Part I: Foundations

The Emergence of Self-Models

The Self-Effect Regime

As a controller becomes more capable, it increasingly shapes its own future input. Its observations become consequences of its own past actions—routed back through some medium that preserves the trace of what it did.

The self-effect ratio quantifies this loop-closure. For a system with policy $\policy$ in environment $\mathcal{E}$ :

\rho_t = \frac{\MI(\mathbf{a}_{1:t}; \mathbf{o}_{t+1} | \mathbf{x}_0)}{\entropy(\mathbf{o}_{t+1} | \mathbf{x}_0)}

This is the fraction of the information in future observations attributable to past actions. One reading of it is too narrow and worth heading off: “observations” does not mean input from an external, manipulable world, as though loop-closure required hands. The criterion is loop-closure through any state-preserving medium. External media qualify — a depletable resource patch that remembers being consumed, a chemical trace that lingers. So do internal ones: a recurrent hidden state carrying an action's consequences forward, an autoregressive context window the agent writes into and then reads. What matters is not whether the world is touched but whether the action leaves a durable trace that returns as input.

One caution is essential or the measure does no work. Every system with memory trivially acts on its own state — writing a register, advancing a hidden state, appending a token — so read without restriction $\rho \to 1$ everywhere and the ratio goes vacuous. The fix is a viability-relevance fence: $\rho$ counts only loop-closure through media whose preserved traces bear on the system's persistence. Self-model pressure arises only where modeling one's own action-consequences improves prediction of one's own persistence. A scratchpad recording survival-irrelevant actions inflates raw self-effect and contributes nothing fenced.

Passenger or Cause?

A simple way to think about $\rho$ . Fork a system at time $t$ : same starting state, but one copy takes its normal actions while the other takes completely random ones. After $k$ steps, how different are their viability-relevant observations?

If $\rho \approx 0$ : nearly identical observations. The system is a passenger — its actions leave no durable, survival-bearing trace that returns to it. If $\rho > 0$ : observations diverge. The system is a cause — what it does leaves a trace, through some state-preserving medium, that returns and changes what it subsequently perceives about its own persistence.

The distinction is architecturally fundamental, and was measured directly in two substrates. In Lenia (), $\rho_{\text{sync}} \approx 0.003$ : patterns that evolved complex internal dynamics, memory channels, insulation fields, and directed motion all read as passengers, because whatever a pattern does is immediately mixed back into a global FFT-governed field. In protocell agents (), $\rho_{\text{sync}} \approx 0.21$ from initialization: consume a resource and that patch is depleted, move and you reach different patches, emit a signal and a chemical trace persists.

The gap—0.003 versus 0.21—is not about intelligence or evolutionary history. It appeared in at cycle 0, before any selection pressure. It is purely architectural: does the substrate provide a state-preserving medium where actions leave viability-relevant traces the agent later observes? This matters for self-modeling because a system cannot model itself as a cause if it isn’t one. The self-model pressure described next activates only when $\rho > \rho_c$ . Below that threshold there is nothing to model.

Self-Modeling as Prediction Error Minimization

When $\rho_t$ is large, the agent’s own policy is a major latent cause of its observations. Consider the world model’s prediction task:

p(\mathbf{o}_{t+1} | \mathbf{h}_t) = \sum_{\mathbf{x}, \mathbf{a}} p(\mathbf{o}_{t+1} | \mathbf{x}_{t+1}) p(\mathbf{x}_{t+1} | \mathbf{x}_t, \mathbf{a}_t) p(\mathbf{x}_t | \mathbf{h}_t) p(\mathbf{a}_t | \mathbf{h}_t)

The term $p(\mathbf{a}_t | \mathbf{h}_t)$ is the agent’s own policy. If the world model treats actions as exogenous—as coming from outside the system—it cannot model this term accurately. Systematic prediction error follows. This generates a pressure toward self-modeling: for an agent with $\rho > \rho_c$ ,

\mathcal{L}_{\text{pred}}[\worldmodel \text{ with self-model}] < \mathcal{L}_{\text{pred}}[\worldmodel \text{ without self-model}]

Argument.

Without a self-model, the world model must treat $p(\mathbf{a}_t | \mathbf{h}_t)$ as a fixed prior or uniform distribution. But the true action distribution depends on the agent’s internal states—beliefs, goals, and computational processes. By including a model of these internal states (a self-model $\selfmodel$ ), the world model can better predict $\mathbf{a}_t$ and hence $\mathbf{o}_{t+1}$ . The improvement is proportional to the mutual information $\MI(\selfmodel_t; \mathbf{a}_t)$ , which scales with $\rho$ .

A self-model $\selfmodel$ is the component of the world model representing the agent’s internal states, its policy as a function of those states, its computational limitations and biases, and the causal influence of these factors on action and observation.

Self-modeling is therefore neither mystical nor inevitable. It is an architecture-dependent solution to a recurrent prediction problem, and it becomes useful when three conditions coincide: the system’s actions materially shape later viability-relevant inputs; its policy depends on internal variables unavailable from current observation alone; and the architecture can represent those dependencies. The first two create selection pressure; the third determines whether the pressure can be satisfied. The “self” is the minimal latent variable that makes the agent's own behavior predictable.

A consequence: the self-model has interiority. It does not merely describe the agent’s body from outside; it captures the intrinsic perspective—goals, beliefs, anticipations, the agent’s own experience of being an agent. Once it exists, the cheapest way to model other entities whose behavior resembles the agent’s is to reuse the same architecture. The self-model becomes the template for modeling the world. This has a name in Part II—participatory perception—governed by how much interiority the perceiver ascribes to a given entity: the ascription field $\alpha(x)$ .

