nucleation

CLAIM-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

4 min read

Properties13
StatusProposed
Version2
Rolethesis
Warrantinterpretive
Tagsagency, cognition, philosophy-of-science
Linedistributed-agency
Bears onQUESTION-086 — Can two systems with matching behaviour and persistence be told apart by whether their own processes generate and revise the constraints that keep them going, and by what intervention?
Rests onCLAIM-414 — Matched ordinary trajectories can conceal differences in an interpretation's efficacy, in the allocation of that efficacy, and in the maintenance of its enabling constraints: in two finite constructions, systems that agree under ordinary operation differ in each, and a specified intervention exposes the difference that passive observation of the matched trajectories does not
GroundsTHEORY-092 — Behavioural capacity does not fix intelligence: a system with the capacity to respond sensibly to every possible input can lack it, so whether behaviour is intelligent, or the mark of any mental property, depends partly on how it is produced, THEORY-689 — A mobile robot built from parallel layers of simple finite-state machines, each linking sensing to action by itself and combined only by one layer suppressing or inhibiting another's signals, avoids obstacles, wanders and heads for distant places with no central world model, goal list or controller; shown at insect scale, on three layers of one physical robot, THEORY-624 — Orseau, McGill and Legg cast whether an identified system is an agent or a device as Bayesian model comparison between descriptions of the function it optimises and descriptions of its input-output mapping, so that the verdict is relative to the observer's prior; on seven toy gridworld trajectories goal-seeking comes out agent-like and wall-following device-like, and they argue that under a speed prior everything may look like an agent
ComplementsCLAIM-355 — In a distributed agent the binding of appraisal, readiness and governance into one response can be real while the single owner a self-model or collective self-description attributes it to is a representational construction, not a further entity; the illusion, where there is one, is in the attribution and belongs to whichever system adopts the representation, and this deflates the self without settling whether experience is real, CLAIM-447 — Model-based wanting is a pursuit process's ranking of the futures it predicts under candidate interventions, P(S | do(a)) against P(S | do(a₀)), while cue-triggered wanting is learned salience that need not compute those futures; a process's causal model, its wanting and its practical efficacy are separable, so learning to want differently, learning that a want will not pay, and learning not to act on it are three different changes
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 “objects to” byCLAIM-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
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 claim

Fix an interface: the agent’s actions, the environment’s observations, and a horizon T. Let agent A maintain an explicit predictive model of its environment and act on it; its policy, marginalized over its internal state, is a kernel π_A(a | h) from histories to actions. Let agent B hold no model, only the table π_B(a | h) = π_A(a | h) for every history h of length below T. Then for any environment the probability of every interaction history h₀:T is the same for A and B, by the chain rule, since each step’s action law and the environment’s response law are the same. No experiment confined to that interface and horizon can tell them apart (A20 §1).

So competence on the interface is not sufficient to establish that a world model is present, or to identify one. That is not to say it is no evidence: under relevant priors and resource assumptions, competence can still favour a model-based architecture, as the compression point below says. “World model” ascribes an organization that explains the behaviour, as “the self” ascribes an owner to it (CLAIM-355). That is the owner’s U20 point. Brooks’s robots are the standing example that competent action can run on the world as “its own best model” (THEORY-689).

Two things must be kept apart:

policy identifies predictive structure ⇏ a particular internal component realizes it.

Richens et al.’s result belongs on the left (CLAIM-513). Mechanistic identification belongs on the right. The evidence for each differs.

  • Compression. A table grows with the number of histories, and a model can be exponentially smaller. Where resources are bounded, as in every physical agent, matched competence under a resource bound can favour a model-based explanation (THEORY-624 casts this as model comparison relative to a prior). It favours that explanation over a particular table. It does not identify the architecture: many compact controllers are not explicit generative world models.
  • Generalization beyond what the implementation was built to cover. A model can serve new objectives or new horizons without relearning, and a table cannot go beyond its domain. This too supports comparing architectures, and no more. A table whose domain already includes the goals generalizes across them, so generalization within the covered family does not discriminate.
  • Evidence that discriminates implementations. Establishing an internal mechanism needs observations that candidate implementations, matched in behaviour, predict differently. Targeted interventions on internals are the main source: alter a component and find the behaviour change as a predictive model’s would, and not as the rivals’ would. Matched trajectories conceal differences that intervention exposes (CLAIM-414).

What it does not say

  • That no agent has a world model. Some do, and an intervention can show it.
  • That the table encodes nothing. Functionally, B’s table determines every prediction A’s model would make that the interface can reveal. Whether that counts as “containing” a model is the question Richens et al. (LIT-1891) answer yes, for goal-conditioned agents competent across a family of composite goals. On that functional reading their theorem holds and is filed as the counter CLAIM-513. On the explicit-component reading, the claim stands.
  • That causal model, wanting and efficacy go together. They are separable (CLAIM-447); this claim is about the first only.

Note of 2026-11-01: a precedent in cognitive modelling

Bechtel and Abrahamsen (LIT-1936, NOTE-1492, THEORY-1290) make this claim’s methodological point about cognitive models in general. Such models are proposed de novo and checked only by fit, and “Empirical evidence can rule out specific models, but usually cannot decide between competing architectures” (p. 332). Their example is the past-tense debate, “irresolvable” without constraint from identified parts (p. 330). Their footnote 5 (p. 329) bears on the claim’s two readings of “contains a model”. A model whose variables track a mechanism describes it and does not perform its operations. A system whose behaviour a predictive model describes need not contain a component that performs prediction. Their remedy is anchoring variables to parts found by experiment. That is the claim’s “evidence that discriminates candidate implementations”.

No status, title or warrant changes.

Backlinks