What the Data Actually Says

What the Data Actually Says

Eight experiments (V11.0–V11.7), hundreds of GPU-hours, thousands of evolved patterns. What has this taught us?

Finding 1: The Yerkes-Dodson pattern is universal and robust. Across every substrate condition, channel count, and evolutionary regime, mild stress increases $\intinfo$ by $60$–$200$. This is not an artifact of any particular measurement. It reflects a statistical truth: moderate perturbation prunes weak patterns while the survivors are, by definition, the more integrated ones. Severe stress overwhelms even well-integrated patterns, producing the inverted-U. This pattern is the clearest positive result in the entire experimental line.

Finding 2: Evolution consistently produces fragile integration. In every condition where evolution increases baseline $\intinfo$ (V11.5: $+10.5$, V11.6: higher metabolic fitness), evolved patterns decompose more under severe drought than unselected patterns. This is not a bug in the experiments—it is a real dynamical phenomenon. Evolution on this substrate finds tightly-coupled configurations where all parts depend on all other parts. Tight coupling is high integration by definition. But it is also catastrophic fragility: when any component fails under resource depletion, the failure cascades through the entire structure. This is the difference between a tightly-coupled factory (high integration, catastrophic failure mode) and a loosely-coupled marketplace (low integration, graceful degradation under stress).

Finding 3: Curriculum training is the only intervention that improved generalization. V11.7 is the sole condition where evolved patterns outperform naive on novel stressors across the full severity range ($+1.2$ to $+2.7$ percentage points). Not more channels, not hierarchical coupling, not metabolic cost—graduated, noisy stress exposure. The substrate barely matters compared to the training regime. This has a direct parallel in developmental biology: organisms with rich developmental histories (graduated stressors from embryogenesis through maturation) develop robust integration. Organisms exposed to a single threat level develop anxiety-like maladaptive responses. The CA experiments reproduce this pattern with surprising fidelity.

Finding 4: The locality ceiling. This is the deepest lesson, visible only in retrospect across the full trajectory. Every V11 experiment uses convolutional physics: each cell interacts only with neighbors within kernel radius $R$, weighted by a static kernel. Information propagates at most $R$ cells per timestep. The interaction graph is determined by spatial proximity and does not change with the system’s state.

This means that $\intinfo$ can only arise from chains of local interactions—there is no mechanism for a perturbation at $(x, y)$ to directly affect $(x’, y’)$ unless $|x - x’| < R$. The coupling matrix in V11.4–V11.5 partially addresses this (it couples distant channels), but it is fixed: the “who talks to whom” graph does not change in response to the system’s state. A pattern cannot choose to attend to a distant resource patch. It cannot reorganize its information flow under stress. It cannot forage.

V11.6 makes this concrete. Adding metabolic cost to a substrate with radius-$R$ perception does not produce active self-maintenance. It produces efficient passivity—patterns that waste less, not patterns that seek more. A blind organism with a metabolic cost dies when local resources deplete, regardless of how well-integrated it is, because it has no way to detect resources beyond its perceptual horizon. The autopoietic gap is not about resource dependence. It is about perceptual range and its state-dependent modulation—which is to say, it is about attention.

Finding 5: Attention is necessary but not sufficient. V12 tested the locality ceiling hypothesis directly by replacing convolution with windowed self-attention while keeping all other physics identical. The results create a clean ordering across three conditions:

• Convolution (Condition C): Sustains $40$–$80$ patterns, mean robustness $0.981$. Life without integration.
• Fixed-local attention (Condition A): Cannot sustain patterns at all—$30$+ consecutive extinctions across $3$ seeds. Attention expressivity without evolvable range is worse than convolution.
• Evolvable attention (Condition B): Sustains $30$–$75$ patterns, mean robustness $1.001$. Life with integration at the threshold.

The $+2.0$ percentage point shift from C to B is the largest single-intervention effect in the entire V11–V12 line. But it is a shift to the threshold, not past it. Robustness stabilizes near $1.0$ rather than increasing with further evolution. The system learns where to attend (entropy dropping from $6.22$ to $5.55$) but this refinement saturates. What is missing is not better attention but individual-level adaptation—the capacity for a single pattern to reorganize its own internal dynamics in response to its current state, within its lifetime, rather than waiting for population-level selection to discover robust configurations post hoc. Biological integration under threat is not just a population statistic; it is a capacity of individual organisms.

Connection to the trajectory-selection framework. This is where the experimental results meet the theory developed above. We defined the effective distribution $p_{\text{eff}} = p_0 \cdot \alpha / \int p_0 \cdot \alpha$ and argued that attention ($\alpha$) selects trajectories in chaotic dynamics. The Lenia experiments have now shown what happens in a substrate where $\alpha$ is fixed by architecture: the system’s measurement distribution is determined by the convolution kernel, which never changes. The system cannot modulate its own attention. It has no $\alpha$ to vary.

