Operator-Mismatch Contrastive Learning for World-Relative Trajectory Coherence

World-model evaluation for model-based reinforcement learning typically asks whether the learned model predicts reward and value well, which can leave planning-relevant errors in the model's latent rollouts unmeasured. We introduce a complementary diagnostic, operator-on-F, that compares a model's k-step latent pushforward to the environment's on an observable subset F, using the model's own predictor. On a TD-MPC2 size sweep over cheetah-run, reward-prediction error stays within [0.028, 0.091] for every model size - only about 3x variation - so an unnormalized reward-fit check has narrow resolution to distinguish them; the (unnormalized) Bellman residual and reward error themselves have weak relationships with return (Spearman -0.10 and -0.30). Operator error spans 0.28 to 2.62 over the same sizes. At 317M the operator error is 2.62 - an order of magnitude above the 0.28-0.36 cluster - and the planning return collapses to 0.9, while reward-prediction error (0.091) is the highest of the five but stays within the same small [0.028, 0.091] range as the rest of the sweep. The rank correlation between operator error and return loss is -0.90 (anchor-bootstrap 95% CI [-0.90, -0.70] at n=5 sizes; leave-one-out removal of any single size leaves it at -0.80 or stronger). The operator also returns informative, architecture-discriminating estimates in a cross-architecture comparison between TD-MPC2 and a pure-SSL latent world model. The operator diagnostic complements value-equivalence rather than replacing it.

Planning-time intervention fidelity is a distinct, measurable property of a learned world model: whether the model's own open-loop transitions move task variables the way matched environment interventions do. In the settings we test, it is neither revealed by reward fit nor ensured by task-anchored training. Across released TD-MPC2 checkpoint sizes, episode return falls as an operator-error diagnostic on task observables grows, while reward-prediction error stays small and nearly flat, and a self-supervised world model trained without task signal preserves the same operator substantially better than a task-anchored model on the shared task. A capture-gated matched-intervention audit then localizes what fails. On Cheetah, three LeWorldModel checkpoints capture the current task query and support decodable real intervention effects; however, their imagined five-step effects are worse than predicting no effect and worse than an environment-endpoint oracle. The failure is task-direction rotation with excess gain, not feature collapse. This severe pattern is conditional: five PreJEPA seeds retain an oracle-relative deficit without it, Finger Spin experiments extend the deficit beyond locomotion with heterogeneous severity across seeds, and shared-bank effect geometry is both candidate- and support-dependent. We also test practice-side questions. In DreamerV3 the posterior distribution, not its sample, carries the current query; ensemble disagreement ranks error only near training support; and a frozen support-aware score degrades held-out error ranking in both tested transfer directions while native disagreement remains informative in both. We conclude that intervention fidelity must be audited directly, capture-first, on the model's native interface.

World models promise a general route to embodied intelligence: learn predictive dynamics once, then reason, plan, and act with them. Increasingly, the representations beneath such models are pretrained on large-scale video, interaction, and multimodal corpora, which raises a question prediction quality alone cannot answer: when is a learned representation actually actionable? We identify a failure mode we call counterfactual collapse: a model predicts visually plausible futures while failing to distinguish interventions with different behavioral consequences. This arises whenever a representation is optimized for perceptual similarity rather than intervention structure, which is precisely the objective under which most large-scale pretrained encoders are learned. We introduce Counterfactual Latent World Models (CLWM), which combine a recurrent belief-state encoder, action-conditioned latent dynamics, and a contrastive counterfactual objective that separates futures induced by distinct interventions even when their observations look alike. Across occluded manipulation, aliased navigation, and long-horizon manipulation, CLWM improves planning success over the strongest baseline (65.1% $\to$ 74.6% on Occluded Push and 67.3% $\to$ 78.9% on Aliased Maze) and reduces exploitative planning failures (18.4% $\to$ 9.7% on Deferred Kitchen), with ablations attributing the gains to hard counterfactual negatives, especially perceptual-alias negatives. Finally, our counterfactual separability metric, which tracks planning success across the five baseline model classes ($r \ge 0.94$), is representation-agnostic: given intervention-outcome labels, it can audit any encoder, pretrained or trained from scratch, before a planner trusts it. We do not yet measure it on large-scale pretrained encoders. Here we establish the metric and its relationship to planning success for world models trained from scratch.