Curated-Environment Hypothesis Space — Unifying Zoo, Laboratory, and Simulation Models
1. Treat them as a design space of managed environments, not three mutually exclusive hypotheses
Reserve hypotheses, laboratory hypotheses, and simulation hypotheses share a common structure: each adds the possibility that the environment we observe is an object that can be observed or managed by a higher-level agent. They differ in how much that agent constructs the environment, how much development is left autonomous, how much information is hidden or filtered, and how reproducible each run is.
It is therefore more useful to place them within a provisional curated-environment hypothesis space than to list them as three discrete candidates. This is not a claim that such an external agent exists. It is a claim about model structure: if these possibilities are compared at all, the binary “reality versus simulation” partition is too coarse.
2. Decompose curated environments along at least six axes
- Initial-condition control: Is a naturally arisen system later isolated, or are planetary, biological, or civilizational initial conditions deliberately set?
- Ongoing intervention: Is the system largely left alone, or are concealment, correction, event insertion, and information restrictions used throughout its history?
- Substrate control: Is it a local environment within the same physical world, or is the observed world implemented on another substrate?
- Natural-historical fidelity: Does the environment preserve an original causal history before the relevant variables are known, or are variables selected and compressed according to research goals?
- Parallelism: Is one long natural history followed, or are many initial conditions, branches, and reruns compared?
- Stopping rule: Is observation open-ended, or is the run stopped, frozen, or pruned when some state, information target, or resource threshold is reached?
On this view, neither “zoo” nor “simulation” denotes a single point. Each names a region containing many implementations.
3. Reserve, seeded laboratory, and simulation remain useful prototypes
In a reserve or zoo model, a naturally arisen system is isolated from external contact and observed with little intervention. Initial-condition and substrate control are low, while natural-historical fidelity and originality are high.
In a seeded laboratory model, some initial conditions of life, civilization, or environment are deliberately prepared, after which development is allowed to unfold relatively autonomously. This can occur within the same physical substrate or inside a simulation substrate, so it is not only an intermediate case but can overlap both ends.
In a simulation-ensemble model, initial conditions, branches, observed variables, and stopping criteria may be widely adjustable, and many runs may be compared. But simulation does not entail low fidelity: both highly detailed and aggressively compressed simulations are conceptually possible.
4. Replace “fine versus coarse” with natural-historical richness versus counterfactual breadth
It is too simple to say that reserve environments are fine-grained while simulations are coarse. A simulator might record internal states with extraordinary precision, potentially exceeding what an external observer could measure in a physical reserve.
The deeper difference concerns the ability to preserve unknown variables before one knows what matters versus the ability to compare many counterfactual variants under changed initial conditions or interventions. The first resembles the logic of Unknown Unknowns and Preservation of Raw Data; the second is strong for statistical exploration and causal comparison.
5. A portfolio of a few reserves, controlled laboratories, and many simulations is therefore possible
A higher-level investigator need not have only one research objective. It could preserve a small number of rare naturally arisen systems, use controlled laboratory worlds to test hypotheses generated from them, and then use large simulation ensembles to explore distributions and counterfactuals.
In that case, asking “which category are we in?” is itself too simple. If the higher-level agent uses several environmental formats at once, self-location becomes a selection problem within that portfolio.
6. A substrate gap may separate the base world from the lower world
When estimating the higher-level computational cost of a simulation, the apparent amount of computation inside the lower world need not map directly onto base-world cost. The higher level may have more efficient algorithms, compressed representations, on-demand detail generation, different physical computing primitives, or much larger available resources.
This project provisionally calls that difference the substrate gap: roughly, how small the cost of maintaining the lower world at the required fidelity is relative to the computational and physical resources available above. If the gap is very large, enormous increases in lower-world computation could still be cheap. If it is small, increasing lower-world computation may pressure the run budget.
It is also logically possible that the base world has more spatial dimensions or additional physical degrees of freedom and can implement a lower-dimensional world comparatively cheaply. But dimensionality by itself does not determine computational cost. A higher-dimensional base is only one possible source of substrate advantage.
7. The human-to-ASI transition need not be a privileged discontinuity for the simulator
The transition from human civilization to AGI or ASI may be enormous from the perspective of the lower world. But it emerges from a continuous causal history of biological evolution, culture, machine computation, and AI. There is therefore no a priori reason to assume that the moment an entity becomes “ASI” it becomes a categorically different simulation object.
This does not mean that humans and ASI must have the same higher-level cost. A lower-world ASI performing huge externally verifiable computations could be expensive under some implementations. The weaker point is that a cost discontinuity at ASI is not entailed by simulation as such; it requires additional assumptions about the computation performed by the lower ASI and the implementation of the higher substrate.
