Scientific, Medical, and Social Worlds › Limitations and Impact

The examples separate computational usefulness from validity for an intervention. A learned force model, a forecast benchmark, or a device simulator can provide useful evidence within a tested scope. None automatically establishes a reliable response to every new forcing, treatment, or policy.

Several limitations remain relevant to the worked experiment and the domain examples:

• Identification. A selected history may leave a shallow fitting-loss profile, as it does for the recovery-rate sweep. Different data, longer observation windows, known mechanisms, or justified experiments can constrain an answer. Where the available evidence does not identify the query, report the assumptions and range of answers rather than treating the selected optimum as uniquely established. Similar error on one history is weaker than equality of observable distributions.
• Extrapolation. A new composition, forcing, patient group, or institution can change relevant mechanisms. A constraint retains only its stated scope: an isolated-system energy balance is not the balance for a driven system. Uncertainty estimates, refusal rules, and held-out regimes can help diagnose a mismatch, but an architecture label cannot certify a new regime.
• Model-form uncertainty. A finite ensemble or Bayesian model average can compare included alternatives. It does not automatically cover omitted mechanisms, and a narrow interval can be conditional on a restrictive family. Explain both the included variation and the exclusions.
• Feedback. Deploying or publishing a model can change a system when agents respond to it. That response is a possible channel to model and test, not a universal claim that every deployment changes every target. Closed-loop physical controllers also create feedback.
• Governance. A prediction used in triage, restrictions, resource allocation, or policy argument can affect people beyond the modeled average. Auditability, affected groups, error costs, and communication belong alongside numerical performance. Part XII takes up these reliability questions.

Scientific world models can reduce the cost of exploring an explicitly modeled system. Differentiable surrogates can support optimization when their derivatives and reference calculations are adequate for the objective. Symmetries can constrain a learned representation, and hybrids can expose the division between specified and learned components. These advantages remain conditional: a fast gradient, an invariant architecture, or a plausible rollout is not an intervention-validity certificate.

A useful model report therefore includes the target query, represented state and observations, action semantics, governing assumptions, validation regime, and remaining uncertainty. Prediction and decision support are related but separate uses. The report should let a reader tell which conclusions follow from the evidence and which depend on untested assumptions.