In an attempt to get layering correct I mostly do not want feature derivation to be possible in the scoring layer of honeybee. If you want timeLoggedOut - timeLoggedIn then the user must first create a feature called timeOnSite rather than referring to the difference in the model.

However, I'm tempted to allow the definition of multidimensional features in the model as a list of feature names (and then a comparison is a list of the same length of the correctly typed comparisons). This is mostly because features will be encoded in sql databases which don't have their own support for multidimensional features.

I'm mostly bearish on this idea but I figured I'd create the issue.

I wonder if you'd be better to have a NumericLT, NumericGT, NumericEQ, StringEQ, BooleanEQ, BooleanNEQ, etc, rather than the enum + union you've got going now. Would feel more "typed" and harder to screw up.

Can't you just use coefficients for this?