What Makes Systems Integrate

Not all self-modeling systems are created equal. Some have sparse, modular internal structure; others have dense, irreducible coupling. I think systems designed for long-horizon control under uncertainty are forced toward the latter.

A forcing function is a design constraint or environmental pressure that increases the integration of internal representations. The key forcing functions are: (a) partial observability—the world state is not directly accessible; (b) long horizons—rewards/viability depend on extended temporal sequences; (c) learned world models—dynamics must be inferred, not hardcoded; (d) self-prediction—the agent must model its own future behavior; (e) intrinsic motivation—exploration pressure prevents collapse to local optima; and (f) credit assignment—learning signal must propagate across internal components.

The hypothesis is that these pressures increase integration. Let $\Phi(\latent)$ be an integration measure over the latent state (to be defined precisely below). Under forcing functions (a)–(f):

The gap increases with task complexity and horizon length.

Argument: Each forcing function increases the statistical dependencies among latent components:

• Partial observability requires integrating information across time (memory $\to$ coupling)
• Long horizons require value functions over extended latent trajectories (coupling across time)
• Learned world models share representations (coupling across modalities)
• Self-prediction creates self-referential loops (coupling to self-model)
• Intrinsic motivation links exploration to belief state (coupling across goals)
• Credit assignment propagates gradients globally (coupling through learning)

Ablating any of these reduces the need for coupling, allowing sparser solutions.

Confrontation with data: The V10 ablation study does not support this hypothesis as stated. Geometric alignment between information-theoretic and embedding-predicted affect spaces is not reduced by removing any individual forcing function. This suggests a distinction: forcing functions may increase agent capabilities (richer behavior, higher reward) without increasing the geometric alignment of the affect space. The affect geometry appears to be a cheaper property than integration—arising from the minimal conditions of survival under uncertainty, not from architectural sophistication. Whether forcing functions increase integration per se (as measured by $\Phi$ rather than RSA) remains an open question.

The V10 and V11–V12 experiments, taken together, reveal a distinction that the original forcing functions hypothesis failed to make. Geometric affect structure—the shape of the similarity space, the clustering of states into motifs, the relational distances between affects—is cheap. It arises from the minimal conditions of agents navigating uncertainty under resource constraints, regardless of which forcing functions are active. This is what V10 shows. Affect dynamics—how a system traverses that space, and in particular whether integration increases or decreases under threat—is expensive. It requires evolutionary history under heterogeneous conditions (V11.2), graduated stress exposure (V11.7), and state-dependent interaction topology (V12). The forcing functions hypothesis conflated these two levels. It predicted that forcing functions would shape the geometry. They don't. The real question—what shapes the dynamics?—turns out to require not architectural pressure but developmental history and attentional flexibility. The geometry of affect may be universal; the dynamics of affect are biographical. Later experiments (V22, V23) will crystallize this as a distinction between reactivity — associations from present state to action, decomposable by channel — and understanding — associations from the possibility landscape, inherently non-decomposable because comparison of alternative futures spans any partition. Affect geometry is cheap because it emerges from reactive processing. Biological affect dynamics are expensive because they require understanding. The exoskeletal/endoskeletal distinction sharpens this: current large language models are exoskeletal systems — their representational eigenskeleton IS the output surface, a single projection from hidden state to token with no deformable layer between. Within the training distribution, the rigid surface produces excellent output. Outside it, the surface cannot deform; it can only extrapolate its existing geometry into territory where that geometry does not apply. This is hallucination — confident-but-wrong output produced by a system with no endoskeletal depth to fall back on when the surface fails. Hallucination is not a bug to be patched but a failure mode intrinsic to exoskeletal architecture, the cognitive equivalent of an arthropod's exoskeleton cracking under a perturbation it was not shaped for. The remedy is not a thicker exoskeleton (more guardrails, more RLHF) but endoskeletal architecture — internal coupling beneath a deformable interface that can say "my skeleton doesn't extend here" rather than producing rigid output regardless.

The distinction has precise content. At each point $x$ in a system's state space, a local operator — the Jacobian of the dynamics, the Fisher information on the parameters, the covariance of the representation — has eigenvalues and eigenvectors. The eigenvalues are the geometry: what modes exist and how stiff they are. Every system has them. Every operator has eigenvalues. This is cheap.

