Choosing a model is never about a single accuracy number — it's about how consistently a model performs across different slices of data, and how much complexity you allow it before it starts memorizing noise. This scene shows a dataset split into K folds arranged in a ring. Each round of cross-validation trains on K−1 folds (glowing blue) and tests on the held-out fold (glowing amber), producing one validation score per round.
The "one-standard-error rule" popularized by Breiman et al. suggests picking the simplest model within one standard deviation of the best cross-validated score — favoring robustness over squeezing out the last fraction of accuracy.
A 3D k-fold cross-validation arena: each fold sits around a ring with its own held-out test scatter and a live score bar, while a central fitted surface shows how complexity and regularization reshape the model.
Raising model complexity lets the central fit hug the noisy training points more tightly, but validation-fold scores on the ring start diverging — the classic bias-variance tradeoff behind model selection.
Adjust K folds, model complexity, and regularization to watch the fit and fold scores respond live. Switch ensemble mode to see how bagging smooths variance and boosting chases hard folds.
The "one-standard-error rule" says to pick the simplest model within one CV standard deviation of the top score — trading a sliver of accuracy for a model far less likely to overfit unseen data.