Why the Stages Compose

Core Pillars of Probly splits the library into transformation, representation, quantification, and evaluation. Three dispatch mechanisms hold those pillars apart:

  • Type dispatch (flexdispatch) routes predict, representer, and quantify to the right backend based on the object handed in.

  • Traverser dispatch (flexdispatch_traverser) walks a network layer by layer, so a transformation is defined per layer type instead of per model.

  • Value dispatch (switchdispatch) maps names to implementations, which is what makes string arguments such as predictor_type="logit_classifier" work.

The practical consequence is the benchmark: one pipeline, 20+ methods, changed one line at a time. If you are adding a method, the pillar structure is also the checklist — a new method needs a transformation, must declare which representation it produces, and inherits stages 3 and 4 for free. That is why the recipe in Adding a New Method is mostly about pillar 1: registering the per-layer transformations for each backend, and then declaring the hooks that tell pillars 2 and 3 what the method hands back.

See also

Adding a New Method for the concrete checklist, and Uncertainty Methods for the catalogue, grouped by the representation each method produces.