As described in https://github.com/scikit-learn/scikit-learn/issues/8443, the Scikit-Learn implementation for the decision tree algorithm by default is not deterministic as it should be. This is due to a design choice here](https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/tree/_splitter.pyx#L381-L401) where even if max_features = n_features, the algorithm still randomly samples up to max_features.
To address this unexpected behavior, the internal random state of the DecisionTreeClassifier in Araucana has been fixed to 1. The global randomness of Araucana (e.g., during oversampling) can still be controlled with the 'seed' parameter.