Biological systems solve this: neural attention (largely implemented through inhibitory gating) dynamically reshapes which signals propagate and which are suppressed. Under moderate stress, attention narrows—the measurement distribution sharpens around threat-relevant features—and this reorganization of information flow preserves core integration while shedding peripheral processing. That is the biological pattern our experiments have been searching for. It requires not just integration (which local physics can produce) but flexible integration (which requires state-dependent, non-local communication).

V12 provides direct evidence for this claim. In the attention substrate, the system’s $\alpha$ is the attention weights, and they evolve: attention entropy decreases from $6.22$ to $5.55$ across 15 cycles as the system learns where to look. The measurement distribution becomes more structured—not through explicit instruction, but through the same evolutionary pressure that failed to produce this effect in every convolutional substrate. The difference is that the substrate now permits modulation of $\alpha$. The modulation is sufficient to reach the integration threshold ($\intinfo$ approximately preserved under stress) but not to clearly cross it ($\intinfo$ does not reliably increase under stress the way it does in biological systems). Attention provides the mechanism; something else—perhaps individual-level plasticity, explicit memory, or autopoietic self-maintenance—provides the drive.

These results crystallize into a hypothesis I will call the attention bottleneck. The biological pattern (integration under threat) cannot emerge in substrates with fixed interaction topology, regardless of the evolutionary regime applied. It requires substrates where the interaction graph is state-dependent—where the system can modulate which signals propagate and which are suppressed in response to its current state. Convolutional physics lacks this; attention-like mechanisms provide it. The relevant variable is not substrate complexity ($C$), not selection pressure severity (metabolic cost), and not training diversity (curriculum)—it is whether the system controls its own measurement distribution.

Status: Partially supported by V12, further advanced by V13. The first clause is confirmed: eight convolutional substrates (V11.0–V11.7) failed to produce integration under stress; fixed-local attention (Condition A) fared even worse. The second clause is partially confirmed: evolvable attention (Condition B) shifts robustness from $0.981$ to $1.001$—the right direction, and the only intervention to cross the $1.0$ threshold. V13 content-based coupling provides additional evidence: robustness peaks at $1.052$ under population bottleneck conditions (see Finding 6).

Finding 6: Content-based coupling enables intermittent biological-pattern integration. V13 replaced V12's learned attention projections with a simpler mechanism: cells modulate their interaction strength based on content similarity. The potential field becomes $\phi_i = \phi_{\text{FFT},i} \cdot (1 + \alpha \cdot S_i)$ where $S_i = \sigma(\beta \cdot (\bar{\text{sim}}_i - \tau))$ is a sigmoid gate on local mean cosine similarity. This is computationally cheaper than attention and provides a minimal test: does content-dependent topology, without learned query-key projections, suffice?

Three seeds, each $30$ cycles ($C{=}16$, $N{=}128$), curriculum stress schedule:

• Mean robustness: $0.923$ across all seeds and cycles
• Peak robustness: $1.052$ (seed 123, cycle 5, population $55$ patterns)
• Phi increase fraction: $30\%$ of patterns show $\intinfo$ increase under stress
• Key pattern: Robustness exceeds $1.0$ only when population drops below $\sim 50$ patterns — bottleneck events select for integration

Two distinct evolutionary strategies emerged across seeds. In one regime (large populations of $\sim 150$–$180$ patterns), the similarity threshold $\tau$ drifted toward zero — evolution discovered that maximal content coupling (gate always-on) works when diversity is high. In another regime (volatile populations oscillating between $13$ and $120$), $\tau$ drifted upward to $0.86$ — selective coupling, where only highly similar cells interact. The selective-coupling regime produced all the robustness-above-$1.0$ episodes.

The deeper lesson is not about content coupling per se. It is about composition under selection pressure. When stress culls a population to a handful of survivors, those survivors are not merely the individually strongest — they are the ones whose content-coupling topology supports coherent reorganization under perturbation. This resonates with a different framing of the problem: what we are watching may be closer to symbiogenesis — the composition of functional subunits into more complex wholes — than to classical Darwinian selection optimizing a fixed design. The content-coupling mechanism makes patterns legible to each other, enabling the kind of functional encounter that drives compositional complexity. Intelligence may not require deep evolutionary history so much as the right conditions for compositional encounter: embodied computation, lethal stakes, and mutual legibility.

The V10 MARL ablation study produced a surprise: all seven conditions show highly significant geometric alignment ($\rho > 0.21$, $p < 0.0001$), and removing forcing functions does not reduce alignment—if anything, it slightly increases it. The predicted hierarchy was wrong: geometric alignment appears to be a baseline property of multi-agent survival systems, not contingent on any specific forcing function. This strengthens the universality claim but challenges the forcing function theory developed in the next section.