Categories of causation are salient or well defined. However, but they may play an important role in organizing are search. Consider these points:

Our universe of cognition and discourse may recognize many chains of events like A=>B=>C=>D=>E where we find it meaningful to say that event A was a contributing cause to event E. Some of the intermediate stages may be hidden, they may involve different forms of force-effect transmission: mechanics, energy, speech, agent goals, weather, seasonality,… To narrow our search, we need ways circumscribing the sets of transmission mechanisms we are currently interested in looking for, their possible time course, whether we expect to see them showing up in our data stream, etc. This set of problem defining info can lead to inner/outer shells of event space within which we can look for mutual information and potential causal models. In some cases, there will be questions about the existence of unobserved connections,and interactive or other mechanisms to go out side of the dataset & try to gather answers.

Edit: This angle provides a good example of the desire for a platform architecture in which additional tasks can be hooked into the stat modeling & analysis – our event filter modeling wants to be informed when patterns/latent features that fit some event filter criteria were positvely added in a stable model. The publish/subscribe mechanism of signals & slots is a good way to accomplish that coordination. The writers of the statistical packages that built a new component don’t have any awareness or connection to the subscribers that are noticing them as event components – much like the builders of integrated desktop applications don’t focus on how the app coordinates with the user changing color preferences, audio filtering, or changing light conditions in the room – those things become part of the infrastructure that is exposed to developers of other layers.

Commonsense reasoning & efficient communication between similar entities requires & benefits from informative priors on the utilities of different kinds of events & the assumption that we and similar entities have a value interest in noticing events with high negative & positive utility.

These priors can be adapted & improved by additional knowledge & refinmement of models for other minds/interest groups, as well as special applications like investing tips.

But we shouldn’t have to program in “leaving the poison open on the kitchen counter or destroying the car were significant events.”