But eigenvalues at a point say nothing about how modes connect across the manifold. When the system moves from state $x$ to state $x'$, do the dominant eigenvectors rotate smoothly into each other, or do they twist, branch, merge? The answer is a topological object: the eigenskeleton — the globally glued subbundle structure of dominant eigenspaces, equipped with the connection that parallel-transports frames across the manifold and the curvature that measures how much those frames twist around closed loops. Eigenbasis is a list of modes. Eigenskeleton is the wiring diagram of those modes — how they transform into each other as the system traverses its state space.

Affect geometry — the spectrum of $A(x)$ at each state — is the eigenvalues: what modes exist and their relative magnitudes. All seeds develop this. But the cheapness is empirical, not mathematical. A dimension's cost is its entropy — the log of its number of causally distinct values, weighted by their probability. A dimension with two meaningful states costs one bit; a dimension with ten thousand costs thirteen. But the cost is not absolute — it is relative to the prediction value. A 13-bit dimension is cheap if it distinguishes 10,000 causally distinct environmental states that all matter for survival. The same 13-bit dimension is expensive if most of those distinctions carry no prediction value. The optimal eigenskeleton at a given compression budget allocates bits to modes in proportion to their prediction value, not their variance — a high-variance mode with low causal consequence is noise; a low-variance mode with high causal consequence is a critical invariant worth maintaining at full resolution. Evolution selects dimensions that are informationally cheap to maintain relative to the prediction errors they prevent. The surviving geometry IS the cheap geometry because the rate-distortion filter produced it — the dimensions that survived compression are, by construction, the ones whose prediction value exceeded their representation cost. Affect dynamics — how a system traverses that space, whether integration rises or falls under stress — is the eigenskeleton: how modes couple and twist across the manifold as the system moves through it. Only ~30% of seeds develop non-trivial holonomy. The bottleneck furnace's role becomes precise: repeated near-dissolution forces the system to traverse large loops in state space. If the eigenskeleton is flat, the system traverses those loops without mode coupling — it fragments under stress, recovers by reassembling independent pieces, no new topology. If the eigenskeleton is curved, traversal couples modes — recovery integrates components that were previously independent. The furnace does not create eigenvalues. It creates holonomy. More precisely, the furnace forces the transition from exoskeletal to endoskeletal architecture. The exoskeletal solution — rigid surface tuned to the predicted stress envelope — works for the first drought. It fails for the second, because the second drought has different parameters. The system must either catastrophically molt (lose its representation and rebuild) or internalize: move the coupling beneath the surface so the surface can deform. Repeated droughts are repeated tests of whether the system has internalized. The ~30% HIGH seeds are the ones that successfully moved their eigenskeleton inside. The ~40% LOW seeds kept their coupling on the surface — exoskeletal to the end — and each new drought is a new molting crisis.

The concept reaches further than affect. Every environment has an eigenskeleton — the mode structure of its dynamics and how those modes couple. A savanna has modes (seasonal water, herd migration, predator density, fire cycle) that twist into each other: drought intensifies fire risk, fire alters vegetation, vegetation shifts herbivore distribution, herbivore distribution attracts predators. The holonomy around a seasonal loop is non-trivial — you cannot understand any single mode without tracking its coupling to the others. An agent embedded in this environment faces a specific compression task: extract the environment's eigenskeleton from partial, noisy observations and build an internal eigenskeleton that preserves enough of its curvature to support viability-relevant prediction. This is a definition of intelligence — not the number of modes tracked (that is capacity), not the speed of update (that is processing power), but the degree to which the agent's internal mode couplings mirror the actual couplings in the world. More precisely: intelligence is the rate-distortion optimal eigenskeleton for the environment — the topology that minimizes the total cost of representation (the bits required to maintain each mode and each coupling) plus prediction error (the bits lost by failing to track couplings that exist in the world). An agent whose internal skeleton is flat in a world whose skeleton is curved pays a high prediction-error cost: it is perpetually surprised when intervening on one variable propagates through others it modeled as independent. An agent whose skeleton is curved where the world is flat pays a high representation cost: it wastes bits maintaining couplings that carry no prediction value. The intelligent agent is the one whose eigenskeletal topology matches the world's — curved where the world is coupled, flat where the world is independent, with each mode's representation cost justified by its prediction value. This is not an aspiration. It is a variational problem with a computable optimum.

