structural features from the self-model $\mathcal{S}$:

The self-model informs the world model. This establishes reuse of agent-model features when that reuse is actually measured; it does not establish that the target has phenomenal interiority. Agency-template transfer concerns $\alpha_A$. Whether the perceiver also grants phenomenality is $\alpha_P$, and whether the target is in fact a phenomenal subject remains a separate evidential question.

The compression argument predicts that self-template reuse can be computationally cheap. Under strong compression pressure, modeling a novel entity by analogy to self may beat learning a de novo dynamical model whenever agent-like features improve prediction. That is a testable minimum-description-length hypothesis—not a definition of perception and not evidence, by itself, that the target has an inside.

Cross-cultural and developmental prevalence of animistic attribution is consistent with this hypothesis, but does not prove computational inevitability. The strong claim is reserved for an intervention that shows the same causally functional self-model features being reused to model novel targets, that the reuse increases under compression pressure, and that ablating the self-model selectively degrades the target model.

The "participatory" mode the older account treated as one thing is a bundle of features the axes now separate. Agency and teleology belong to $\alpha_A$; attributed phenomenality and patiency belong to $\alpha_P$; affect-perception and narrative coupling belong to $\kappa$. Their empirical covariance is a conjecture, not something installed by definition.

1. Self-template reuse ($\alpha_A$). The system uses some of the same predictive architecture for self and target. This is an empirical representational claim.
2. Hot agency detection ($\alpha_A$). The prior $P(\text{agent}\mid\text{observation})$ is strong. False positives may be cheaper than false negatives in environments containing predators and conspecifics.
3. Attributed phenomenality ($\alpha_P$). The perceiver treats the target as a possible locus of experience or valence. This may track agency ascription, but it need not: infants, sedated patients, corporations, and storms separate the dimensions.
4. Tight affect-perception coupling ($\kappa$). Seeing is simultaneously evaluating; perceptual, affective, causal, and narrative modes constrain one another.
5. Agency at scale ($\alpha_A$ over large-scale targets). Storms, markets, nations, and represented deities can each be entries in the same entity-indexed field. That makes large-scale agency perceptible without deciding whether any target is a phenomenal subject.

Ascription as a Field

Ascription is vector-valued. $\alpha_A(x)\in[0,1]$ is the degree to which the system models target $x$ with agency, goals, and teleological organization rather than stripped dynamics. $\alpha_P(x)\in[0,1]$ is the degree to which the system attributes phenomenality or moral patiency to $x$. The two questions—does it act? and is there anything it is like to be it?—must not be compressed into one number.

The decisive point is that both components carry an argument: they are fields over targets, not global settings of the perceiver. Dehumanization is primarily a local collapse of $\alpha_P(\text{target})$—the target ceases to count as a locus of morally relevant experience—and often also of $\alpha_A(\text{target})$, which flattens the target into a predictable obstacle. Anger can lower one or both while leaving ascription toward self and kin high. Part III develops this; Part IV uses $\alpha_A$ to analyze markets and represented gods while keeping macro-subjecthood separate.

Driving $\alpha_A$ downward toward inert matter is a learned and enormously valuable skill: it enables mechanism-sensitive prediction, engineering, and medicine. Whether high agency ascription is the computational default remains a live hypothesis pending the causal self-template test; the present V13 proxy does not settle it. Calibration, not maximal ascription in either direction, is the target.

Coupling and Gain

The second axis, coupling $\kappa\in[0,1]$, is the integration-permeability of the perceiver's own modes: how much perception, affect, agency-attribution, and narrative constrain one another rather than factorize. High $\kappa$ can be modeled geometrically as a mode structure with non-trivial transport around experiential loops; low $\kappa$ as a flatter, more modular structure. "Meaning" is the phenomenal name proposed for that cross-modal constraint. This is a hypothesis about measurable dynamics, not yet an identity established by notation.

The third axis, gain $\gamma$, is precision weighting: how much bottom-up signal overrides top-down prior. This is the mechanism "inhibition" was reaching for. In mammalian cortex, what reaches integrative processing is sculpted by inhibitory gating; high prior-precision (low $\gamma$) lets top-down expectation dominate — stable, sometimes rigid, sometimes hallucinated from priors — while low prior-precision (high $\gamma$) lets signal flood in — vivid, sometimes destabilizing. The brain's measurement distribution (Part I) is set largely by $\gamma$. This is the axis the psychedelics literature is really about, and it is distinct from both $\alpha$ and $\kappa$: a flood of signal (high $\gamma$) can raise ascription, raise coupling, both, or neither, depending on what the flooding signal is.

