CLAIM-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

The claim

World-modeling competence. Let 𝒬 be a declared family of questions about environmental dynamics. Let q ∈ 𝒬 have correct answer P_q, and let L be a proper predictive loss. An organization O has world-modeling competence WM_{B,T,ε}(O; 𝒬) when there is a single procedure κ, implementable by O, such that

sup_{q ∈ 𝒬} E[L(κ(q), P_q)] ≤ ε,

with κ operating within budget B and deadline T.

The single-procedure condition matters. An outside oracle must not be allowed to pick a different best answerer for each query.

The families are not interchangeable:

• observational predictions need not identify responses to intervention;
• interventional distributions need not fix counterfactuals at the level of the individual case.

Nothing in the definition requires an explicit generative model. A neural predictor, predictive-state statistics, a database, a simulator or a distributed system of specialized processes could each meet it. And like the record’s other capabilities, it need not grow monotonically. A controller can get better at short-horizon action-conditioned prediction while it gets worse at long-horizon forecasting. So the object is a family of query-relative capabilities, not one score of world-model possession.

Where the two claims it rests on sit.

• CLAIM-513 is about the behaviourally encoded model: the policy determines the dynamics, recoverable by an external algorithm.
• CLAIM-519 is about the realized mechanism: behaviour does not identify it.

They do not conflict once each is read as being about its own notion. Recoverability from a policy does not imply that the agent can produce the model on demand, simulate with it, or update it when wrong.

What it does not say

• That an explicit world model marks superior agency. Its worth depends on the queries it serves, what it costs and the further inquiry it opens.
• That the Richens result is an observer’s fiction. The recovered model is constrained by the behaviour. It is not arbitrary.
• That competence is no evidence of a model. It does not establish one. Under resource assumptions it can still favour one (CLAIM-519).

Note of 2026-11-01: which “realization” interventions can show

The claim calls a representational world model “one possible realization” of world-modeling capabilities. It also says behaviour can determine environmental structure “without identifying any internal mechanism”. The constitution debate read today bears on the next step, identifying the mechanism. If “realization” means that the capability is produced through an internal predictive mechanism over time, interventions on that mechanism can show it (CLAIM-528, reading 1). If it means that the capability at a time consists in the mechanism, interventions cannot. Every intervention on the capability also acts on some part of the mechanism along another path, and the inference is abductive (THEORY-1288, THEORY-1294). The claim’s “producing, maintaining and revising” world-modeling capabilities are diachronic, and so of the first kind.

No status, title or warrant changes.