Design Plugin Architecture for UQ Methods

[kiplangatkorir]

Problem

As UEQ grows, adding new UQ methods directly to core risks:

• Tight coupling
• Slower review cycles
• High maintenance burden
  A lack of clear extension points also discourages external contributions.

Proposal

Design and implement a plugin architecture that allows UQ methods to be added as first-class extensions without modifying core logic.

Scope

• Abstract base classes for:
  • UQ methods
  • Calibration strategies
  • Evaluation hooks
• Well-defined lifecycle:
  • fit → calibrate → predict → evaluate
• Plugin discovery mechanism (entry points or registry)
• Backward compatibility with existing UEQ methods

API Sketch

class BaseUQMethod:
    def fit(self, model, X, y): ...
    def calibrate(self, X_cal, y_cal): ...
    def predict(self, X): ...

UQ(model, method="my_custom_uq")

Acceptance Criteria

• New UQ methods can be added without touching core code
• Clear, documented extension points
• Minimal performance overhead
• Example plugin included in repo
• Architecture diagram in documentation