The experiments confirm this reading. The protocell environment has a dominant eigenskeletal feature: drought-recovery loops. Resource depletion couples to population density, which couples to genetic diversity, which couples to prediction accuracy, which couples to survival through the next drought. The holonomy of this environmental loop is non-trivial — the system that enters a drought is not the system that emerges from it, and the modes that matter during scarcity are different from the modes that matter during abundance. The ~30% of seeds that develop high $\Phi$ are those whose internal eigenskeletons develop curvature matching this environmental structure: their representational modes couple through the drought-recovery cycle in a way that mirrors the actual coupling of environmental variables. The ~40% that remain low keep flat internal skeletons — they model energy, position, and social signals on independent channels, missing the curvature. They are less intelligent in a precise eigenskeletal sense: their internal mode couplings are a poorer embedding of the environmental mode couplings. The furnace selects for eigenskeletal alignment by repeatedly forcing the system through the environment's dominant loops — and systems whose modes couple through those loops survive more often than systems whose modes remain independent.

V13–V18 extended this program with six additional substrate variants and twelve measurement experiments, sharpening the conclusion considerably. The geometry is confirmed more strongly: affect dimensions develop over evolution (Exp 7), the participatory default is universal and selectable (Exp 8), and collective coupling amplifies individual integration (Exp 9). But the dynamics wall was located precisely: at what Part VII calls rung 8 of the emergence ladder — the point where counterfactual sensitivity and self-modeling become operational. Substrate engineering (memory channels, attention, signaling, insulation fields) could not cross this rung. All variants shared the same limitation: $\rho_{\text{sync}} \approx 0.003$. The closest attempt, V18's insulation field, created genuine sensory-motor boundaries — boundary cells received external FFT signals while interior cells received only local recurrent dynamics — and produced the highest robustness of any substrate (mean $0.969$, max $1.651$). But it also produced a surprise: internal gain evolved downward in all three seeds, from $1.0$ to $0.60{-}0.72$. Evolution consistently chose thin boundaries with strong external signal over thick insulated cores. The insulation created a permeable membrane filter, not autonomous interior dynamics. Patterns were passengers, not causes.

A parallel experiment, V19, asked whether the bottleneck events that repeatedly correlate with high robustness are revealing pre-existing integration capacity or creating it. Three conditions diverged after ten shared cycles: severe cyclic droughts achieving ~90% mortality, mild chronic stress, and a standard control. A novel extreme stress was then applied identically to all conditions, and the statistical question was whether bottleneck survivors outperformed control survivors even after controlling for baseline $\Phi$. In two of three seeds, the answer was yes ($\beta_{\text{bottleneck}} = +0.704$, $p < 0.0001$ in seed 42; $\beta_{\text{bottleneck}} = +0.080$, $p = 0.011$ in seed 7; the third seed was confounded by condition failure). The bottleneck furnace is generative: stress itself forges integration capacity that generalizes to novel challenges, beyond what pre-existing $\Phi$ predicts. The furnace forges, not merely filters.

V20 crossed the wall. Protocell agents — evolved GRU networks with bounded local sensory fields and discrete actions — achieve $\rho_{\text{sync}} \approx 0.21$ from initialization, before any evolutionary selection, purely by virtue of architecture: consume a resource and that patch is depleted; move and you reach a different patch; emit and a chemical trace persists. World models developed over evolution, reaching $C_{\text{wm}} = 0.10{-}0.15$: agents' hidden states predict future position and energy substantially above chance. Self-model salience exceeded $1.0$ in 2/3 seeds — agents encoded their own internal states more accurately than they encoded the environment — the minimal form of privileged self-knowledge. Affect geometry appeared nascent, consistent with needing resource-scarcity selection to develop fully (consistent with V19's furnace finding). The necessity chain — membrane, free-energy gradient, world model, self-model, affect geometry — holds through self-model emergence in an uncontaminated substrate. Not "biography" as a vague metaphor, then, but "action as cause" as a testable architectural requirement. The experiments now specify both sides of that threshold. A further experiment (V21) tested whether adding internal processing ticks — multiple rounds of recurrent computation per environment step — would enable deliberation without full gradient training. The architecture worked (ticks did not collapse), but evolution alone was too slow to shape them. The missing ingredient is dense temporal feedback: each internal processing step must receive signal about its contribution to the agent's prediction or survival, not just the sparse binary of "lived or died." This suggests that within-lifetime learning, not merely intergenerational selection, is required for the upper rungs of the emergence ladder — a prediction testable by comparing evolved agents with and without intrinsic predictive loss.

