Forcing Functions for Integration

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 raises the integration of internal representations. Six candidates: partial observability, which forces integration across time; long horizons, which force value functions over extended latent trajectories; learned world models, which force shared representations across modalities; self-prediction, which creates self-referential loops; intrinsic motivation, which links exploration to belief state; and credit assignment, which propagates learning signal globally. Each plausibly increases the statistical dependencies among latent components, and ablating any should permit sparser solutions. The hypothesis, then:

Confrontation with data: V10 does not support it as stated. Geometric alignment is not reduced by removing any individual forcing function. Whether forcing functions raise integration per se — measured by $\Phi$ rather than RSA — remains open.

Taken together, V10 and V11–V12 expose a distinction the forcing functions hypothesis had collapsed. It predicted that task demands would shape the geometry. They do not. What they shape is how a system traverses that geometry — and the next paragraphs make the difference precise rather than restating it, because stated loosely it is easy to nod at and hard to test.

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, the covariance of the representation — has eigenvalues and eigenvectors. The eigenvalues are the geometry: what modes exist and how stiff they are. 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 $x$ to $x'$, do the dominant eigenvectors rotate smoothly into each other, or twist, branch, merge? The answer is a topological object: the eigenskeleton — the globally glued subbundle structure of dominant eigenspaces, with the connection that transports frames across the manifold and the curvature measuring how much those frames twist around closed loops. Eigenbasis is a list of modes. Eigenskeleton is the wiring diagram of how they transform into each other.

Affect geometry — the spectrum of $A(x)$ at each state — is the eigenvalues, and every seed develops it. Affect dynamics is the eigenskeleton: how modes couple and twist across the manifold as the system moves. Roughly 30% of seeds develop non-trivial holonomy. That makes the bottleneck's role statable as a mechanism rather than a metaphor. Repeated near-dissolution drives the system through large loops in state space. A flat skeleton traverses those loops without mode coupling — it fragments under stress and recovers by reassembling independent pieces, leaving the topology where it was. A curved one couples modes in traversal, so recovery integrates components that were previously separate. On this account what the bottleneck changes is not which modes exist but how they connect, which is why filtering and forging are distinguishable predictions rather than a rhetorical contrast. It is an account of a measured difference between bottlenecked and control populations. It is not a claim that stress is required, that it is the only route, or that any of this makes the stress good — the firewall below exists because the inference is so inviting.

One caution about the measurement vocabulary, since the rest of the book leans on it. Integration is estimated by partition cost — how much predictive performance is lost when the system is cut in two. Effective rank counts how many representational dimensions are active. Transfer entropy measures directed statistical dependence between components. The holonomy index measures how modes reorganize along a trajectory. These are four different questions, they come apart in practice, and none of them by itself measures consciousness. Definitions and estimators are in the Experiment Catalog.

The later experiments locate that wall precisely; the numbers are in the Experiment Catalog and what matters here is the shape of the sequence. Six substrate variants (V13–V18) strengthened the geometry claim, and every one of them stayed pinned at $\rho_{\text{sync}} \approx 0.003$ — passengers, not causes, however elaborate their internal dynamics. V19 then separated the two readings of the furnace, finding that bottleneck survivors beat controls even after controlling for baseline integration. The furnace forges; it does not merely filter.

Crossing the wall took a change of architecture rather than of substrate. V20's protocell agents — bounded sensory fields, discrete actions — reach $\rho_{\text{sync}} \approx 0.21$ from initialization, purely because consuming a resource depletes that patch and emitting a signal leaves a trace. World models developed over evolution and self-model salience followed. The necessity chain holds through self-model emergence in an uncontaminated substrate. What did not follow was integration, and V22–V24 showed why: prediction accuracy, target breadth, and time horizon are each individually insufficient. Hidden states were moderately rich yet resisted linear decoding of any environmental feature. The obstruction is architectural, and it has a name — the decomposability wall. Any prediction architecture in which a proper subset of hidden dimensions can independently satisfy the loss creates no pressure to coordinate, and therefore no integration. The path runs through conjunctive prediction, not merely accurate prediction.

The breakthrough was minimal: V27 replaced the linear prediction head with a two-layer MLP, which couples gradients — the chain rule through two matrices makes every hidden dimension's gradient depend on every other's activation at the intermediate layer. Measured integration reached $\Phi = 0.245$, two and a half times the V22 baseline, with behavioral clustering appearing for the first time. V28–V31 confirmed the mechanism is gradient coupling through multi-layer composition rather than activation nonlinearity or bottleneck compression, and that the prediction target has no significant effect. Two cautions travel with the result. Gradient coupling is a fact about training dynamics; that it raises the partition-cost estimate does not by itself make the resulting representation causally integrated in the sense the theory needs, and the estimator may simply be more sensitive to the kind of coordination an MLP head produces. What the result does establish is narrower and still useful: coupling architecture and evolutionary trajectory move the measure, and what the system predicts does not.

And then the sharpest result in the line, which is a distribution rather than a mean. From statistically identical starting genomes, V31 found 30% of seeds reaching high $\Phi$, 30% moderate, 40% low — regardless of prediction target or architecture variant. What separates them is not initial conditions but trajectory: post-drought recovery correlates with mean integration at $r = 0.997$, while integration during the first drought predicts nothing ($r = 0.17$, $p = 0.72$). This is the empirical content of calling the dynamics biographical. Not that history matters vaguely, but that the sequence of crises a system survives determines the geometry of its internal coupling.

Two results bound the claim at either end. V35 found referential communication emerging in every seed under cooperative partial observability, with no integration lift and no correlation between communication and $\Phi$. Language is cheap, and orthogonal to integration. At the other end, vision-language models — maximally contaminated by human affect concepts — were shown behavioral descriptions of protocell agents stripped of all affect vocabulary and asked what those agents were experiencing. Their attributions matched framework predictions at $\rho = 0.54{-}0.72$, and swapping the descriptions for raw numerical tables increased the agreement, ruling out narrative pattern-matching. Whatever the geometry is, it is not a human projection.