What latent predictive learning warrants inferring about the world, and what it makes unrecoverable downstream
Fourth installment in the series “Encoding, transduction and world models.” It follows part 3/3 (chapters 6-9) and extends “Learning What Cannot Vary”. The full article, with its 31 references and its seven-axis architecture diagnostic, is available in the PDF alongside.
A learned representation does not describe a world. It decides which differences survive pretraining and which disappear. Under a pretext objective — predict a masked target, minimize a distance in a jointly learned latent space — the apparatus fixes a retention policy. What it does not retain, no supervised head recovers downstream: it projects what has survived, it does not resurrect what has been destroyed. This irreversibility, and not the quality of the learned invariances, is the fact on which architecture decisions depend in regulated environments. JEPA serves as the case study because its objective is explicitly a predictability objective, which exposes the retention policy instead of concealing it.
Four sets suffice to locate the problem: I(X), the information in the observation; I_pred, what the apparatus makes predictable; I_Z, what the representation retains; I_Y, what matters for the decision. Pretraining pushes I_Z toward I_pred, since nothing in the objective encourages preserving what does not contribute to latent prediction. Three regions result, only one of which is dangerous: I_Y ∖ I_pred, the decision-critical information weakly predictable from its context. In medicine this configuration is structural, not accidental. The sentinel event, the atypical presentation, the low-prevalence lesion are exactly what the context does not allow one to anticipate. What is noise for the pretext problem is sometimes the clinician’s signal. And the governing factor is not rarity but weight in the expected objective: a representational tail risk, at present absent from every reported metric.
A predictive structure can be perfectly stable and causally wrong: the site shortcut survives the invariance objective, a device artifact being precisely what varies least within a site. Causal discrimination is not a matter of data volume but of experimental apparatus — distinct environments, declared identifiability assumptions. To this is added endogenous drift: a model that guides triage modifies the population examined, hence the subsequent distribution; it no longer observes the world it learned but the one it helped produce. The governance consequence is a requirement current validation practice does not cover: documenting what pretraining has made unrecoverable, through targeted probing of the frozen representation on known critical distinctions. The intelligence of a deployed system lies neither in what it observes nor in what it stabilizes, but in the aptness of the differences it has chosen not to destroy. That choice is not a property of the model. It is an architect’s decision, and it must be documented as such.
The next installment, “A Persistent Memory Is Not a Biographical Memory,” addresses the axes of memory persistence and self-model left empty here.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.