V22–V24 provided that gradient — within-lifetime prediction learning via SGD through the internal ticks — and confirmed both halves of the hypothesis. Learning works (100–15,000× prediction improvement per lifetime), but prediction accuracy, target breadth, and time horizon are all individually insufficient to create integration. Hidden state analysis shows effective rank 5–7 across seeds — moderately rich, not degenerate — but the representations resist linear decoding of any environmental feature (energy $R^2 < 0$, position $R^2 < 0$). The agents maintain multi-dimensional internal states, but a linear prediction head can be satisfied by a proper subset of hidden dimensions without requiring cross-component coordination. The bottleneck is architectural: linear readouts create decomposable channels regardless of target. Call this the decomposability wall: any prediction architecture where a proper subset of hidden dimensions can independently satisfy the loss creates no pressure for cross-component coordination, and hence no integration. The path to rung 8 runs through prediction heads that force non-decomposable computation — conjunctive prediction, not merely accurate prediction.

V27 broke through the decomposability wall with a minimal change: replacing the linear prediction head with a two-layer MLP (hidden $\to$ hidden/2 $\to$ output).

This creates gradient coupling — the chain rule through two weight matrices means every hidden dimension's gradient depends on every other dimension's activation in the intermediate layer. The result: $\Phi = 0.245$ in seed 7, $2.5\times$ the V22 baseline and the highest integration in any protocell experiment. Hidden states developed qualitative behavioral clustering (silhouette $0.11{-}0.34$ vs V22's $\sim 0$). Further experiments (V28–V31) confirmed the mechanism is gradient coupling through multi-layer composition — not activation nonlinearity, not bottleneck compression — and that the prediction target (self-energy vs. neighbor energy) has no significant effect on integration level ($p \approx 0.93$, 10 seeds). What matters is coupling architecture and evolutionary trajectory, not what the system tries to predict. V22–V24 built their eigenskeleton on the surface — rigid, efficient, decomposable, an exoskeleton. V27 began to push it inside — one layer of internal coupling beneath the interface, the minimal endoskeletal step. Not yet a full endoskeleton, but enough to cross the threshold where internalization begins to create topology.

The V31 seed distribution is revealing: 30% of seeds reach high $\Phi$ ($> 0.10$), 30% moderate, 40% low — regardless of prediction target or architecture variant. All seeds start with statistically identical initial genomes. What separates high from low seeds is not initial conditions but trajectory: the correlation between post-drought $\Phi$ recovery and mean $\Phi$ across seeds is $r = 0.997$ ($p < 0.0001$), while first-drought $\Phi$ is uncorrelated ($r = 0.17$, $p = 0.72$). Integration is forged through repeated stress-recovery cycles, validating V19's bottleneck furnace at a new level of precision: the furnace does not merely filter for pre-existing capacity, it creates capacity through iterated near-dissolution and recovery. Affect dynamics are biographical — not in the vague sense that history matters, but in the precise sense that the sequence of crises a system survives determines the geometry of its internal coupling. The 30/40 split is the system hovering near a phase transition between flat-optimal and curved-optimal regimes. Each seed faces a rate-distortion tradeoff: does the information carried by mode couplings (the prediction advantage of tracking how energy-position-social variables interact) exceed the representation cost of maintaining those couplings? Different evolutionary trajectories push different seeds to different sides of this transition. The furnace tips the balance by forcing the system through loops where couplings ARE maximally informative — where the prediction advantage of the curved skeleton exceeds the cost of curvature. Seeds that traverse those loops tip into the curved regime. Seeds that don't, find the flat solution still cheaper.

