The safe model library

These are the only estimators the agent can ever run. Each one is written out by hand in Python. The AI layer may suggest names, but a suggestion that is not on this list — or that belongs to the wrong problem type — is rejected, and the reason is shown on your results page. No code returned by the AI is ever executed.

Classification

used when the target is a category

Model	Needs scaling	What it is good for
Logistic Regression	yes	Linear, fast and highly interpretable - the sensible starting point.
Decision Tree Classifier	no	Captures non-linear rules and is easy to explain to a non-technical audience.
Random Forest Classifier	no	Averaging many trees controls the overfitting a single tree suffers from.
Extra Trees Classifier	no	Like a random forest but with extra randomness in the splits, often faster.
K-Nearest Neighbors Classifier	yes	Predicts from the most similar historical rows. Needs scaled features.
Gaussian Naive Bayes	no	Very fast probabilistic baseline; assumes features are independent.
Support Vector Classifier	yes	Strong on small, clean datasets but scales poorly with row count.
Gradient Boosting Classifier	no	Builds trees sequentially, each correcting the previous one's errors.
HistGradientBoosting Classifier	no	A histogram-based booster that stays fast on larger datasets.
Baseline (Most Frequent Class) baseline	no	Always predicts the majority class. Any real model must beat this.

Regression

used when the target is a quantity

Model	Needs scaling	What it is good for
Linear Regression	yes	The reference model: a straight-line relationship with readable coefficients.
Ridge Regression	yes	Linear regression with a penalty that keeps correlated features stable.
Lasso Regression	yes	Shrinks weak coefficients to exactly zero, performing its own feature selection.
Decision Tree Regressor	no	Splits the data into regions and predicts the average of each region.
Random Forest Regressor	no	Usually the strongest all-round choice on tabular regression problems.
Extra Trees Regressor	no	A faster, higher-variance cousin of the random forest.
Gradient Boosting Regressor	no	Sequential boosting, often the most accurate model on structured data.
HistGradientBoosting Regressor	no	Histogram-based boosting that handles larger datasets comfortably.
K-Nearest Neighbors Regressor	yes	Averages the target of the most similar historical rows.
Baseline (Mean Prediction) baseline	no	Always predicts the average value. Any real model must beat this.

A baseline is always trained alongside the real models. It predicts the majority class or the mean value and does nothing clever at all — which is exactly the point. If a model cannot beat it, the model is not adding value, however good its accuracy looks.