Known Issues

These issues were inferred from the inspected codebase.

predict_proba path is incomplete

KFCProcedure.predict_proba() calls self.fstep_.predict_proba(...), but FStep currently implements predict() only. Add FStep.predict_proba before relying on probability outputs.

Random state is passed to stateless combiners

CStep._build_combiner() injects random_state into combiner_params if it is missing. Stateless combiners such as MeanCombiner and MajorityVoteCombiner do not define random_state in their constructors, so this can raise a constructor error depending on usage.

Potential fix: only pass random_state if the target constructor accepts it.

Internal split is fixed at 50/50

KFCProcedure.fit() uses train_test_split(..., test_size=0.5). This is simple and helps calibrate the C-step, but it reduces the number of samples available for local model training.

Potential improvement: add a calibration_size parameter.

NaN prediction risk in F-step

FStep.predict() initializes predictions with NaN. If a predicted cluster has no corresponding trained local model, the NaN can propagate to the C-step.

Potential improvement: add fallback strategies such as nearest available cluster, global fallback model, or cluster-prior fallback.