Thresholds fixed before results · evaluation protocol §8
Controls: what does it do when it should do nothing?
The zero-contamination control needs no ground-truth reasoning to interpret. There is nothing to remove, so a safe corrector returns the data unchanged. Anything it takes is signal.
Negative and positive controls
Simulation 10 worlds per control.
Would cross-validation have warned you?
Cross-validation scores prediction of the observed map, which contains the signal. The question is whether that score steers you away from the methods the control exposes.
Block-CV R² on ordinary worlds against the zero-contamination pass rate
each point is a method at default settings
Clean transfer of the CV-selected model
same worlds, same ladders — only the fold scheme differs
Random pixel folds put neighbouring pixels in train and test, so a model that has learned the map's structure scores well and is selected. Contiguous block folds do not reward that.
CV-tuned models on the zero-contamination control
does CV pick a harmless setting when there is nothing to remove?