Research limitations

WSM provides Python access to model weights and records evaluation comparisons. A scripted edit or lower public loss does not establish capability improvement, generalization, or independent edit discovery. Private grading requires a separate process with data unavailable to the editing program.