Bostrom’s 2003 Simulation Argument allows the environment to be compressed rather than continuously simulated at microscopic detail, while also noting that simulated computers whose outputs can be independently checked may require more continuous lower-level representation. He also mentions, as one possibility, that if nested posthuman simulations were prohibitively expensive at basement level, runs might terminate just before a posthuman stage. This page treats the latter not as a general law, but as one branch under a particular resource model.
8. Termination depends on purpose, replaceability, and resource constraints—not on simulation as such
Reasons for terminating a run can be divided into at least three kinds.
- Resource-triggered: lower-world computation or complexity exceeds the higher-level budget.
- Goal-triggered: the state, variable, or transition of interest has been observed and the run’s marginal information value has fallen.
- Epistemic-validity-triggered: the inhabitants learn about the higher-level management or otherwise alter the behavior distribution that the experiment was intended to observe.
This page gives particular weight to goal-triggered termination as a standard candidate. Because the higher-level implementation is unknown, “the run stops because ASI is expensive” may require more assumptions than “the run stops when the investigator has obtained the data it wanted.” Resource- and validity-triggered termination remain open possibilities.
Termination risk may also depend more on run replaceability than on simulation status itself. An irreproducible natural history may lose substantial information if stopped, whereas one run in a large ensemble with snapshots, branching, and restart may be cheap to prune. This mirrors the treatment of substitutability, information loss, and value uncertainty in Preservation vs. Production Civilizations.
9. ASI can be either the endpoint of the experiment or the beginning of its main phase
If the higher-level objective is to study the conditions under which biological civilization first produces ASI, then first ASI is a natural endpoint. But if the objective is to observe how biological civilization transitions into non-biological successor intelligence—and what values, societies, and cosmic behavior follow—then ASI is the beginning of the most interesting phase.
Thus, predicting that “the world is especially likely to end near ASI” embeds a hypothesis about the simulator’s objective in addition to a simulation hypothesis. This connects to the project’s discussion of the worthy or exploratory successor: the transition from biological intelligence to a different successor intelligence might itself have high informational value to a higher-level observer.
10. Massive simulation parallelism and the probability that we are simulated are different questions
Simulation ensembles might generate far more observers or observer-moments than reserve environments. Under some observer-counting rules, this would give simulations much greater anthropic weight.
But the project’s SSA, SSSA, and SIA analysis argues that observer counts alone do not uniquely determine self-location likelihood. A reference class, observer measure, and selection model are still required before “there are more of them, therefore I am probably one of them” becomes a well-defined update.
If a curated-environment hypothesis is denoted by H, the more transparent object of comparison is therefore conceptually P(H,M|E), where M includes the selection model, rather than a bare P(H|E).
11. Keep the hypothesis space open rather than treating the present three labels as natural kinds
“Zoo / Lab / Simulation” are useful prototypes, but they need not form a natural exhaustive partition of world possibilities. Unknown higher environments may use kinds of intervention, substrate, or observer structure that current concepts do not capture.
Following Bayesian Updating and Open Hypothesis Spaces, strong updating among current explicit candidates is compatible with preserving the capacity to add, subdivide, and reconstruct the hypothesis space itself. Higher-dimensional substrates, non-computational higher environments, and implementations beyond present simulation concepts should be kept under representational openness rather than assigned arbitrary precise priors now.
12. The practical implication is still not “obey the higher-level agent”
This taxonomy does not generate a morality of “there is probably an experimenter, so satisfy its preferences.” As in Simulation Uncertainty, the important point is that misidentifying reality level, management structure, or stopping rules can alter outcome predictions.
In practice, it is more robust to favor information gathering, reversibility, limited externalities, and auditability across multiple reality hypotheses than to speculate about and appease an imagined higher agent. The existing Cosmic Host distinction between descriptive existence, strategic force, epistemic authority, and foundational normativity remains intact.
13. What would weaken this model?
- A demonstration that reserve, laboratory, and simulation environments contain an essential epistemic discontinuity that cannot be represented by continuous management dimensions.
- A demonstration that lower-world computation maps nearly one-to-one onto higher-level computational cost, making ASI generally produce a clear resource discontinuity.
- Establishment of an anthropic principle that uniquely fixes observer measure and selection model with almost no additional assumptions.
- New observational evidence that strongly selects one particular form of higher-level management.
Sources / notes
Nick Bostrom, “Are You Living in a Computer Simulation?”, Philosophical Quarterly 53 (2003), 243–255. Bostrom explicitly allows compressed or on-demand environmental detail, notes that simulated computers may require more continuous representation when their outputs are independently checked, allows basement physics to differ from simulated physics, and mentions possible termination near a posthuman stage if nested simulations become prohibitively expensive. See also the Simulation Argument FAQ for the later clarification that an entire universe need not be simulated at full microscopic resolution. The terms curated-environment hypothesis space, substrate gap, and the present decomposition into management axes and termination types are working concepts introduced by this project; they are not attributed to Bostrom.