Two estimands, two questions
A risk model asks what is likely to happen under the observed operating process. An intervention-effect model asks how the outcome would differ under a specific action relative to a defined alternative.
Conflating the two produces recommendations that sound quantitative but have no defensible counterfactual.
What can be claimed before causal evidence
Where randomised or credible quasi-experimental evidence is unavailable, an engine can present bounded, versioned scenarios and sensitivity ranges. It must label them as non-causal and keep them out of realised-value claims.
Completion probability, direct cost, side effects and action feasibility remain useful even before heterogeneous treatment effects can be identified.
Measurement is part of the product
Action assignment, approval, completion and delayed outcome evidence must be captured consistently. Without them, the organisation cannot learn which interventions work, for whom and under which conditions.
A mature programme moves from scenario evidence to tenant-calibrated treatment evidence without rewriting the product’s action contract.