adaptive-machine-learning / CapyMOA / 35929972803

84%

push

github

test: lock seed-reproducibility contract for stochastic classifiers

A systematic seed-determinism audit (all 26 classifiers in the test
matrix, ElectricityTiny, windowed + per-instance prediction traces)
confirmed random_seed is correctly forwarded to MOA learners: identical
seeds reproduce prediction traces exactly, and stochastic ensembles
(ARF, LeveragingBagging, OnlineBagging, SRP) genuinely consume the seed
(212-296/300 per-instance proba vectors differ across seeds).

These tests keep that property from regressing: a refactor dropping the
setRandomSeed call in a MOA wrapper base would otherwise pass the
existing suite (accuracy assertions tolerate it) while silently breaking
reproducibility of user research.

Deterministic learners are intentionally excluded (MajorityClass,
NoChange, HoeffdingTree, ... ignore the seed by design); the
different-seed test enforces that exclusion stays accurate.

Assisted-by: codebuff:deepseek-v4-flash

30 of 30 new or added lines in 1 file covered. (100.0%)

9635 of 11816 relevant lines covered (81.54%)

0.82 hits per line