Second-Read Divergence: Context Reuse in JEPA-WM

Long-horizon agents read the same context repeatedly. An internal intervention that is correct on the first read may no longer represent the intended change on the next. We study this problem in JEPA-WM, a world model that plans by predicting latent futures from a sliding window of two frames and actions. Each action is therefore read twice: as the newest input, then as history. Overwriting an action’s internal encoding at the first read reproduces the changed-input rollout exactly. At the second read the model sees the original action, and the rollouts diverge; we call this second-read divergence. The discrepancy persists to the planning horizon and changes the selected candidate in 42% of Reach and 59% of Reach-Wall starting states, while the excess cost of the selected candidate under the model’s own objective averages about 2%. Overwriting both reads restores exact agreement. Single-block and random-direction controls indicate that neither the location nor the magnitude of the edit compensates for an uncovered second read. An internal edit that is meant to reproduce an input change must therefore be verified at every step where the model reads that input, through to the decision the model makes.