World models based on Joint-Embedding Predictive Architecture (JEPA) have demonstrated emergent physical understanding through Violation-of-Expectation (VoE) paradigms. However, the "surprise" metric used to evaluate these models conflates statistical novelty with genuine causal reasoning. This paper introduces Mind the Ladder, a diagnostic benchmark and metric suite for testing causal fidelity in latent world models. The framework operationalises Pearl's Ladder of Causality (Level 1: Association, Level 2: Intervention, Level 3: Counterfactuals) directly in the latent space of a trained world model, making it architecture-agnostic. Three novel metrics are proposed: AAP Surprise Ratio, Structural Invariance, and AAP Consistency Advantage all grounded in the LeWorldModel (LeWM) architecture. The benchmark is validated on the Glitched Hue Two Room environment, which tests causal disentanglement between spurious correlations and true causal mechanisms. Results show that VoE surprise alone is insufficient: a model can exhibit high surprise for physical violations while still failing Level 3 counterfactual tests.
Di Prodi Paolo (2026) studied this question.