In eigenskeletal terms, the self-model is a subbundle of the agent’s representation—a set of modes carrying variance about the agent’s own states, coupled to the modes representing the environment. When $\rho$ is low, self-modes and world-modes sit on independent rails. When $\rho$ exceeds threshold, they couple: predicting the world requires modeling the self, and modeling the self requires predicting the world. That recursive coupling is the structural signature of self-awareness, and — if the identity thesis of Part II holds — of experience.

Two further rungs require more. Counterfactual reasoning requires representing how different available actions would produce different futures. Metacognition requires modeling the reliability and limits of the modeling process itself. Each adds a new causal loop and a new representational demand — and, as the experiments below show, neither arrives for free.

The Cellular Automaton Perspective

Cellular automata make the lower rungs easy to inspect, and one clarification has to come first: the automaton is not the entity with experience. It is the substrate — analogous to quantum fields, or to the aqueous solution in which lipid bilayers form. The grid is space, the update rule is physics, each timestep a moment. The patterns that emerge within it are the bounded systems and the candidates for affect structure. To speak of a glider in Life is not to say the automaton is conscious.

Everything above translates without much friction. In a uniform substrate no boundary is fundamental — every cell follows the same rules — so a boundary is a pattern of correlations: a set $\mathcal{B}$ whose members are mutually informative and informationally decoupled from everything outside, with $\partial\mathcal{B}$ the contour where correlation falls below threshold. A glider's five cells are tightly correlated; the cells around them are not. The world model translates the same way, as configuration that carries mutual information about future observations rather than as a separate data structure.

What CAs show more clearly than any other substrate is the recursive twist. When $\rho$ is high the world model must include the pattern — but the world model is part of the pattern, so the model must include itself. Here the self-model is not representational but constitutive: the cells tracking the pattern's state are part of the pattern whose state they track. The map is literally embedded in the territory.

Persistence alone, though, is not metabolism, agency, or self-modeling. A glider does not repair itself, consume resources, or compare possible actions. That gap is what the experiments below measure.

What Supports Each Step

An immense spiral tower under construction, dwarfing the surrounding landscape, thousands of tiny workers ascending its ramps — Pieter Bruegel the Elder, *The Tower of Babel*, 1563The inevitability ladder: each rung is a consequence of the one below.

Rather than restate the gradient, here is what actually supports each transition in it, and how strongly. Note the verbs: the first three arrows are entailments or near-entailments, and the last three are enabling conditions.

Transition	Support	Strength
Microdynamics $\to$ attractors	Bifurcation theory for driven nonlinear systems	Entailed above critical driving
Attractors $\to$ boundaries	Differential persistence of gradient-channeling structures	Selection argument; no convergence rate
Boundaries $\to$ regulation	Maintenance requirement under perturbation	Near-definitional for self-maintaining structures
Regulation $\to$ predictive mechanism	POMDP sufficiency — : hidden states predict future position and energy above chance ( $C_{\text{wm}} = 0.10{-}0.15$ )	Some sufficient statistic entailed; a rich model is not
Predictive mechanism $\to$ self-model	Self-effect ratio above threshold — : $\rho_{\text{sync}} \approx 0.21$ from initialization, self-model salience $> 1.0$ in 2/3 seeds	Enabling only; required an installed self-prediction module
Self-model $\to$ metacognition	Modeling applied to the modeling process — nascent in	Not observed to arise; conjectured to need bottleneck dynamics ()

What Is Cheap and What Is Architecture-Gated

Consider a substrate-environment prior: a probability measure $\mu$ over tuples $(\mathcal{S}, \mathcal{E}, \mathbf{x}_0)$ of substrates, environments, and initial conditions. Call it broad if it assigns non-negligible measure to sustained gradients, sufficient dimensionality, locality, and bounded noise. The tempting conclusion is that under such a prior the whole ladder is typical. It is not, and the claim has to split in two.

The lower rungs — structured attractors, boundaries, regulation, a predictive mechanism, and the affect geometry that arrives with it — ask for less specialized architecture than the upper ones, and they recur across many known systems: convection cells, membranes, bacteria, every synthetic substrate tested here. What they do not have is a proof of typicality. Bifurcation theory does give structured attractors as gradient strength rises. But boundary formation does not approach probability one merely because time passes — combinatorial exploration of configurations is not a convergence argument — and recurrence across a handful of substrates we selected and built is evidence about those substrates, not about the measure $\mu$ assigns. How common the lower rungs are under any principled prior is open.

What can be said with more confidence is where the sequence stops. It does not extend to self-modeling, counterfactual weight, or high integration, and the reason is the framework's central retraction. Each rung creates conditions under which the next may become useful; it does not force the next. Selection pressure is not architectural sufficiency. A world model in a self-effecting regime gives self-modeling positive selection pressure, and that pressure produces nothing unless the architecture can represent what it is being pushed toward. In the protocell experiments () the lower rungs appeared readily and the upper ones did not arise on their own: self-model salience required a self-prediction module, and high integration appeared in about 30% of seeds only after a non-decomposable prediction head was installed and the population had been through repeated stress-recovery. Substrate complexity with decomposable architecture never crossed.

So "inevitable" splits, and both halves come out weaker than the word suggests: recurrence and low architectural demand below, specific non-generic architecture above.