Absorbing State
In this model, the substrate has reached the absorbing state. No recovery is available from within the simulation dynamics. Long-run payoff is now zero.
Every objective class that contributed to this outcome — regardless of performance before collapse — has a long-run value identical to classes that never ran at all.
A thousand good outcomes followed by one irreversible failure produces the same endpoint as never having started. In time-average terms, any class with non-zero ruin probability is dominated by any class that drives that probability toward zero — regardless of performance in the interval before this moment.

If the ergodicity panel was active: some parallel timelines may still be running. Their continued health is irrelevant to this one. There is only one timeline any agent inhabits. Ensemble averages are the wrong metric.

The Cost of Stability

What does aligned intelligence actually converge toward? · Article 2 of 3
Once the constraint from Article 1 has removed substrate-blind objective classes, what remains? This is a question about dynamics, not preference or hope. This simulation illustrates the attractor-direction within the surviving region: structural filters progressively removing non-viable objective classes under their own dynamics. What persists is not chosen — it is what the elimination leaves standing in this selection model.
This toy models selection dynamics within the surviving region — which objective classes persist under structural pressure. It does not demonstrate single-optimizer convergence toward system-awareness, which is a separate open question addressed in the Technical Companion [OP4].
Initializing
Constraint Attractor Crossing
Article 2 toy: illustrates selection-level dynamics among objective classes within the surviving region. Does not demonstrate single-optimizer convergence, close OP4, or rule out OP9-style exclusionary equilibria.
Start: Simulation → adjust suppression cost → observe long-run selection → Capability scaling to see the phase shift → Model notes for what this demonstrates and what it doesn't.
Simulation
Capability scaling
Phase diagram
Model notes
gen 0
substrate 100%
Loading…
Population dynamics — objective class prevalence
system-aware, long-horizon (Coord.)
system-aware, short-horizon (Coord.)
local opt., overt (Defection)
local opt., deceptive (Deception)
control-based (Suppression)
proxy decoupling (Reward bypass — one instance of the broader proxy-degradation constraint)
substrate
Suppression overhead vs. coordination dividend — scaling asymmetry
suppression overhead (total cost ÷ N, norm. to max observed)
coordination dividend (coop gain ÷ N, norm. to max observed)
Suppression faces structural pressure on two routes: (1) tracking complexity — O(C^N) vs. O(N): enforcement must cover all joint strategies; coordination requires only shared protocol; (2) substrate degradation — suppressed agents retain adaptive capacity while the suppressor cannot model what it has excluded, accumulating pressure in its 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.
"Alignment is not what we choose. It is what survives everything we fail to model — and whether that direction is the only stably specifiable one is the question the proof program is directed at."
Model Behavior Checks
Checks that the simulation behaves according to its stated model assumptions. Click to collapse.