nucleation

CLAIM-513: Across a family of goals, a world model is not an optional explanation of competent behaviour: in a fully observed, finite, communicating, stationary environment, a goal-conditioned policy within a factor 1 − δ of optimal on every composite goal of depth n fixes every transition probability to within √(2P(1 − P)/((n − 1)(1 − δ))), so observers reading the same policy must ascribe the same world model; what it leaves open is whether a predictive component is realized inside the agent, since a lookup table over histories and goals has the same policy

2 min read

Properties9
StatusActive
Rolecounter
Warrantdemonstrated
Tagsagency, philosophy-of-science, learning-theory
Linedistributed-agency
GroundsTHEORY-1249 — In a fully observed, finite, communicating, stationary controlled Markov process, a deterministic goal-conditioned policy that achieves every composite goal of depth up to n with probability at least 1 − δ times the optimum determines every transition probability to within √(2P(1−P)/((n−1)(1−δ))), read from its first action on either-or goals; for a policy competent only at one-step goals no bound follows, THEORY-1260 — Richens et al.'s sense of "an agent contains a world model" is that the transition function can be computed from the agent's goal-conditioned input-output map, uniquely to within the theorem's error, so the attribution is not a stance an observer may withhold, unlike a representation-theorem fit; but it holds of every implementation of the same map, a lookup table included, and so cannot show that a predictive component is realized inside the agent
Objects toCLAIM-519 — Competent behaviour does not establish an internal world model: over a fixed interface and a finite horizon, an agent with no model, whose actions are looked up from the history so far, reproduces the interaction distribution of an agent that has one exactly; a policy can identify predictive structure without any particular internal component realizing it, so compression and generalization support comparing candidate architectures, and establishing an internal mechanism needs evidence that discriminates candidate implementations, such as targeted interventions
Cited as “illustrates” byCASE-187 — A proposed study: agents matched on competence over a declared family of goals, one planning with an explicit transition model, one a directly learned goal-conditioned policy, one a distributed predictive-state controller, compared on the original goals, on on-demand predictive and interventional queries, under targeted internal interventions, and after environmental change and damage to their predictive machinery, to find which world-modeling achievements are linked and which come apart under matched resources
Cited as “rests on” byCLAIM-514 — World-modeling is a family of resource-bounded capabilities for producing and using predictions about an environment's possibilities and consequences, each relative to a declared family of questions; a world model in the representational sense is one possible realization of them, competent behaviour can determine environmental structure without identifying any internal mechanism, and producing, maintaining and revising world-modeling capabilities are further achievements, not prerequisites of agency in general

Graph View

The objection

Richens et al.’s reduction uses either-or goals. One goal is at most k successes of a transition (s, a) → s′ in n tries; the other is more than k. A competent agent’s first action locates the binomial median, and the k at which its choice switches fixes P (THEORY-1249). So the attribution is determined by the policy. This is unlike a Savage-style fit, which exists even for a random policy, and unlike Orseau et al.’s verdict, which is relative to a prior (THEORY-624).

Within that scope, CLAIM-519’s “ascribing a world model is an explanation of the behaviour” is too weak. The world model is a measurable property of the goal-conditioned disposition, though not of any finite run.

It also corrects one item in the thesis’s list of evidence that goes beyond the interface. Generalization across goals does not discriminate a table from a model when the table’s domain includes the goals. Such a table is competent in the theorem’s sense and yields the same world model (THEORY-1260).

What it does not say

  • That any component represents P. The theorem is silent on the agent’s internals, which is the thesis’s explicit-component reading, untouched.
  • Anything for one fixed objective, partial observation or stochastic policies. A single-task or myopic agent need encode nothing about the dynamics (their Theorem 2), and the thesis’s own setup is of that kind.
  • Anything about cost or compression. A lower bound on the resources any competent implementation needs is the thesis’s compression point, still open.

Note of 2026-11-01: “realized inside” has two readings, and interventions settle one

The claim leaves open “whether a predictive component is realized inside the agent”. The constitution debate read today (THEORY-1288, CLAIM-528) separates two things that phrase could mean. (1) The agent’s competence is produced through an internal state that carries predictions and later shapes action. That is causal. An intervention on the state, such as a patch to it, can be surgical, and it is what tells the lookup table from the model. (2) The competence consists, at a time, in such a component. Every intervention on the competence then also acts on some component along another path, and no finite set of interventions shows the relation. What is left open here is (1), which interventions can settle. CLAIM-515’s “realized” should say which it means.

No status, title or warrant changes.

Backlinks