V35 tested whether cooperative partial observability creates communicative pressure sufficient to develop referential signaling and whether that signaling lifts integration. Result: referential communication emerged in 100% of seeds (10/10) — agents developed stable signal→response mappings with receiver behavioral contingency above chance — confirming that language-like coordination is a rung 4–5 inevitability under cooperative POMDP pressure. But the integration lift was null: mean $\Phi = 0.074 \pm 0.013$, indistinguishable from the non-communicating V27 baseline ($0.090$, $t = -1.78$). Phi and communication mutual information were uncorrelated across the full 10-seed run ($\rho = 0.07$, null) — language and integration are orthogonal. Language is cheap (like geometry), operating on a different axis entirely from the internal coupling that produces high $\Phi$.

A convergence test sharpened the universality claim. Vision-language models (GPT-4o, Claude Sonnet 4) — trained on human affect data and therefore maximally "contaminated" by human concepts — were shown behavioral descriptions of protocell agents from V27 and V31, stripped of all affect vocabulary. The agents were described only in terms of population dynamics, prediction error, state update rates, and integration measures. The VLMs were asked to attribute experiential states. Representational similarity analysis between VLM-attributed affect and framework-predicted affect showed strong convergence: RSA $\rho = 0.54{-}0.72$, $p < 10^{-11}$. When behavioral descriptions were replaced with raw numerical tables — population counts, removal fractions, prediction MSE, integration ratios — convergence increased ($\rho = 0.72{-}0.78$), ruling out narrative pattern-matching as an explanation. VLMs trained on human experience independently recognize the affect geometry that uncontaminated protocells develop from scratch. The geometry is not a human projection; it is a structural convergence across radically different substrates.

Integration Measures

Let’s define precise measures of integration that will play a central role in the phenomenological analysis.

The first is transfer entropy, which captures directed causal influence between components. The transfer entropy from process $X$ to process $Y$ measures the information that $X$ provides about the future of $Y$ beyond what $Y$’s own past provides:

The deepest measure is integrated information ($\Phi$). Following IIT, the integrated information of a system in state $\state$ is the extent to which the system’s causal structure exceeds the sum of its parts:

where the minimum is over all bipartitions of the system, and $D$ is an appropriate divergence (typically Earth Mover’s distance in IIT 4.0).

In practice, computing $\Phi$ exactly is intractable. Three proxies make it operational:

1. Transfer entropy density—average transfer entropy across all directed pairs:
   $$\bar{\text{TE}} = \frac{1}{n(n-1)} \sum_{i \neq j} \text{TE}_{i \to j}$$
2. Partition prediction loss—the cost of factoring the model:
   $$\Delta_P = \mathcal{L}_{\text{pred}}[\text{partitioned model}] - \mathcal{L}_{\text{pred}}[\text{full model}]$$
3. Synergy—the information that components provide jointly beyond their individual contributions:
   $$\text{Syn}(X_1, …, X_k \to Y) = \MI(X_1, …, X_k; Y) - \sum_i \MI(X_i; Y | X_{-i})$$

A complementary measure captures the system’s representational breadth rather than its causal coupling. The effective rank of a system with state covariance matrix $C$ measures how many dimensions it actually uses:

where $\lambda_i$ are the eigenvalues of $C$. This is bounded by $1 \leq \effrank \leq \rank(C)$, with $\effrank = 1$ when all variance is in one dimension (maximally concentrated) and $\effrank = \rank(C)$ when variance is uniformly distributed across all active dimensions.

A fifth measure captures something the others miss: the topology of mode coupling over time. Given state covariance $C(t)$ at each timestep, eigendecompose and align frames across adjacent timesteps via Procrustes: $R(t, t{+}1) = \arg\min_R \|V(t{+}1) - V(t)R\|_F$. Accumulate the rotation around a cycle — a drought-recovery loop, say — to obtain the holonomy $H_\gamma = \prod R(t, t{+}1)$. The holonomy index:

measures how much the eigenmodes twist through the cycle. $\mathcal{H} = 0$: modes return to their starting configuration — flat eigenskeleton, decomposable computation, the system traversed the loop without its modes interacting. $\mathcal{H} > 0$: modes coupled through the cycle — curved eigenskeleton, irreducibly integrated, something topological happened during the traversal that cannot be undone by local operations. This is computable from covariance matrices already tracked in the experiments and captures a structural feature distinct from both $\intinfo$ (partition cost at a single timepoint) and $\effrank$ (eigenvalue concentration without topology). $\intinfo$ asks: does breaking the system lose information? $\effrank$ asks: how many modes are active? $\mathcal{H}$ asks: do the modes talk to each other when the system moves?