Foundation-model agents are often certified with safety logic that sees only the current state, treating model proposals as ordinary actions. This work shows that abstraction can be unsound for certification even when it covers every proposal. In a finite robust interface, the state-only viability kernel is contained in the physical projection of the history-augmented kernel, and the collapse is lossless only when every proposal-conditioned fiber has a common robust-safe intervention. The gap can be maximal with constant-size proposal and history alphabets.
Because the proposal is observed before intervention, conditioning on it can also restore feasibility when it separates latent modes with incompatible safe interventions. The authors extend the one-step result over time with exact finite beliefs, separating indefinite viability from finite worst-case progress. Controlled model-in-the-loop tests reproduce the predicted obstructions when telemetry, effect , or intervention authority is removed or restricted.
That is interesting because it identifies exactly when proposal-history correlation matters for certification, not because it adds a new fixed-point calculus.