What Makes Systems Integrate

Not all self-modeling systems are alike. Some have sparse, modular internal structure; others have dense, irreducible coupling. Systems designed for long-horizon control under uncertainty appear 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 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 raise agent capabilities (richer behavior, higher reward) without raising the geometric alignment of the affect space. Affect geometry appears cheaper than integration—arising from the minimal conditions of survival under uncertainty, not from architectural sophistication. Whether forcing functions increase integration per se (measured by $\Phi$ rather than RSA) remains an open question.

The V10 and V11–V12 experiments together reveal a distinction the original forcing functions hypothesis missed. 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, whatever forcing functions are active. This is what V10 shows. Affect dynamics—how a system traverses that space, whether integration rises or falls 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 the two levels. It predicted forcing functions would shape the geometry. They don't. 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 are biographical. Later experiments (V22, V23) crystallize this as reactivity — associations from present state to action, decomposable by channel — versus understanding — associations from the possibility landscape, non-decomposable because comparing 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 — 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 from a system with no endoskeletal depth to fall back on when the surface fails. Not a bug to be patched but a failure mode intrinsic to exoskeletal architecture, the cognitive equivalent of an arthropod's shell 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 causally distinct values, weighted by probability. A dimension with two meaningful states costs one bit; with ten thousand, thirteen. The cost is not absolute — it is relative to prediction value. A 13-bit dimension is cheap if it distinguishes 10,000 causally distinct environmental states that all matter for survival, expensive if most of those distinctions carry no prediction value. The optimal eigenskeleton at a given compression budget allocates bits to modes by prediction value, not variance — a high-variance mode with low causal consequence is noise; a low-variance mode with high causal consequence is a critical invariant worth full resolution. Evolution selects dimensions 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. Only ~30% of seeds develop non-trivial holonomy. The bottleneck furnace's role becomes precise: repeated near-dissolution forces the system through large loops in state space. If the eigenskeleton is flat, it traverses those loops without mode coupling — fragments under stress, recovers by reassembling independent pieces, no new topology. If curved, traversal couples modes — recovery integrates components that were independent. The furnace does not create eigenvalues. It creates holonomy. More precisely, it 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, whose parameters differ. 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 test whether it has internalized. The ~30% HIGH seeds 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, herbivores attract predators. The holonomy around a seasonal loop is non-trivial — no single mode is intelligible without tracking its coupling to the others. An agent embedded here faces a compression task: extract the environment's eigenskeleton from partial, noisy observations and build an internal one that preserves enough curvature for viability-relevant prediction. This is a definition of intelligence — not the number of modes tracked (capacity), not the speed of update (processing power), but how well the agent's internal mode couplings mirror the world's. More precisely: intelligence is the rate-distortion optimal eigenskeleton for the environment — the topology minimizing representation cost (bits to maintain each mode and coupling) plus prediction error (bits lost by failing to track couplings that exist). A flat internal skeleton in a curved world pays a high prediction-error cost: perpetual surprise when intervening on one variable propagates through others modeled as independent. A curved skeleton in a flat world pays a high representation cost: wasted bits on couplings that carry no prediction value. The intelligent agent matches the world's topology — curved where the world is coupled, flat where it is independent, each mode's cost justified by its prediction value. Not an aspiration. 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, to genetic diversity, to prediction accuracy, to survival through the next drought. The holonomy of this loop is non-trivial — the system that enters a drought is not the system that emerges, and the modes that matter during scarcity differ from those that matter during abundance. The ~30% of seeds that develop high $\Phi$ are those whose internal eigenskeletons develop curvature matching this structure: their representational modes couple through the drought-recovery cycle the way the environmental variables actually couple. The ~40% that stay low keep flat skeletons — modeling energy, position, and social signals on independent channels, missing the curvature. They are less intelligent in a precise eigenskeletal sense: their internal couplings are a poorer embedding of the environmental ones. The furnace selects for eigenskeletal alignment by forcing the system through the environment's dominant loops — and systems whose modes couple through those loops survive more often than those whose modes stay independent.

V13–V18 extended this with six more substrate variants and twelve measurement experiments, sharpening the conclusion. The geometry is confirmed more strongly: affect dimensions develop over evolution (Exp 7), the participatory default is universal and selectable (Exp 8), collective coupling amplifies individual integration (Exp 9). But the dynamics wall was located precisely: at what Part VII calls rung 8 — where counterfactual sensitivity and self-modeling become operational. Substrate engineering (memory channels, attention, signaling, insulation fields) could not cross it. All variants shared the limit $\rho_{\text{sync}} \approx 0.003$. The closest attempt, V18's insulation field, created genuine sensory-motor boundaries — boundary cells received external FFT signals, interior cells only local recurrent dynamics — and produced the highest robustness of any substrate (mean $0.969$, max $1.651$). The surprise: internal gain evolved downward in all three seeds, $1.0$ to $0.60{-}0.72$. Evolution consistently chose thin boundaries with strong external signal over thick insulated cores. The insulation made 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 reveal pre-existing integration capacity or create it. Three conditions diverged after ten shared cycles: severe cyclic droughts at ~90% mortality, mild chronic stress, and a standard control. A novel extreme stress was then applied identically, and the question was whether bottleneck survivors beat control survivors even after controlling for baseline $\Phi$. In two of three seeds, 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 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 selection, purely by architecture: consume a resource and that patch depletes; move and you reach a different patch; emit and a chemical trace persists. World models developed over evolution, reaching $C_{\text{wm}} = 0.10{-}0.15$: hidden states predict future position and energy well above chance. Self-model salience exceeded $1.0$ in 2/3 seeds — agents encoded their own internal states more accurately than the environment — the minimal form of privileged self-knowledge. Affect geometry appeared nascent, consistent with needing resource-scarcity selection to develop fully (and 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 vague metaphor but "action as cause" as a testable architectural requirement. A further experiment (V21) tested whether 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 step must receive signal about its contribution to prediction or survival, not just the sparse binary of "lived or died." This suggests within-lifetime learning, not merely intergenerational selection, is required for the upper rungs — 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 each individually insufficient to create integration. Hidden states show effective rank 5–7 across seeds — moderately rich, not degenerate — yet 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 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 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'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 ($p \approx 0.93$, 10 seeds). What matters is coupling architecture and evolutionary trajectory, not what the system predicts. V22–V24 built their eigenskeleton on the surface — rigid, efficient, decomposable, an exoskeleton. V27 began pushing 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 starts 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 genomes. What separates high from low 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 new 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 cost of maintaining them? Different trajectories push different seeds to different sides. The furnace tips the balance by forcing the system through loops where couplings ARE maximally informative — where the curved skeleton's prediction advantage 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 for referential signaling, and whether that signaling lifts integration. Referential communication emerged in 100% of seeds (10/10) — stable signal→response mappings with receiver behavioral contingency above chance — confirming language-like coordination as 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), 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, maximally "contaminated" by human concepts — were shown behavioral descriptions of protocell agents from V27 and V31, stripped of all affect vocabulary, described only by 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 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. VLMs trained on human experience independently recognize the affect geometry uncontaminated protocells develop from scratch. The geometry is not a human projection; it is structural convergence across radically different substrates.

Integration Measures

Precise measures of integration follow, playing a central role in the phenomenological analysis.

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

The deepest measure is integrated information ($\Phi$). Following IIT, for a system in state $\state$ it 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?