We have filters (they score how well a transformation is expected to do), transformations and validators.

Filters get the guessed data type in multiformat and the data itself. for example: image/png, video/mp4, image/ etc. That way we can score
negatively say when we need to do transformations that need floating point precision on lossy formats.

An example filter for a transformation of a vertical split would see if they are similar in the axis we want to split on.

Filters return Optional[float] so they can discard themselves if it's certainly not going to work.

Transformations return a multiformat and data. data should be returned as an nparray always. shape can be whatever

Validators check in some way if we found anything. they should probably also return an Optional[float].