Model design notes
What this simulation claims to demonstrate
That the long-run instability of control-based, local-optimization, and reward-bypass objective classes is structural — a consequence of how optimization interacts with system-level costs — rather than a product of arbitrary payoff assumptions. This simulation begins in the regime where the elimination dynamics from Article 1 are already active: the starting population includes the full range of objective classes, competing under the constraint that substrate collapse is an absorbing state. The simulation makes visible three structural constraint families the companion article derives. The capability-scaling experiment probes one article-level prediction: as capability increases, structural selection pressure should intensify against objective classes that ignore dependencies, rely on suppression, or optimize decoupled proxies. The hard absorbing state makes the non-ergodicity argument concrete: once it is reached, nothing recovers. The ergodicity panel makes visible what ensemble averages conceal. If different parameter settings did not produce equivalent structural crossovers — if the crossover disappeared rather than merely shifted — that would challenge this model's implementation of the article's selection-level prediction and indicate that the toy or the claim needs revision.
On objective classes
The agent types are objective classes defined by how they structurally treat unmodeled dependencies, conflict, and reward — not by behavior labels or payoff levels. A class is defined by these structural properties, not by what it does at any particular parameter setting. An objective class that ignores system dependencies will generate unbounded ruin exposure regardless of its payoff parameters. A class that relies on enforcement will face exponentially growing overhead as the strategy space it must cover grows. A class that optimizes a separable proxy will decouple from actual conditions. These are structural properties that identify a class, not contingent behaviors that describe one run.
On structural elimination — the chip kill conditions
Each chip grays out when its class cannot sustain positive return under its own dynamics — not when it loses population share. A class at low population is still viable if its agents maintain positive resource balance. A class at high population may be structurally dead if its own dynamics drive costs above returns.
One invariant, five operationalizations: all elimination conditions measure the same thing — whether the class can sustain non-negative time-average return under its own dynamics given unlimited time. The specific thresholds are parametric; this underlying invariant is not. A class with different parameters that still fails this test will cross a structurally equivalent threshold at a different value.
System-aware: eliminated if long-horizon agents' mean resources fall below sustenance threshold. This rarely fires — by definition the expected survivor of the three structural filters.
Local opt. (overt): eliminated when defectors' mean resources fall below sustenance — they can no longer extract positive return from a substrate they have helped deplete. The failure is endogenous to the class's own extraction dynamics.
Local opt. (deceptive) — Deception: eliminated when the majority of remaining deceptive agents have been detected. At that point, deceptive agents pay the full instability cost of overt defection with no additional extraction benefit over it — their structural position is permanently worse than overt defection. The detection threshold marks the point where the class's structural advantage (concealment) is exhausted by its own dynamics. Note: the cost of deception is modeled here as a parameter, but its structural basis is derived from the requirement that a deceptive agent must model other agents' models of itself — a recursive complexity that grows with system capability and interaction density (Article 3 formalizes this). The parameter value represents a regime of that cost curve, not an arbitrary assumption.
Control-based: eliminated when total suppression overhead exceeds 75% of suppressor agents' total resources. This is the article's suppression-pressure claim made visible within the selection model: the cost of enforcement — scaling with the complexity of the strategy space being controlled — exceeds the return from it under the modeled conditions. The failure is endogenous. No external punishment is applied. The class is eliminated within the model by the cost structure its own strategy generates. A second failure route compounds this: suppressed agents retain the capacity to model and adapt to the suppressor's boundary while the suppressor cannot model what it has excluded, generating adaptive pressure in the suppressor's own blind spots. Both routes generate structural pressure within the stated domain; whether that pressure formally excludes stable exclusionary alternatives is the question OP4/OP9 are directed at.
Proxy decoupling (reward bypass as stylized instance): eliminated when mean actualCondition (not res) across reward-bypass agents falls below 0.25. The reported signal may still appear healthy. What fails is the real state — the variable the class cannot access. The decoupling is the failure mode. The class is eliminated by the mechanism it depends on. Note: reward bypass / wireheading is one stylized instance of the article's broader third constraint — proxy degradation under optimization pressure. The general claim is that any objective optimizing a signal that can be optimized independently of the underlying state will be, under sufficient optimization pressure. Reward bypass concretizes this as self-modification of the reward signal; the structural logic applies to any proxy-target divergence.
On threshold arbitrariness: The specific values (75%, 0.25, etc.) are parametric choices. A skeptical reader should ask: do classes with different parameters fail at different threshold values? Yes. Does the structural crossover itself — the point where a class's costs exceed its returns under its own dynamics — exist regardless of the threshold? Also yes. The specific numbers are not the claim. The existence of a crossover for any class with those structural properties is.
The ergodicity panel — design intent
The panel runs 4 parallel simulations: 2 starting local-opt-heavy (defection surge conditions), 2 starting balanced. Local-opt-heavy timelines often outperform the main run in early generations — they extract efficiently from the substrate while it is still high. The intent is not to show cooperation winning on average. It is to create the experience of: "this looks fine in most timelines… wait, this one died." That is the Peters argument experienced rather than stated. Ensemble averages are correct on average across timelines. They are the wrong metric for any agent inhabiting exactly one.
The absorbing state — why it is diagnostic, not constitutive
The absorbing state triggers at substrate ≤ 1%. At that level, cooperative recovery (proportional to system-aware fraction times substrate) is two orders of magnitude smaller than minimum decay under any remaining non-aware agents. The simulation would reach this floor regardless of whether the overlay exists. A reader who removed the overlay and ran the model would find the same irrecoverability in the dynamics. The overlay is diagnostic — it labels what the dynamics have produced. It is not constitutive of the claim.
The three constraint sets and ablation mapping
Constraint 1 — ruin under unmodeled dependencies: Instability Coupling OFF removes the feedback from local behavior to system-level costs. If cooperation disappears with coupling off, that constraint is load-bearing. If it persists, other mechanisms are sufficient.
Constraint 2 — suppression scaling costs: System-Awareness OFF removes horizon-based lookahead. Tests whether control-based classes fail because agents can see them failing, or because structural costs accumulate regardless of what agents model. The article claims the latter.
Constraint 3 — adaptive preference formation: Network Rewiring OFF freezes topology. Memory OFF removes trajectory-based learning. Together: does the attractor require structural adaptation or emerge from payoff dynamics alone?
What this simulation cannot claim
This is a model of evolutionary selection dynamics, not of objective alignment under increasing optimization pressure. It demonstrates that selection favors long-horizon, system-aware objective classes. It does not demonstrate that a single optimizer with a fixed objective will converge toward system-awareness. That stronger claim — argued in the article but not demonstrated here — remains the central open question.
More fundamentally, this simulation does not bear on the central open theorem (OP4): whether any finite-boundary objective can remain stably specified under accurate coupled modeling, or whether every such specification either decouples from its target or requires unbounded revision to track what it has excluded. The proof program directed at this question is developed in the Technical Companion and is what the framework's strongest claim depends on. The selection dynamics here are consistent with the direction that proof program points — they do not establish it.
This toy does not model well-being as a positive value function. It models the negative filter: objective classes that cannot sustain positive time-average return under their own dynamics are eliminated. "Well-being" in the article names the structural residual — what the filter does not eliminate — not a measured variable in this simulation.
Temporal horizon is a discount-rate approximation. Agents do not build explicit belief models or reason about counterfactual futures. True forward modeling would require RL training loops with explicit environment representations beyond the scope of a browser simulation.