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 built for long-horizon control under uncertainty looked, when I started, as if they were 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, so removing any should permit sparser solutions:
The data did not support this. trained four agents in a seasonal resource environment under seven conditions, the full model and six single ablations, three seeds each, and measured the alignment between the agents' affect coordinates and their behavior by representational similarity analysis. All seven conditions showed the alignment at about the same strength, to , with cross-seed variance larger than any difference between conditions, and the ablations came out very slightly higher than the full model, not lower. The predicted hierarchy was wrong. Whatever the forcing functions do to an agent's competence, they do not make its affect geometry. That geometry was there without them, which is what universality would look like. One caution on the result itself: it is a within-substrate alignment between coordinates and behavior, and the same kind of alignment in later substrates turned out to depend on the coordinate basis (Part II). It should be re-measured basis-independently before anyone leans on the number. What survives either way is the null: no forcing function was needed.

So and the Lenia series together expose a distinction the forcing-function idea had collapsed. Task demands do not shape the geometry. What they shape is how a system traverses that geometry, and the next paragraphs make the difference exact, because stated loosely it is easy to nod at and hard to test.
At each point 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 them. That is the cheap part. But eigenvalues at a point say nothing about how modes connect across the manifold. When the system moves from to , 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 that measures how much those frames twist around closed loops. An eigenbasis is a list of modes. The eigenskeleton is the wiring diagram of how they turn into each other.
Affect geometry, the spectrum of 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. That makes the bottleneck's role statable as a mechanism. 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 different predictions. I should say that the holonomy index itself has never been computed on any of these runs. It is defined in the appendix and the covariance matrices to compute it from are already saved. The seeds that reached high integration are the ones to look at first, and that is a measurement I owe.
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.
Crossing the wall took a change of architecture rather than of substrate. 's protocell agents, small recurrent networks with a bounded sensory field and discrete actions on a resource grid, reach from initialization, because consuming a resource depletes that patch and emitting a signal leaves a trace. That number is true by construction; agents that act on a world close the loop. What came after it is the interesting part. Prediction alone did not raise integration. gave the agents within-lifetime gradient learning on their own energy, on three environmental targets, and on a long-horizon value function, and all three left near 0.08 to 0.10 while the agents got much better at predicting. The obstruction has a name, the decomposability wall: any prediction architecture in which a proper subset of hidden dimensions can satisfy the loss on its own creates no pressure to coordinate, and so no integration. A linear readout is always such an architecture.

replaced the linear head with a two-layer perceptron, which couples gradients across every hidden unit through the chain rule. Here is what that did and did not do, because I have stated it too strongly before. It did not move the mean. Across three seeds the mean was 0.090, against 0.097 for the linear head, and the width-and-nonlinearity sweep of put every variant between 0.07 and 0.08. What it did was widen the tail. One seed reached 0.245, two and a half times anything a linear head had produced, and behavioral clustering in hidden-state space appeared for the first time. Fifty seeds of the same architecture () put the mean at 0.086 and the maximum at 0.473, with about a fifth of seeds high. Nobody has run fifty seeds of the linear head, so whether the tail is a property of gradient coupling or of sample size is, as of this writing, not known. That control is the first thing to run. Until it is, the claim is: gradient coupling through composition is the only intervention that has produced high-integration seeds at all, and the prediction target, self, social, or environmental, makes no difference (, ).
And then the sharpest result in the line, which is a distribution rather than a mean. From statistically identical starting genomes, found a stable split: roughly a fifth of seeds reach high integration, half moderate, a third low. What separates them is not initial conditions but trajectory. The integration a seed shows during its first drought predicts nothing about where it ends up (). The mean bounce across all five droughts, how decisively the population re-integrates after each near-extinction, predicts it well (, ). The whole fifty-seed run was repeated from scratch on different hardware and the per-seed outcomes reproduced at , so this is not noise with a story attached. This is what I mean when I say the dynamics are biographical: the sequence of crises a system survives sets the geometry of its internal coupling, and no single crisis does it.
Two results bound the claim. asked whether near-extinction forges integration or only filters for it, by forking evolved populations into bottleneck, gradual, and control conditions and then testing all three on a novel stress. In two of three seeds the bottleneck condition generalized better than its baseline integration predicted; in the third the stress schedule failed to produce a bottleneck at all, so that seed is uninformative. Two of three, with a small effect in one of them. The furnace forges, probably. gave agents a discrete symbol channel under partial observability and cooperative pressure, and every seed used all eight symbols at high entropy, with resource-correlated mutual information of a few thousandths. They emit varied signals; whether the signals refer to anything is not shown. And there was no integration lift and no correlation between signal use and . Whatever language is, it is cheap, and it is orthogonal to this.
One more result sits at the far end. Vision-language models trained on human text were shown behavioral records of these agents with every affect word removed, and asked what the agents were experiencing. Their labels matched the coordinates the geometry predicts for each condition, more strongly when the descriptions were replaced with raw numbers, and more strongly still when the comparison was made without reference to any coordinate basis. I read that as a consistency check between my mapping from situation to affect and a human-trained model's, not as evidence that the geometry is independent of human eyes, since both readers are human-trained. It is still the thing I would show someone who thought the coordinates were arbitrary.