Craver (2007) has argued that constitutive relevance can be discovered by mutual manipulability, based on interventions. However, the requirements on interventions make mutual manipulability of mechanisms and their constituents impossible.

Craver, Glennan, and Povich’s (2021) thesis that constitutive relevance can be reduced to “causal betweenness”, between the input and the output condition of a mechanism, is not adequate for multi-level mechanisms. Their account of the construction of models for mechanisms leads to the paradoxical result that there are no levels and no multi-level mechanisms, but only causal chains of fundamental level activities.

Models of multi-level mechanisms can be constructed on the basis of empirical information although the relevant experiments directly provide only information about causal relations, because this causal information can bear on variables at different levels. A multi-level model is built in two steps. 1) first, partial purely causal models are built for each hypothetical constituent variable i, on the basis of top-down and bottom-up experiments that modify or measure i in a level-specific way, 2) second, those partial models are merged in a comprehensive model containing both causal and constitution relations between variables, on the basis of information about the level of each variable and spatio-temporal constraints.