5.8 JEPA, world models, and learned quotients

Fixed doctrine.

World-model and joint-embedding predictive architectures replace direct reconstruction with prediction in a learned representation space [ Ha and Schmidhuber , 2018 , LeCun , 2022 , Assran et al. , 2023 ] . Their hypothesis world contains an encoder, a latent transition or predictor, and a family of readouts. Masked sensory fragments or action-conditioned trajectories provide the presentation. The query doctrine asks whether a latent future, relation, or planning-relevant feature can be predicted.

The doctrine presupposes that a useful quotient of observations exists, that the selected context determines the chosen target at the intended resolution, and that prediction in the latent category preserves the distinctions needed by downstream probes. Representation learning selects a realization of this quotient; it does not by itself prove that the quotient is sufficient for unanticipated decisions, interventions, or modalities.

Categorically, the encoder proposes a quotient: distinctions in observation space that no admitted latent query detects are intentionally identified. This is not a defect if the quotient is sufficient for the declared tasks. It becomes a UOCL failure when a later query, intervention, or modality needs a distinction that the representation destroyed and no conservative refinement can recover it. Accommodation then means changing the representation doctrine, not merely fitting the predictor more accurately.

Doctrinal audit.

The structural doctrine contains encoders, latent predictors, readouts, and their admitted compositions; masking or action-conditioned sensory experience supplies the presentation; representation prediction supplies the solver; and latent or control-sufficient equivalence supplies the quotient. A newly important distinction erased by the encoder is evidence against the current doctrine of representation, not simply another prediction residual.