Scientific, Medical, and Social Worlds › Summary

Scientific, medical, and social applications ask world models to connect observations with specified mechanisms and actions. The distinction between a successful fit and an identified intervention runs through each domain.

• Molecular structure prediction, force estimation, dynamical rollout, material stability, and synthesis are distinct tasks. A potential and numerical integrator together determine a molecular trajectory. Applicable symmetries constrain the representation without guaranteeing accuracy.
• Weather forecasts depend on initial state and horizon; climate questions concern distributions under specified forcing as well as relevant initial conditions and coupled processes. GraphCast's forecast benchmark and NeuralGCM's tested hybrid simulations have particular scopes, not blanket extrapolation guarantees.
• Physiological models need a stated observation process, parameter scope, and intervention. The volume-correct compartment equations preserve their drug-mass balance. Predicting an outcome is not identifying a treatment contrast; consistency, exchangeability, positivity, and interference assumptions must match the estimand.
• Social models may include expectations and adaptation. The Lucas critique concerns invariance of policy-dependent decision rules. Randomized and quasi-experimental designs are possible under different conditions; none is an assumption-free substitute for a specified causal argument.
• Constraints should match represented boundaries and numerical implementation. Observation, parameter, model-form, and numerical uncertainty have no universal ordering. Explain what intervals and ensembles include, and test the regimes needed for the intended query.
• Possible misuse includes presenting conditional numerical precision as established intervention evidence, treating unvalidated patient simulations as advice, or presenting generated social responses as representative population data. This chapter gives no clinical or social-policy recommendation.

The synthetic intervention begins at week 25 after observing weeks 0–24. An action-responsive SIR fit and an autoregression without action inputs produce different forecasts for weeks 25–44. The fitted SIR also misses the true recovery coefficient, and selected refits with similar errors disagree on intervention totals while all having net declines in this window. These observations concern the constructed experiment, not every statistical model or real outbreak.

The next chapter, The World-Model Evaluation Ladder, begins Part XI by asking what evidence different evaluation levels provide. The state, observation, action, identification, and validation distinctions developed here give that ladder concrete questions to test.