Crossing the Causal Ladder: Architectural Separation of World Model and Policy Achieves Level 2 Causal Inference and Level 3 Novel Intervention Generalisation

Paper 1 established that environmental pressure drives causal representation to a precise equilibrium, but the entangled GRU belief model failed the intervention test: do(C) accuracy remained at 40%, permanently below the 50% chance baseline. This paper presents v3 a minimal architectural change that separates the belief state into a dedicated world model (z_w, 51,456 parameters) updated only by environmental observations, and a policy module (z_p, 62,980 parameters) that reads z_w but cannot write back to it. Gradient isolation is verified by construction. Training uses an identical three-condition environment with a new three-phase curriculum. Results: do(C) accuracy climbs from 40% (below chance) to 85% (35pp above chance); CLW Benchmark B-full scores reach Level 2 across all three environments with the non-linear probe (CLW-1: 0.61, CLW-2: 0.71, CLW-3: 0.78); CLW-1 novel intervention tests reach Level 3 (C1: 1.00, C2: 0.78). The causal ladder has been climbed two rungs by changing one variable. Environmental pressure is necessary but not sufficient architectural separation of world model and policy is the missing condition for Level 2 causal inference. Builds on: https://doi.org/10.5281/zenodo.19118615