Counterfactual coverage metrics in parameter space

Develop metrics that quantify coverage of counterfactual interventions in parameter space θ, weighted by the causal relevance of each parameter to the downstream quantity of interest, to serve as the analogue of dataset coverage for instrumented data.

Background

Instrumented data enables executable counterfactuals via interventions on mechanistic and confounder parameters. However, there is no established analogue of dataset coverage that reflects how thoroughly these interventions span causally relevant regions of parameter space.

A formal coverage metric would guide corpus design, benchmarking, and downstream validation strategies.

References

Nine open questions will determine whether instrumented data matures into a recognised substrate for scientific machine learning. Counterfactual coverage metrics. What is the analogue of dataset coverage in \theta-space, weighted by causal relevance to the downstream quantity of interest?

— Instrumented data for causal scientific machine learning  (2606.07865 - Wilke, 5 Jun 2026) in Section 7, Methodological questions for the community, Item 2

Validating world-class assumptions linking observed and unperformed interactions and developing tractable objectives remain open.

— When Do Intrinsic Rewards Lead to Exploration?  (2610.02159 - Viteri et al., 1 Oct 2026) in Section 5, final paragraph of Section 5, "Related work and limits"