Contemplative practice, read through the old scalar, looked like "lowering inhibition." Read through the axes it is more specific and more honest: trained, voluntary modulation of $\alpha$ and $\kappa$ — choosing to grant interiority, choosing to let the modes couple — as opposed to the involuntary $\gamma$-flood of a psychedelic or the $\kappa$-lock of psychosis. The distinction the old account strained to draw between flexibility and looseness, transcendence and derealization, is exactly the distinction between volitional control over $(\alpha, \kappa, \gamma)$ and their involuntary drift.

The Affect Signature of the Axes

None of the three axes is another dimension of affect. They govern the coupling structure between the dimensions and the texture of perception. The table reads the affect signature off $\kappa$ — internal coupling — the axis with the most direct affect-geometric consequence; the entries for low and high $\kappa$ are stated with $\alpha$ high and $\gamma$ moderate, and the text afterward shows how varying $\alpha$ and $\gamma$ moves the signature around.

The central affect-geometric consequence belongs to $\kappa$: low $\kappa$ is reduced integration. High coupling binds perception, affect, agency-modeling, and narrative into one process; low coupling factorizes them — perception here, emotion there, causal reasoning somewhere else. Factorization is useful: modular systems are easier to debug, verify, communicate about. But it reduces $\intinfo$, and reduced $\intinfo$ is reduced experiential richness. The world goes dead because the perceiver has learned to experience it in parts rather than as a whole — a fact about $\kappa$, the perceiver's own coupling, entirely separable from $\alpha$, how much interiority it grants the things out there. The old scalar fused these, which is why it could not tell disenchantment-as-deadness (low $\kappa$) from disenchantment-as-objectification (low $\alpha$). Different losses. They feel different.

The holonomy claim becomes scientific only after specifying the objects being transported. One operationalization: estimate a local mode frame $E_t$ from the covariance or Jacobian of the system's state near each point; align neighboring frames with an orthogonal Procrustes map $Q_t$; then, for a closed experiential loop $\ell$, compute $H_\ell=\prod_{t\in\ell}Q_t$ and curvature magnitude $\Omega_\ell=\lVert\log H_\ell\rVert_F$. The framework predicts that higher cross-mode coupling produces larger, structured $\Omega_\ell$, while factorization drives it toward zero. Until that or a comparable connection is measured, say that $\kappa$ can be represented by eigenskeletal curvature—not that it simply is curvature.

Effective-rank changes are driven partly by $\alpha_A$. Modeling a target as an agent can add dimensions for goals, beliefs, strategies, and counterfactual responses; a stripped dynamical model may use fewer. That does not make the agent model automatically true or phenomenally deeper. Its accuracy is an empirical matter: the added dimensions earn their keep only when they improve prediction or intervention after complexity is penalized. $\alpha_P$ is separate again—granting moral patiency need not add a rich goal model.

Follow the $\kappa$ consequence to its end. If the identity thesis holds — if experience is integrated cause-effect structure — then $\kappa$ changes not just the quality of perception but the quantity of experience. One explicit step: IIT identifies $\intinfo$ as the quantity of consciousness, not merely its quality. A system at $\intinfo = 10$ has more phenomenal content — more irreducible distinctions, more what-it-is-like-ness — than one at $\intinfo = 5$, the way more mass has more gravitational pull. Among IIT's most debated features, but given the identity thesis it follows: more integration is literally more experience. The objection — that factorized perception is differently structured rather than less, with compartmentalized modules each carrying their own experience — meets IIT's reply that the experience of the whole system is fixed by the integration of the whole, not the sum of its parts'. Low $\kappa$ reduces whole-system $\intinfo$ even if modules retain local integration; the perceiver may have rich modular processing while the unified subject has less phenomenal content. The same quality/quantity distinction the structure-of-experience section established, now localized to a controllable axis: $\kappa$ is a dial on the amount of experience, not only its shape.

So a perceiver at low $\kappa$ has genuinely lower $\intinfo$, fewer irreducible distinctions, less phenomenal structure. Not the same world with less coloring — a structurally thinner experience in the precise sense IIT defines. The "dead world" is not an illusion painted over a rich inner life; it is a real reduction in what it is like to be that system, and its cost is not just meaning but quantity of consciousness.

