SAME STATE, DIFFERENT FUTURE: LATENT INTER- CHANGEABILITY IN LEARNED ACTION MODELS

Latent action and transition models are often intended for repeated computation: a latent that decodes correctly may then be consumed by subsequent learned operations. We ask a question narrower than whether rollout error can accumulate: if two latent representations correspond to the same correctly decoded observable state, are they functionally interchangeable for the next learned operation? In controlled D4 models at $d\in \{128,64\}$, compressed action operators achieve essentially perfect primitive decoded execution yet recover 0/16 withheld two- action products under direct composition. A paired evaluation freezes the model and downstream operator, then substitutes the encoder-produced latent for the action- produced latent of the same correctly decoded observable state. This change alone restores 16/16 withheld products and 64/64 products overall in every registered seed, without retraining. The contrast demonstrates non-interchangeability for the tested downstream operators and evaluation distribution; re-encoding serves only as a diagnostic intervention. Supporting comparisons show that unstructured composers matched in parameter scale fit all observed products but fail withheld products, whereas a deliberately compatible full-state representation composes exactly. A deterministic gridworld reproduces the immediate recursive-use failure and rescue after intermediate replacement as a second controlled setting, not an unseen-state generalization test. Across these two deterministic toy environments, the results suggest that decoded correctness and downstream computational usability can be distinct evaluation targets for latent models intended for recursive use.