Belief-Intersection Forcing Shield

1. Exact integration point: place the shield between a recurrent, transformer, or world-model policy and the environment. Use an RNN or learned world model to maintain a particle belief $\hat B_t=\{x_t^{(1)},\ldots,x_t^{(K)}\}$ from the observation history. The policy outputs logits over candidate forcing masks $U\subseteq E_f$; the shield filters those masks using a learned or symbolic transition model. 2. Pseudocode: compute $\hat B_t$; for every particle $x^{(k)}$, enumerate candidate masks $U$; retain $U$ if every unforced event is harmless or disabled and every resulting successor avoids $X_{bad}$; intersect the retained masks across particles; choose the highest-policy-logit mask in the intersection; if it is empty, execute a conservative emergency action and record an inconsistency certificate. 3. Computed quantities: the intersection rule and safety test come from the forcing-consistency mechanism. Beliefs, transition probabilities, unsafe labels, and candidate ranking are estimated or learned. For stochastic models, replace universal checking by a confidence bound over particles and require estimated violation probability below $\delta$. 4. First cheap experiment: use a small partially observed gridworld or CartPole with hidden velocity and an event-based action set containing a forcible emergency brake. Compare a point-estimate shield, ordinary action masking, and the belief-intersection shield. Sweep observation aliasing and belief size $K$. The falsifiable prediction is a sharp feasibility boundary: safety failures should remain near zero while $A_B(o)$ is nonempty, then increase abruptly when the empirical fraction of empty intersections rises. At matched model error, the robust shield should have lower hidden-state violation rate, while empty-intersection frequency should predict failures with at least 0.9 precision.