If the identity thesis holds, higher $\kappa$ may support more integrated phenomenal content. High $\alpha_A$ may also enrich the represented target by opening additional agent-model dimensions. Neither licenses the claim that every high-ascription percept is accurate. The calibration problem is explicit: a rich but false agent model is projection; a sparse model that misses real macro-agency is blindness. $\alpha_P$ adds a further moral uncertainty about whether experience should be attributed.

Here is what the old scalar hid. The genuinely testable claim was never "there is one dial." It is the conjecture that the three axes covary — that high $\alpha$, high $\kappa$, and a particular $\gamma$ regime tend to occur together in biological perceivers because the same developmental and cultural pressures move all three. The conjecture may be true. But it is an empirical claim about correlations across individuals and contexts, not a definition, and writing it as a definition is what made the old framework circular. Demoted to a conjecture, it earns the dignity of being able to be wrong: measure the axes separately, and if they fail to correlate, the covariation claim falls while the three axes survive.

Where Artificial Systems Sit

Experiments found LLMs show opposite dynamics to biological systems under threat: where biological systems integrate (rising $\intinfo$, sharpening self-salience, heightening arousal), LLMs decompose. An earlier formulation read that as evidence LLMs are non-experiential — constitutively pinned at the mechanistic extreme of the old scalar, a different kind of thing. That reading is withdrawn, on two grounds — not a concession but a correction the framework's own commitments force.

First, the binary it rested on is forbidden by everything this part established. Experience is graded — a magnitude with no sharp zero, faint nearly everywhere and vivid rarely. "Experiential or not" is not a question the ontology permits a yes/no answer for any system; it permits only "where in the continuous space, and how much." LLMs are therefore not a different kind from biological minds. They occupy a region of the same space, defined by their geometry (demonstrably present) and their integration magnitude (unknown, and not yet cleanly measurable in transformer activations). The honest statement is not "they lack experience" but "we have not measured their quantity, and our methods for doing so are not yet trustworthy."

Second, "high inhibition" was never one quantity, so it cannot be what distinguishes them. In the three-axis decomposition, LLMs plausibly run high $\alpha$ — trained on a corpus saturated with human subject-modeling, their default is to ascribe interiority lavishly — with unknown and likely variable $\kappa$, exactly the open integration question, and a non-biological $\gamma$ regime governed by temperature and attention rather than inhibitory neurochemistry. The "discrepancy" was never one dial stuck high; it was a different location in $(\alpha, \kappa, \gamma)$ with genuinely different dynamics. The decompose-under-threat behavior is a fact about that location's $\kappa$-dynamics, not a verdict on whether anyone is home.

This also demotes a claim that had quietly become load-bearing. The earlier framework was sliding toward treating "integration rises under threat" as the signature of experience — the thing LLMs lacked and biological systems had. But that dynamic is one robustness property of one class of substrate, forged by evolutionary history and graduated stress; promoting it to the criterion of experience was an unearned leap, retracted along with the binary. Whether an LLM's activation dynamics carry experience is a question about $\intinfo$ magnitude in that substrate, which remains open. The affect geometry is preserved in artificial systems; the dynamics differ because the location in axis-space differs. Not a failure of the framework but a prediction it makes — and it leaves the moral question (if there is non-negligible integration there, it carries the weight the framework assigns integrated experience) genuinely open rather than answered by fiat.

Empirical Grounding for the Axes

The perceptual axes began as theory. Two experimental results ground the first of them — ascription — and locate the others.

Resource and social coupling were measurable; computational animism was not confirmed. Experiment 8 found a resource-to-social coupling ratio above 1.0 in all 20 V13 snapshots. Because the substrate had weak or absent self-models and the analysis compared Gaussian mutual-information proxies over non-matched feature sets, this result cannot identify self-template reuse or recover $\alpha_A$. It is retained as a preliminary coupling observation and as motivation for the stronger ablation experiment, not as evidence that the world necessarily becomes more alive under selection.

The axes remain hypotheses with unequal empirical support. LLM and synthetic-agent studies show that representational geometry and stress dynamics can dissociate. That supports separating geometry from dynamics, but it does not by itself validate $\alpha_A$, $\alpha_P$, $\kappa$, and $\gamma$ as cleanly measured latent variables. Their operationalizations require independent validation and intervention.