Probing JEPA Latent Spaces Against Simulated Ground Truth: Fidelity, Sufficiency, and Linguistic Properties

Probes are used to claim that a representation encodes a property, but a probe that succeeds may reflect its own capacity rather than the representation. We study probing of JEPA latent spaces with two controls: simulated systems with known state, where a probe's fidelity can be measured exactly, and linguistic properties (negation, tense, and syntactic depth) with matched control tasks. A pilot with a random encoder shows selectivity that grows with width; the JEPA probing results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.