World models assess explanatory-reach through action-conditioned prediction

Type: kb/types/note.md · Tags: foundations, computational-model, self-improving-systems

Reach-assessment is not limited to natural-language judgment or formal symbolic proof. A learned predictive world model can assess explanatory-reach when a candidate commitment's claimed generality is represented as predicted consequences under actions, interventions, or environment shifts, and those predictions are tested against the class of changes the commitment claims to cover.

This is a distributed-parametric route. The retained artifact is not a theorem, causal graph, or natural-language explanation; it is a learned representation plus predictor. Its explanatory-reach is visible when the predictor continues to support correct anticipation or control outside the exact observations that fitted it.

LeCun's JEPA line is close to this route. Meta's V-JEPA article frames the predictor as an early physical world model: it predicts masked video regions in an abstract representation space rather than reconstructing pixels. Meta's V-JEPA 2 release makes the reach-assessment shape more explicit: after action-conditioned training, the predictor can imagine consequences of candidate robot actions, score them against a goal, and replan in new environments. The separate publication abstract supports the reported action-conditioned post-training and zero-shot deployment conditions, but not those planner details.

That is not the same as formal symbolic systems assessing explanatory-reach through causal and proof obligations. A formal symbolic route checks consequences inside an explicit causal model, theorem, invariant, or proof obligation. A learned world-model route probes a latent predictor by asking whether its action-conditioned rollouts keep working under specified interventions or shifts.

The formalization boundary reappears here in a different representational form: a world model can support reach-assessment only as far as its learned state, action conditioning, training distribution, and evaluation regime cover the intended claim. A model that predicts familiar videos, or succeeds on one robot setting, has not thereby assessed explanatory-reach for arbitrary physical reasoning. The explanatory-reach enters through the counterfactual test surface: what happens when the agent considers an action, shift, or unseen setting the commitment says should still be covered?

So world models belong beside causal/proof obligations, not inside them. If a system's commitment is stored in natural-language form, semantic judgment is still needed. If it is stored as a symbolic causal or proof artifact, formal machinery can carry the assessment. If it is stored as a learned predictive artifact, action-conditioned prediction and shift testing can carry part of the assessment.

Reach can be assessed, but nothing can be localized

A learned world model and a retained theory are both model-based learning: an internal model stands between evidence and behavior change. Since the predictor can assess reach, reach-assessment does not separate them. What separates them is the localization axis of representational form. A failed rollout says the predictor is wrong somewhere, not which commitment is wrong, and the only repair is more fitting. A theory's failure can be attributed to a part, and the repair edits that part while leaving the rest, which is what theory refinement requires of its object. The choice between retaining a theory and training a predictor is therefore not a choice about who can judge reach. It is a choice about whether a later correction must reach one commitment without disturbing the others, which is the bet the sample-efficiency conjecture makes. An inspectable causal model or simulator program is both a world model and a theory, and sits on the localized side.

Relevant Notes: