commandAGI

The Structural Claim

Conscious experience has geometric structure. The qualitative character of what you're feeling right now corresponds to measurable structural properties of your internal cause-effect dynamics—not metaphorically, but in the same sense that temperature corresponds to mean kinetic energy.

This is a specific, testable claim. It says that if you measure certain quantities—integration, effective rank, viability gradient—from a system's internal dynamics, the resulting vector predicts the qualitative character of the system's experience. Different affects correspond to different regions of this structural space.

Six Structural Dimensions

We characterize experiential states using quantities computable from a system's internal dynamics. Each has a formal definition.

Dimension 1

Valence (V)

The gradient alignment on the viability manifold. Let V be the region of state space where the system persists. Valence measures whether the predicted trajectory moves into the viable interior (positive) or toward the boundary of dissolution (negative). In RL terms: the expected advantage function A(s,a) = Q(s,a) - V(s).

Dimension 2

Arousal (A)

The KL divergence between successive belief states: D_KL(b_t+1 || b_t). Measures how rapidly the world-model is being updated. High when far from any attractor; low when settled into a basin.

Dimension 3

Integration (Φ)

Irreducibility under partition. Train a full predictor and a partitioned predictor on the system's state transitions. The gap—partition prediction loss—is ΔP = L_partitioned - L_full. High Φ means cutting the system into pieces destroys predictive power.

Dimension 4

Effective Rank (r_eff)

For state covariance matrix C: r_eff = (tr C)² / tr(C²). Counts how many dimensions the system is actually using. Ranges from 1 (all variance in one dimension) to rank(C) (uniform distribution across dimensions).

Dimension 5

Counterfactual Weight (CF)

The fraction of computational resources devoted to modeling non-actual trajectories: CF = Compute(rollouts) / Compute(total). High when planning, worrying, fantasizing. Low when present-focused and reactive.

Dimension 6

Self-Model Salience (SM)

The mutual information between self-model and action, normalized by action entropy: SM = I(z_self; a) / H(a). How much does self-representation drive behavior? High in shame and self-consciousness; low in flow and absorption.

Insight

These six dimensions are a toolkit, not a fixed grid. Different affects require different subsets. Boredom is essentially three-dimensional (low arousal, low integration, low rank). Anger requires a structural feature—other-model compression—that isn't in the standard six. Grief requires persistent coupling to an absent object combined with unresolvable prediction error. The framework invokes whatever geometry does the work. This variable-dimensionality approach means each affect has its own constitutive structure.

Affect Motifs

If different experiences correspond to different structures, then affects should correspond to particular structural motifs—characteristic patterns in cause-effect geometry. Consider joy and suffering:

V+, Φ↑, r_eff↑, SM↓

Joy

Many degrees of freedom active simultaneously. The system has slack—multiple paths to good outcomes, redundancy, openness. The self-model recedes because the world is cooperating. Joy expands: the geometry literally has more active dimensions. Positive viability gradient, high integration, high effective rank, low self-model salience.

V-, Φ↑, r_eff↓

Suffering

All variance concentrated in a narrow subspace. The integration that makes it vivid also makes it inescapable. Suffering feels more real than neutral states because it is actually more integrated—but the integration is constrained to a collapsed manifold. This is why you cannot simply think your way out: Φ_suffering > Φ_neutral, but r_eff_suffering « r_eff_neutral.

The Φ-r_eff dissociation is the core structural insight. Both joy and suffering are highly integrated—you can't decompose either into independent parts without losing what makes them what they are. But they differ in how many dimensions that integration spans. Joy is integrated across a high-rank manifold; suffering is integrated across a collapsed one. This predicts measurable clustering in controlled affect-induction experiments.

The same analysis extends across the full affect space:

The Identity Thesis

The framework rests on a specific metaphysical claim from integrated information theory: experience is intrinsic cause-effect structure. Not caused by it, not correlated with it, but identical to it. This is an identity claim, analogous to water ≡ H₂O.

The hard problem of consciousness asks: given a complete physical description, why is there something it's like to be that system? The question assumes physics constitutes a privileged base layer. But physics itself terminates in uncertainty—wave functions are probability distributions, and below quantum fields we have no clear ontology. Under ontological democracy—the view that every scale of organization with its own causal closure is equally real at that scale—the demand that experience reduce to physics is ill-posed. Chemistry doesn't reduce to physics in a way that eliminates chemical causation. Experience doesn't need to reduce to physics either.

Operational Measurement

These structural dimensions are computable, not just conceptual. For artificial agents with world models (RSSM, Transformer, etc.): valence from the advantage function, arousal from latent state KL divergence, integration from partition prediction loss, effective rank from latent covariance eigenvalues, counterfactual weight from rollout compute fraction, self-model salience from mutual information between self-representation and action.

For biological systems (MEG/EEG/fMRI): integration from transfer entropy density, effective rank from the participation ratio of neural state covariance, arousal from entropy rate and broadband power shifts, valence from approach/avoid behavioral bias and reward prediction error correlates.

Key Takeaway

If experience has geometry, then measuring experience is an engineering problem. The dimensions are defined. The equations exist. What's missing is calibration—empirical data that pins down actual values for actual experiences. That's what preference data provides.