What is JEPA, and why does a governance lab care?

Joint-Embedding Predictive Architecture is Yann LeCun's proposal for how a system should learn to predict: not by generating the next token or the next pixel, but by predicting the *representation* of a missing piece of a scene from the representation of what surrounds it. I-JEPA does this for images; V-JEPA extends it to video. The wager is that most of what makes a prediction useful — the structure, not the texture — survives compression into an embedding, and that training a predictor in that space is both cheaper and more robust than training it to reconstruct raw pixels.

Why would a lab that spends most of its time on deterministic engines and promotion gates care about a self-supervised world model? Because our argument across this research program has been that language models are one expert among several, not the substrate — and the honest version of that argument has to say what *is* missing, not just what a solver or an SMT search can already cover. Formal methods are exact about what a system is permitted to do. They are silent about what a system should *expect* to happen next. A world model is a candidate answer to that second question, and JEPA's bet — that prediction in representation space generalizes better than prediction in output space — is directly relevant to a Formation that has to simulate a plan before committing to it.

We have not built anything here. This is the frankest card on this page: JEPA sits in our schedule as the December 2025 paper in the series, and as of today it is unwritten. We are including it because a genuine research lab should show you the question before it shows you the answer, and because the honest way to say "we don't have this yet" is to name it rather than omit it.