DEMONSTRATION

Try SAR

The default case is the four-feature function from the five-minute tutorial, g = A + 0.5·B + C + 0.4·D − A·C + A·B·D. Coalitions use K=4, M=128, n_fit=96, and seed=17. A repeated mask keeps the same score. This is not a model experiment and it does not run GPT-2. Recompute uses the public NumPy estimator.

Pairwise share is a held-out estimate. Custom numbers may give SAR2 above SAR1 or a negative share. That estimate is returned as computed and is not a property of the population least-squares projection.

Back to the tutorial · Use with AI agents

Recompute Restore default

SAR1, SAR2, pairwise share, rank fidelity, and ordering error appear here.