16.12 Open problems

Simulator adequacy.

A system may learn a perfect theory of a flawed simulator. Multi-simulator comparison, empirical residuals, and explicit discrepancy models are needed.

Intervention semantics.

Many code edits change several mechanisms simultaneously. Typed mechanism-replacement interfaces are needed if an action is to support a causal interpretation.

Structural observability.

A missing mechanism may have no public observer. The system must distinguish nonidentifiability from absence and decide whether a new sensor or derived measurement is admissible.

Scale separation.

Infinitesimal probes are local, while drought, evolutionary change, and atmospheric transport are multiscale. Higher-order Weil profiles and finite sketch extensions must be connected without pretending every change is infinitesimal.

Transport to nature.

The strongest simulator-relative result remains a hypothesis until supported outside the executable world. A mature system must produce a feasible empirical admission plan with discriminating outcomes, uncertainty, cost, and ethical constraints.

. Can an agent use a simulator not merely to search a supplied model, but to construct a new causal distinction that remains useful when the simulator, regime, or real experiment changes?

The simulator supplies interventions and executable counterfactuals, but real scientific programs are also assembled from papers, reports, datasets, and partially conflicting claims. Chapter 17 removes the privileged executable world and asks how a document collection can be treated as an implicit, incomplete presentation of a theory whose proposed extensions must ultimately reconnect to experiment.